1 CLIENT SIMILARITY AND ITS IMPLICATIONS FOR AUDIT QUALITY AND PRODUCTION COSTS By STEPHEN V. BROWN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2012
2 2012 Stephen V. Brown
3 ACKNOWLEDGEMENTS I am especially grateful to my dissertation committee for their assistance: W. Robert Knec hel (chair), Stephen Asare, Praveen Pathak, and Jenn ifer Wu Tucker I would also like to thank workshop participants at Arizona State University, Florida International University, Southern Methodist University, the University of Florida, the University of Illinois at Urbana Champaign, the University of Kent ucky, the University of South Carolina, and the University of Virginia for their feedback on portions of this paper.
4 TABLE OF CONTENTS page ACKNOWLEDGEMENTS ............................................................................................... 3 LIST OF TABLES ............................................................................................................ 6 LIST OF FIGURES .......................................................................................................... 7 ABSTRACT ..................................................................................................................... 8 CHAPTER 1 INTRODUCTIO N ..................................................................................................... 10 2 AUDITORCLIENT COMPATIBILITY AND SELECTION OF AUDIT FIRM ............. 14 Introductory Remarks .............................................................................................. 14 Hypotheses and Prior Literature ............................................................................. 17 Auditor Client Compatibility and Selection of Auditor ....................................... 17 Auditor Client Compatibility and Auditor Switching ........................................... 20 Effect of Changes in Auditor Client Compatibility ............................................. 21 Mea surement of Auditor Client Compatibility .......................................................... 22 Financial Statement Similarity .......................................................................... 23 Narrative Disclosure Similarity .......................................................................... 26 Patterns in Client Similarity ............................................................................... 29 Validation of Similarit y Measures ..................................................................... 31 Analysis of Auditor Client Alignment ....................................................................... 33 Sample ............................................................................................................. 33 Typical Auditor Client Alignment ...................................................................... 33 Likelihood of Auditor Change ........................................................................... 35 Auditor Choice Conditional on Decision to Switch Auditors .............................. 37 Changes in Audit Quality Conditional on Auditor Client Fit .............................. 38 Discretionary accruals ................................................................................ 39 Accounting and Auditing Enforcement Releases ....................................... 39 Robustnes s and Sensitivity Analyses ..................................................................... 41 Alternative Inputs for Financial Statement Similarity ........................................ 41 Clients Switching from Non Big4 Auditors ........................................................ 42 Accounting System Comparability .................................................................... 43 Concluding Remarks ............................................................................................... 44 3 SPECIALIZATION THROUGH CLIENT COMMONALITY AND ITS EFFECT ON AUDIT PRODUCTION COSTS ............................................................................... 56 Introductory Remarks .............................................................................................. 56 Hypotheses and Prior Literature ............................................................................. 60 Production Costs and Audit Fees ..................................................................... 60
5 Specialization and Audit Fees .......................................................................... 61 Opportunities to Lower Produc tion Costs ......................................................... 62 Incentives to Lower Production Costs .............................................................. 64 Inconsistent Signals of Commonality ................................................................ 65 Sample .................................................................................................................... 66 Financial Statements ........................................................................................ 66 Narrative Disclosures ....................................................................................... 69 Analysis of Similarity ............................................................................................... 71 Audit Fee Model ............................................................................................... 71 Hypothesis 1 ..................................................................................................... 72 Hypothesis 2 ..................................................................................................... 73 Hypothesis 3 ..................................................................................................... 76 Alternative Measures and Sensitivity Analyses ....................................................... 77 Larger Reference Groups ................................................................................. 77 Minimum Reference Group Size ...................................................................... 78 Accounting System Comparability .................................................................... 78 Concluding Remarks ............................................................................................... 80 4 CONCLUSION ........................................................................................................ 89 APPENDIX A EXTRACTION OF ANNUAL REPORT ITEMS ....................................................... 93 B CALCULATION OF NARRATIVE DISCLOSURE SIMILARITY SCORE ................. 96 LIST OF REFERENCES ............................................................................................. 100 BIOGRAPHICAL SKETCH .......................................................................................... 107
6 LIST OF TABLES Table page 2 1 Descriptive statistics. ............................................................................................ 47 2 2 Correlations between similarity and difference measures. ................................... 49 2 3 Auditor selection based on auditor client compatibility. ........................................ 50 2 4 Probability of auditor change. ............................................................................... 52 2 5 Probability of receiving an AAER. ........................................................................ 54 3 1 Audit fee model variables. .................................................................................... 82 3 2 OLS regression of audit fees on client similarity. .................................................. 84 3 3 Regression of fees on client similarity conditional on auditor incentives. ............. 85 3 4 Regression of fees on client similarity conditional on disclosure consistency. ..... 87 A 1 Narrative disclosure sample selection process. ................................................... 95 B 1 Calculation of narrative disclosure similarity measures. ...................................... 99
7 LIST OF FIGURES Figure page 2 1 Trend in normalized similarity score over auditor tenure. ..................................... 46
8 Abstract of D issertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy CLIENT SIMILARITY AND ITS IMPLICATIONS FOR AUDIT QUALITY AND PRODUCTION COSTS By St ephen V. Brown December 2012 Chair: W. Robert Knechel Major: Business Administration T he current work focuses on the implications of the commonality among clients of an auditor. In Chapter 2, I examine the degree of compatibility between clients and thei r auditors to test whether companies systematically prefer specific auditors based on this criterion. Using both financial statements and narrative disclosures, I introduce two new measures of compatibility based on how similar a client is to other clients of the same auditor. My results strongly support the idea that auditor client compatibility can predict the particular auditor a client will choose to engage. When compatibility is lower, clients are more likely to change auditors and pick a new auditor w ith relatively high auditor client fit. Interquartile changes in compatibility increase the probability of switching auditors by as much as 19 percent. Audit quality, as captured by discretionary accruals, increases as auditor client compatibility increases. However, SEC enforcement actions, indicative of a severe audit failure, become more prevalent as compatibility improves. In Chapter 3, I use the similarity measures from Chapter 2 to proxy for the opportunities to specialize that arise from greater client commonality, finding strong evidence that higher client overlap is associated with lower audit fees. This relationship is incrementally stronger in industries for which the auditor has greater economic
9 incentives. Because the financial statements and narrative disclosures are distinct disclosure channels, I explore the effect on audit fees when the two channels portray inconsistent messages about the degree of client commonality. When the financial statements are relatively unusual compared to peer clients, but the narrative disclosures do not reflect this dissimilarity, I expect the auditor to assess higher audit risk. Consistent with this prediction, the auditor charges higher audit fees under this condition. On the other hand, when the narrative disclosures are more unique than the financial statements reflect, audit fees are lower, which I argue is due to the greater, more useful firm specific information contained in the text relative to the financial statements.
10 CHAPTER 1 INTRODUCTION There is little doubt that relationships between companies are important. A tighter bond between a firm and its suppliers can reduce costs, improve the timeliness of deliveries, and decrease the timeto market of new products. Closer relationships among competitors can improve profitability, as in the case of an oil cartel, or give rise to antitrust litigation. Managers with close ties to political figures might receive preferential treatment for their firm, while those who steer clear of government lobbying could pay a substantial financial cost. A substantial amount of research in accounting examines the relationships between variables such as client size and audit costs but does not fully explore the relationships between observations. Such a limitation is implicit in any study that includes only characteristics of the observation in the empirical model while excluding measures of how the observation relates to entities beyond itself. In this w ork I investigate the strength of the bond between clients and their auditor, specifically testing its implications for which auditor a client will engage, the resulting quality of the audit process, and audit production costs. I focus on the auditor client relationship in particular due to its importance for both regulators and researchers. Generally Accepted Auditing Standards (GAAS) in the U.S. and the International Standards on Auditing both require the auditor to be independent of the client in fact and in appearance, a feature at the core of auditings value propositi on. The closer the relationship between the auditor and client, the more likely independence issues will arise. This concern is reflected in recent discussions about mandatory auditor rotation by the Public Company Accounting Oversight Board (PCAOB). The PCAOB is gathering public comments on a potential new requirement
11 that auditors be changed after a given number of years due to fear that the relation auditor client bond will strengthen to the point of impairing independence of the audit process. While I d o not specifically address the issue of independence in this w ork my results have value for regulators considering the costs and benefits of mandatory auditor rotation. Another important feature of the auditor client relationship explored in many prior studies is specialization of the audit process. The typical line of reasoning is that auditors can customize (specialize) their normal audit process for any client or set of clients. Customization can take the form of modifications to the audit wor kflow, employee training, shifting personnel onto engagements to maximize the usefulness of prior onthe job experience, or in a variety of other ways. This specialization is typically costly for the auditor, but has potential benefits, such as higher audi t quality, more efficient audits, and improved marketability of the product to new customers. While a significant amount of research has examined specialization, the main weakness of the empirical approach is a lack of adequate proxies for specialization. Papers in this area normally rely on strong assumptions about the meaning of market share within an industry to proxy for specializationassumptions that may be difficult to justify (Gramling and Stone 2001). To address this concern, I develop a more direc t measure of specialization based on explicit features of the auditors client base. A major innovation in the current work is the implementation of two measures of similarity between one company and another; one measure is based on the companys financial statements and the other on its narrative, textual disclosures contained in its annual report. Taken together, these two approaches give a comprehensive view of
12 how a companys economic status and disclosure decisions imply a given relationship with other firms. For the financial statement similarity measure, I use an existing algorithm from the cluster analysis literature in a novel way within an accounting setting. For the narrative disclosure score, I extend a measure already used in an accounting contex t, adapting it to allow comparisons among multiple companies. Both of the measures have multiple potential uses within an auditing context and for accounting research more broadly. For the purposes of this w ork I use the similarity of one company to the existing clients of an auditor as a proxy for how much the company has in common with the auditors current client base. Chapter 2 considers the basic question of whether clients tend to select an auditor with which they are more compatible and the implications of auditor client compatibility for audit quality. I use the similarity measures as a proxy for auditor client compatibility based on the argument that when a company has relatively more in common with an auditors clients, it has a better fit with that auditor than an alternative with less inter client similarity. This chapter develops the similarity measures, validates them, tests their sensitivities to various implementations, and compares them with alternatives. I find that clients on average prefer a more compatible auditor, and they are more likely to switch to a new auditor when compatibility is lower. More importantly, this compatibility has implications for audit quality. One broad measures of audit quality is increasing in auditor clie nt compatibility, meaning that the quality is higher when the client fits better with that auditor. On the other hand, I also find evidence that severe audit failures are more likely with the compatibility is high.
13 In Chapter 3, the similarity measures proxy for potential specialization by an auditor because greater commonality among its client base provides more opportunities to customize the audit process for a particular group of clients, with a higher possible payoff to that specialization. I find support for this prediction in the form of lower audit fees for clients having more in common with an auditors other clients. This result is somewhat unique in the audit literature, since most specialization studies find higher audit fees, purportedly a result of higher audit quality by specialist auditors. However, my approach is a more powerful and direct measure of potential specialization than existing proxies based on industry market share. The uniqueness of my measure provides support for an alternative outcome of specialization more efficient audit processes. Chapter 4 concludes the current w ork by summarizing the key findings of the other chapters and discussing their implications for both regulators and researchers.
14 CHAPTER 2 AUDITORCLIENT COMPATIBIL ITY AND SELECTION OF AUDIT FIRM Introduct ory Remarks Clients have preferences about the audit process, its outcomes, and the nature of the relationship with their auditor. I define auditor client compatibility as the ability of the auditor to satisfy these preferences, given its own preferences and constraints.1 If client preferences vary across companies, and auditors have varied abilities to meet clients needs, the degree of fit between the two entities will also vary. Prior literature has examined variations in this compatibility in a broad sense, such as the choice of a Big4 auditor versus a smaller firm or an industry specialist versus a nonspecialist. In this paper, I examine the typical auditor client compatibility observed among Big4 firms, the occurrence of auditor changes when compatibility is relatively lower, and the implications of fit for audit quality.2 I find clients tend to be with auditors where the compatibility is better; depending on the proxy used, anywhere from 51 to 59 percent of cli ents are with the two best fitting auditors among the Big4. I observe the same selection preferences among clients switching to a new auditor. And the worse the fit with the incumbent auditor, the greater the probability the client will choose to switch to a new auditor. When examining the association between fit and audit quality, I find some evidence that discretionary accruals are lower when the auditor client compatibility is better. On the other hand, there is a greater occurrence of SEC enforcement ac tions for these higher levels of auditor client compatibility 1 In this paper, I use the term fit interchangeably with compatibility. 2 Chapter 3 of this w ork addresses the association between auditor client fit and audit fees.
15 The degree to which a client and a specific auditor are well matched is generally not externally observable. However, to the extent a company has similar audit preferences as other companies, t hey would presumably choose the same auditor, subject to various constraints on that choice. Therefore, in this study I compare the similarity of a company to other current clients of the auditor as a proxy for how well that company fits into the auditors client base. If the company of interest is very similar to other clients already audited by the audit firm, the auditor is likely to have developed expertise and cost advantages related to that type of client. Therefore, when the similarity to existing clients is high, I consider this a more compatible fit than those situations where there is very low similarity between the company and the auditors existing clients. I specifically introduce two measures of inter company similarity, one based on financia l statements and one on narrative, textual disclosures. Each source of information provides variation in what managers are disclosing and how they choose to disclose it. The financial statement similarity proxy relies on the Mahalanobis distance, used extensively in the cluster analysis literature to divide objects into groups based on sets of numbers associated with each object. The set of numbers I use is motivated by financial components known to be important in an audit context, including proxies for ef fort, complexity, and risk. The narrative disclosure similarity measure extends the pairwise similarity score introduced in Brown and Tucker (2011) as a proxy for year over year changes in MD&A. As the source of narrative disclosures, I use the business description, MD&A, and footnotes contained in the mandatory annual report. Using each
16 measure, I calculate the similarity of each client year to other clients in the same auditor industry year (the reference group). The primary contribution of this study i s the introduction of two measures of how similar a company is to a reference group of other companies. The financial statement measure, based on the Mahalanobis distance, has been used in other contexts before, but not in the accounting literature. While the financial statement similarity score is used in Brown and Tucker (2011), I extend this formerly pairwise, year over year measure to allow for comparison of one entity to a group. Defranco et al (2011) develops a measure of accounting system comparabil ity based on the relationship between earnings and returns. My measure can be used in settings outside of financial statement comparability and can be adapted to whichever financial statement variables are important in a given context. While the current pa per focuses on auditors and their clients, there are many other potential applications of these measures in any fields of research where either financial statement information or narrative disclosures are available. W ithin the audit context I contribute t o the lack of literature on the fit of specific auditors and their clients. Some literature has looked at misalignments given types of auditors, such as audit firm size (e.g., Shu 2000; Landsman et al 2009). However, there is a much more limited literatur e on the compatibility of specific auditors with a specific client. What literature exists tends to be narrowly focused, such as research on the effects on a client of hiring a former audit partner (e.g., Lennox and Park 2007). I broaden the existing research on auditor choice by considering the suitability of a specific auditor for a client. This research is relevant to the debate on mandatory auditor
17 switching, since forcing an auditor change could have negative implications for the engagement if the clie nt is currently with a first best choice of auditor. My findings are also important for researchers who are considering aspects of the auditor client relationship, since prior literature typically only contains controls for auditor types (e.g., Big4/nonBi g4, industry specialist/nonspecialist) rather than considering the specific auditor being engaged. The rest of the paper proceeds as follows. The next section develops the hypotheses and discusses prior literature. Following that section is the rationale and foundation for the similarity measures, a demonstrat ion of how they are calculated, and a discussion of observed trends. A description of the design and results of the empirical tests follows while the next section examines these results for their sensitivity to changes in the similarity measures. The final section contains the conclusion. Hypotheses and Prior Literature Auditor Client Compatibility and Selection of Auditor To make predictions about the choice of an auditor based on auditor client comp atibility requires two conditions: (1) variation in client preferences regarding the audit and auditor, and (2) variation in auditors abilities to satisfy those preferences. If auditors are all essentially equivalent (no auditor variation), clients would randomly choose between them. On the other hand, if auditors vary but clients all have the same preferences, then all the clients would strictly prefer the one auditor that best suited their uniform preferences, which is a condition not observed in the U .S. audit market. Therefore, the only situation which would lead to predictable patterns in auditor choice is the one in which both clients and auditors have variation within their respective groups.
18 The literature has documented substantial evidence of variation in client preferences and auditor capability. For example, a large, multinational client is more likely to choose a BigN auditor (Chaney et al. 2004) at least in part because a smaller auditor does not have the resources and cap ability of auditing such a company. A large number of studies have also focused on industry specialization as a differentiator of both demand and supply. Industry specialists are those auditors that have invested significantly in developing expertise in auditing a particular industry, typically proxied for by having higher market share in that industry. Specialists are typically better at detecting errors (Owhoso et al. 2002) and are associated with clients having higher earnings response coefficients (Balsam et al. 2003; Gul et al. 2009) and lower discretionary accruals (Krishnan 2003) They are also better at improving audit quality through knowledge spillover from nonaudit services they provide (Lim and Tan 2008) Beyond variation in quality, there are also cost structure differences among specialists. While most studies have found specialists charge higher audit fees (Gramling and Stone 2001) there is also the possibility of cost savings through the same expertise (Cahan et al. 2008; Craswell et al. 1995; Willenborg 2002) Audits by industry specialists also tend to be more efficient (Cairney and Young 2006) Client preferences for certain quality levels and cost structures will lead them to cho ose an auditor with structural characteristics that best meet their needs.3 While most prior literature has focused on broad categories of auditor (e.g., Big4 or nonBig4, specialist or nonspecialist), some studies have examined client 3 Inves tor preference can also play a role in the decision. S witches to larger auditors and specialist auditors are associated with positive market reactions (Fried and Schiff 1981; Knechel et al. 2007; Nichols and Smith 1983)
19 motivations for choosing a specific auditor. For instance, Lennox and Park (2007) find that a company is more likely to engage a particular auditor when a former employee of that auditor is now on the management team of the client. Research on client auditor disagreements about accounting treatments finds that a client is more likely to switch auditors in the presence of more conservative accounting treatments, presumably in an effort to find an auditor who is more amenable to the companys preferences (DeFond and Subramanyam 1998; Krishnan 1994) Bamber and Iyer (2007) find that auditors who more strongly identify with the client will be more likely to allow the desired accounting treatment. While this opinion shopping may sound disreputable, Dye (1991) shows t hat the firm may simply be trying to better communicate its internal information to the market. The broad conclusion is that interpersonal relationships and opinion shopping are just two ways a client can find one auditor to be a more compelling alternativ e than others. Given sufficient variation among clients and auditors, each of them will attempt to maximize their respective utilities by choosing a counterparty that best matches its preferences. While auditors can appear very similar on the surface, ther e are likely to be subtle differences that make one auditor a better fit than another. Considering all a clients needs and preferences, one specific auditor will likely provide a higher net benefit to the client.4 Therefore, I predict in alternative form: H1: Clients will tend to be with an auditor having a higher degree of auditor client compatibility than one with a poorer compatibility 4 The decision is also subject to various constraints. For example, CocaCola is unlikely to choose the same auditor as PepsiCo due to competitive concerns. I do not specifically address this issue in this paper, but the effect will be to shift a client away from its apparent best fit, thus working against my findings.
20 Auditor Client Compatibility and Auditor Switching Under H1, clients on average will prefer higher compatibility with their auditors; I now examine the outcome when the fit is suboptimal. Johnson and Lys (1990) demonstrate companies are more likely to switch auditors as the clients operating, investing, and financing activities change over time. They interpret this higher likelihood of switchi ng as a rational, efficient response to temporal changes in the companys audit preferences. In effect, the auditor client compatibility that was utility maximizing in the past has shifted such that another auditor may now be a better fit. Shu (2000) examines auditor client fit based on whether the client is with a BigN auditor when an empirical model would predict a nonBigN auditor, or vice versa, finding clients are more likely to change au ditors when there is a mismatch between the two. I extend this concept to examine whether such mismatches with a specific auditor are more likely to lead to auditor switches. Based on the degree of fit between the auditor and client, I expect that poor com patibility is more likely to lead to an auditor switch than better compatibility Therefore, I predict in alternative form: H2: Clients having relatively poorer compatibility with their current auditor will be more likely to change auditors. Once a client has made the decision to change auditors, it will need to choose from the remaining available auditors. For example, if the company always limits itself to Big4 auditors, a maximum of three auditors remain, subject to further constraints such as the auditor choices of close competitors and nonaudit service providers. Combining the first two hypotheses leads to the expectation that, after deciding the benefits of changing auditors will exceed the switching costs, a company will generally prefer a
21 new auditor that has a better compatibility from among the remaining options.5 Therefore, in alternative form, I expect: H3: A client switching auditors will tend to choose a new auditor that has a relatively higher degree of compatibility among the nonincumbent au ditors. Effect of Changes in Auditor Client Compatibility If compatibility affects audit quality and cost, then the degree of auditor client compatibility should lead to changes in observable audit process outcomes. Hammersley (2006) shows matched specialists defined as those operating within their industry of expertiseare more likely to process experimental cues regarding misstatements than mismatched specialists. Low (2004) shows that a similar industry based mismatch affects the audit planning and risk assessments occurring before the audit even begins. These experimental studies imply that audit quality will be higher when there is a better fit between the auditor and client. Johnstone, Li, and Luo (2011) find that clients within the sa me supply chain a measure of relatedness, if not similarity tend to have higher audit quality. On the other hand, a sizable literature finds decreased audit quality when the auditor more closely identifies with the client. In essence, the auditor client co mpatibility has become so good that the auditors independence is compromised. Lennox (2005) shows that companies are more likely to receive a clean audit opinion if the auditor becomes affiliated with the client by hiring its former auditors. Menon and Williams (2004) find clients employing former audit partners in executive positions tend to report higher levels of discretionary accruals. Based on existing evidence about the relation 5 I do not predict that the new auditor will necessarily have a better fit than the previous auditor, primaril y because of the endogenous nature of compatibility a client may appear to become a better fit for an auditor over time as the auditors preferences modify the clients observable financial statements and related disclosures.
22 between compatibility and audit quality, I expect that switching to an auditor with a better fit will change audit quality in the years following such a change. However, I do not make a signed prediction. Therefore, in alternative form: H4: Audit quality is associated with auditor client compatibility Measurement of Auditor Client Compatibility When clients are closely related to one another, auditors have the opportunity to specialize in those companies for both reputational and audit production reasons.6 While prior literature typically uses an all or nothing industry membership test to organize clients into similar groups, I introduce continuous proxies for the degree of compatibility between a company and an auditors existing c lient base. I base these measures on two existing streams of academic research that study inter entity similarity, the two being differentiated by the nature of the underlying data. The first stream is cluster analysis, along with related fields such as factor analysis, and is primarily concerned with grouping similar entities together based on a small set of numeric data (e.g., financial statement accounts). The second stream is the information retrieval literature, which uses documents as the underlying data (e.g., narrative disclosures). I develop client year measures of similarity for both financial statement and narrative textual disclosures since they represent different, but complimentary, signals. While the accounting systems and underlying economics are not separately observable, their joint effect is reflected in the financial statements. The narrative disclosures also provide economic and accounting system information, but contain additional information about managements interpretation of past events and expectations about the future. 6 Gramling and Stone (2001) s ummarize the industry specialization literature, which generally finds differences in both quality and audit fees for industry specialists versus nonspecialists.
23 The narrative disclosures are especially flexible, giving the opportunity for management to communicate more firm specific information or otherwise influence the markets view of the company. These two signals are eac h necessary to better understand the other and, taken together, provide a more comprehensive view of how the clients of an auditor relate to one another. Financial Statement Similarity Algorithms used in cluster analysis of numeric data include the Euclidean distance, city block distance, Chebychev distance, and Mahalanobis distance (Hair et al. 2006) Of these, the Mahalanobis distancesquared ( D2) is particularly sophistic ated in its ability to weight each variable equally according to its individual scale, as well as account for covariances among the various components. Introduced in M ahalanobis (1936) the D2 statistic imposes few restrictions on the underlying variables, only requiring nondegenerate distributions. After scaling and acc ounting for covariance of the variables, an observations distance from the group is larger when the variables for the observation are jointly more unusual. The D2 measure is the generally preferred algorithm in cluster analysis, when available (Hair et al. 2006) Outside of the accounting literature, D2 has been used in management to compare the distance between countries along multiple dimensions, including economic, fin ancial, political, cultural, demographic, and geographic location (Berry et al. 2010) Climatologists have used the measure to look for boundaries between different regional climates (Mimm ack et al. 2001) And chemists have used it for multivariate calibration, pattern recognition and process control (De Maesschalck et al. 2000) Prior accounting literature has used the D2 measure to a limited extent. Rege (1984) uses it to test the effectiveness of a discriminant function in classifying data into
24 two groups of likely and unlikely takeover targets. If the distanc e between the two groups is significant then the discriminant function is considered effective. Iyer (1998) also employs the statistic in a discriminant function context to maximize the distance between subgroups. Guilding and McManus (2002) use the measure to test for potentially influential outlying observations. However, all these studies use the measure in a statistical context and do not examine the properties of the distance itself.7 Due to the D2 measures power and flexibility, I use it as my primary measure of financial statement similarity. Because there is limited theoretical guidance on which variables might be appropriate for determining financial statement similarity, I use financial variables having well known significance in an audit context. Client size, complexity, and risk are important aspects of the audit (Hay et al. 2006) I include company size as the log of total assets ( SIZE ) to capture the scope and potential importance of the client. The combination of inventory and receivables proxies for audit risk inherent to the company ( I RISK ). Total accruals is a known audit risk factor and possible indic ator of audit quality, so TACC is calculated as in Defond and Subramanyam (1998) Cash and equivalents ( CASH ) is a proxy for liquidity, while return on assets ( ROA ) measures profitability; both are indicators of higher client stability and lower risk. Al l measures except SIZE are scaled by total assets.8 7 The abnormal accrual model could also be conceptualized as a distance measure. It empirical ly models the relationship between total accruals and various explanatory variables to attempt detection of accrual levels that are unusual looking relative to peer companies. 8 Other potential inputs, various permutations of those inputs, and alternative measurement approaches are described in the robustness section.
25 To calculate the D2 distance for my sample, I gather these five variables from Compustat, only using observations having assets greater than $1 million, with no fiscal year end change, and not in the util ities or financial services industries. The sample begins in 1997, when EDGAR data is first widely available for my narrative similarity scores, and ends in 2009. I include former Arthur Andersen clients only after they have not engaged Andersen for at least one year to limit the potential confounding effect of this onetime event. The concept of similarity is with respect to some group of other objects and is undefined for a single observation on its own. I call these other observations the reference group, which I define as other clients in the same auditor industry year. 9 I exclude any reference groups that do not have at least five observations; the similarity score is unlikely to be reliable if there are too few observations in the group. Because the r eference groups are rarely large enough for nonBig4 auditors, I explicitly limit the sample to Big 4 clients. Finally, I do not allow companies in the reference group in the year that they switch auditors. Each observation in the sample has n = 5 financial statement variables, which are contained in the transposed vector, xT = (x1, x2, xn) = (SIZE, IRISK, TACC, CASH, ROA) An observation is contained in an auditor industry year group having mean 9 Hogan and Jeter (1999) document the increasing importance of auditor restructuring along industry lines at the national level to take better advantage of internal teams of experts. While cli ent compatibility can also vary at the auditor office level, there are few offices with enough clients to calculate my similarity measures within an industry. Calculating the scores at the officeindustry level would lead to a 63% reduction in sample size (and a 30% reduction if calculated at the officesector level).
26 values of the same variables, T 12n) Finally, the group has a covariance matrix, for the five inputs. The Mahalanobis distancesquared is then calculated as:10 The scores are calculated within a GICS industry due to the general lack of comparability across industries I generate the ( x ) portion of the measure by subtracting the auditor industry year mean for each variable from the variables for the company year level to account for different sc ales and covariances across industries and over time that are unlikely to vary significantly between auditors.11 Because D2 is a measure of dissimilarity, I convert it to a similarity measure by taking the inverse. The natural log reduces skewness and outli ers: Narrative Disclosure Similarity The information retrieval literature has developed numerous methods for measuring the similarity of two documents, often in the context of matching a users Internet search query to the closest applicable web pages (Singhal 2001) Assuming the ability to map a document into a numeric representation, the D2 measure is also conceptually possible in a document context. However, practical considerations limit its 10 The Euclidian distance between observations is a special case of the Mahalanobis distance. If the covarianc 2 simplifies to the familiar [(x T(x ], which is the Pythagorean theorem if the vector has length two. 11 Multicollinearity is not problematic as it would be in a regression. However, as the vari ables approach nearly perfect multicollinearity, the covariance matrix will not be invertible, which can be a concern when using a small vector of variables in a small industry. While unusual in my sample, I exclude industry years that do not have at least ten company year observations (twice the number of variables). This restriction is almost always met given the earlier restriction of at least five observations with an auditor industry year.
27 usefulness. When mapping a list of words from narrative text into variable vectors, the high dimensionality makes calculation of the D2 measure impractical.12 Given these computational challenges, I instead use the Vector Space Model (VSM) from the document retrieval literature that can better accommodate large sets of long text. The VSM maps a document into a numeric vector (Salton et al. 1975) There are numerous ways to calculate the similarity of these document vectors. The most common approach is to calculate the cosine of the angle between any two vectors (Singhal 2001) an approach used in the accounting literature by Brown and Tucker (2011) They use the VSM cosine statistic to measure year over year dissimilarities in MD&A as a prox y for changes in narrative disclosure. Because they are interested in the differences between just two documents at a time, they only calculate pairwise similarity scores. In contrast, I aggregate these pairwise scores to get a measure of the similarity be tween one narrative disclosure and the disclosures issued by a reference group of clients within the same auditor industry year. Due to its relative ubiquity in information retrieval and its presence in existing accounting literature, I use the VSM based c osine similarity score as my measure of narrative disclosure similarity.13 To calculate the similarity of clients along a narrative disclosure dimension, I use important items from the annual report to ensure the disclosures are reviewed by the 12 For example, a covariance matrix using the 98,519 unique words for the MD&A in my sample would contain 9.7 billion elements. The matrix would then need to be inverted. The large size of the matrix arises not primarily from the number of documents, but from the unique words used in those documents. While the two are positively correlated, calculating the D2 measure on a subset of documents would not generally address the computational difficulties. 13 While I cannot feasibly calculate the D2 statistic for documents, I can do the reverse and calculate the VSM similarity for financial statements. Using such an approach, Jaffe (1986) uses vectors of different categories of patent applications to examine R&D spending overlap within industries I avoid this approach because the VSM does not account for variances and covariances of the variable components, which reduces its statistical power.
28 auditor, not voluntary, and considered important by the capital markets and regulators. Within the 10K, the longest disclosure items tend to be the business description, Managements Discussion & Analysis (MD&A), and the financial statement footnotes. Excluding exhibits, there are an average of 6,338 words in the business description, 7,054 in the MD&A, and 8,623 in the footnotes (see Table B 1), comprising 17, 18, and 21 percent, respectively, of the length of the typical 10K.14 I use multiple 10K items because ther e is considerable variation in the characteristics of each disclosure along several dimensions: (1) subject matter, (2) timehorizon, and (3) audit requirements. First, the topics discussed in each item are different. Item 101 of Regulation S K requires the 10K item 1 contain a detailed narrative description of the business, including industrial and geographic segments, principal products and services, R&D spending, and competitive conditions. Item 303(a) requires that the MD&A contain a discussion of liquidity, capital resources, results of operations, off balance sheet arrangements, and contractual obligations. Even though there is some topical overlap, the footnote content is typically determined by GAAP. Second, the MD&A is intended to be an interpretat ion of past and future operations through the eyes of management (SEC 2003) Given certain conditions, any forwardlooking statements receive Safe Harbor protection ( Item 303(c) of Regulation S K ). In contrast, regulations do not broadly require footnotes to explicitly contain interpretive or forward looking statements. Third, the footnotes are audited, while the business 14 The footnotes and MD&A seem particularly important to stakeholders, given the large number of accounting standards requiring or encouraging specific footnote disclosures and the relatively frequent guidance by t he SEC on MD&A (e.g., S EC 1987; 1989; 2003) Prior studies have demonstrated the usefulness of footnotes (e.g., Shevlin 1991; Amir 1993; Wahlen 1994; Riedl and Srinivasan 2010) Other research has shown some of the potential information contained in MD&A (e.g., Feldman et al. 2010; Feng Li 2010; Sun 2010) Of the three narrative disclosures, the business description is relatively unexplored except in studies of the full 10K as a single document (e.g., Li 2008).
29 description and MD&A are not audited but are only reviewed for material mis statements and consistency with facts known to the auditor (AU Sections 550; 551). While all narrative disclosures are somewhat flexible, the lack of an explicit audit of the business description and MD&A gives management the greatest flexibility to choose the topics and quality of discussion. Because all three have distinct characteristics, I use each as a separate source of narrative disclosure. For the narrative disclosure sample, I use 10Ks filed electronically via the SECs EDGAR system for fiscal ye ars 1997 through 2009. As in the financial statement sample, the disclosures in the text samples are by Big4 clients having at least five other observations available for comparison within the same auditor industry year reference group. Appendix A describes the selection and extraction process, which yields 33,355 business description, 31,280 MD&A, and 14,439 footnote observations. Treating the three narrative disclosure items of the annual report as separate data sets, I calculate the similarity score for each using an extension of the approach in Brown and Tucker (2011) that allows for a comparison of a company to its peers. The process, summarized in Appendix B, produces three variables SIMBUS, SIMMD&A, and SIMNOTESthat proxy for the degree of similarity between one client and other clients in the same auditor industry year. Higher similarity scores correspond to greater auditor client compatibility Patterns in Client Similarity Panel A of Table 2 1 contains descriptive statistics for the financial state ment and narrative disclosure similarity measures. Higher similarity scores indicate greater
30 similarity in relation to the reference group.15 SIMFS is always negative because of the log transformation. The sample contains the largest number of observations for this similarity measure since the financial statement components used to construct SIMFS are available for nearly all company years in Compustat. The narrative disclosure similarities ( SIMBUS, SIMMD&A, and SIMNOTES) are approximately centered around zero.16 Similarity scores are significantly higher for companies in the top quartile of size than for those in the bottom quartile (untabulated), indicating that bigger clients tend to be at the core of the auditors portfolio in terms of their similarity. The four similarity scores are not directly comparable to one another because of variations in how they are calculated (e.g., a score of 0.20 for MD&A is not necessarily larger than a score of 0.15 for footnotes). To compare across measures, I standardize each of them to have a mean of zero and standard deviation of one. In Figure 2 1, I plot these values against auditor tenure the number of years with the current auditor. A tenure value of zero corresponds to the year before an auditor change and a value o f one is the first year after a switch. For visual comparability, all scores are adjusted to begin at zero when auditor tenure is zero. The graph ends in year ten of the audit engagement since a decreasing number of observations after this point leads to h eightened volatility in the graph. All similarity measures increase over the length of the auditor client relationship, indicating auditor client compatibility improves over time. Auditors might find this trend 15 The average auditor industry year reference group size is 38 clients for financial statements, 23 for the business description, 22 for MD&A, and 12 for the footnotes. 16 As noted in Appendix B, I maximize the sample size for making the length adjustment by using all available observations, including nonBig4 auditors. After restrict ing the study sample to only Big4 clients, the mean is slightly above zero.
31 beneficial to the extent that it improves the quality or reduces the effort involved in the audit engagement. Likewise, clients might benefit from adopting best practices arising from the auditors expertise developed in similar client engagements. The business description experiences a rapid increas e in similarity during the first two years of the engagement, after which point the similarity becomes more stable. MD&A similarity also increases quickly in the first two years and then rises more slowly until approximately year seven. The trends in financial statement and footnote similarities are highly correlated, which is not surprising since those disclosures are intended to be closely aligned by regulation. Both of these measures increase gradually over time, with no sudden jump in the early years of the relationship. To test for the statistical significance of these trends, I compare the means of similarity for engagements with a short tenure (less than four years) to those with a long tenure (nine or more years). In all cases, the similarity scores are significantly higher for longer tenure clients than for shorter ones. I perform a related test using year over year changes in similarity and find that the annual changes for long tenure clients are not as large as the changes seen in newer clients.17 T hese patterns all demonstrate that auditor client compatibility is not just a static component of the relationship, but at least partially a function of the length of auditor tenure. Validation of Similarity Measures Table 2 2 shows the Pearson pairwise correlations among the similarity measures. I limit the correlations to those observations with scores available for all four disclosures 17 The t statistics for the difference in means of the financial statements, business description, MD&A, and footnotes are: 24.10, 4.61, 5.13, and 8.72, respectively. The correspond ing t statistics for the difference in changes are: 6.55, 2.94, 2.29, and 1.63.
32 (narrative and financial), although the unrestricted correlations are similar. The four similarity scores are all positi vely correlated with one another, indicating they measure related constructs. The correlations are higher among the three narrative disclosures (ranging from 0.61 to 0.72) than they are with the financial statements (from 0.05 to 0.11). As a means of validation, the table also contains the correlations between the similarity measures and various proxies for client differences which I expect to be negative. For each variable used to produce SIMFS ( SIZE IRISK TACC CASH and ROA ), I calculate the clients absolute difference from the mean for the auditor client year, calling them | SIZEDIFF |, |IRISKDIFF |, | TACCDIFF |, |CASH DIFF |, and |ROA DIFF |.18 The correlations with SIMFS are all significant, ranging from 0.25 to 0.42. More importantly, the significant cor relations between these difference variables and all the narrative disclosure scores are also negative, indicating the proper functioning of the narrative scores even though they did not explicitly include any financial statement variables. Of all the diff erence variables, | ROA DIFF | and | TACCDIFF | have the highest negative correlation with SIMFS, so client profitability and accruals appear to be important determinants of financial statement similarity. On the other hand, | IRISKDIFF | has the most negative correlation with the three narrative similarities, consistent with the importance of the numerous risk related disclosures in annual report items (e.g., Kravet and Musl u 2010; Campbell et al. 2010) 18 These difference variables should be negatively correlated with SIMFS by design, but these correlations document the SIMFS measure is working as expected. Performing a series of univariate correlation tests is also substantially different from the joint difference measure produced using the D2 technique.
33 For an alternative measure of client difference, I use the absolute value of unexpected discretionary accruals. Following DeFond and Jiambalvo (1994) I estimate accruals with a cross sectional modified Jones model run within SIC 2digit industries. The residual from thi s model is DACC, the unexpected discretionary accruals. As expected, all the similarity scores are significantly negatively correlated with the absolute value of DACC As a final validation, I expect that auditor client compatibility will not change dramat ically over short time periods because of the general stability in financial statements, related disclosures, and client portfolios. Large changes in the similarity measure from year to year are unlikely if the similarity measure is capturing the desired construct. In untabulated analysis, I calculate the autocorrelation coefficient for SIMFS (0.55), SIMBUS (0.93), SIMMD&A (0.92), and SIMNOTES (0.92), demonstrating a high degree of timeseries stability in all four measures. Analysis of Auditor Client Align ment Sample For the hypotheses tests, I collect additional data from Compustat for each observation with at least one similarity score available. The other variables, summarized in Table 2 1, begin in 1997, corresponding with the earliest availability of t he narrative disclosure data, and end in 2009. Auditor tenure is calculated based on the current auditor information in Compustat beginning in 1974. Typical Auditor Client Alignment If clients and auditors randomly choose to enter into an audit engagement without regard to auditor client alignment, one would expect approximately a 25% probability that a client would be with each of the Big4 auditors. According to the first hypothesis, a
34 client is more likely to be with an auditor having other clients similar to itself and are least likely to be with an auditor having a less similar set of clients. Table 2 3, Panel A contains a summary of auditor client alignment using each of the similarity scores. For financial statements, more clients are with the auditor having the best fit (26%) than with the worst fit (24%). When examining the business description, 27% of clients are with the auditor they are most aligned with, but only 23% are with the auditor with which they are least aligned. The MD&A pattern is even stronger, with 30% of clients being with the most aligned auditor and just 21% with the least aligned auditor. The strongest pattern occurs when using the footnotes, where 32% of clients are using the most similar auditor and only 20% are with the least si milar auditor. In each case, the probabilities monotonically decline as auditor client alignment decreases. To evaluate the statistical significance of these patterns, I compare the average auditor client alignment rank with the expected rank under the nul l of a random distribution.19 Since the ranks range from one to four, the null hypothesis would predict an average rank of 2.5.20 Comparing the average rank of each score to the null of 2.5 gives a test of the tendency of clients to be with a more closely al igned auditor than with a less aligned one, without requiring them to necessarily be with the most similar auditor. Table 2 3, Panel B shows the average rank of the incumbent auditor based on auditor client alignment. The average ranks are 2.47 (t = 5.67) when using the financial 19 An alternate approach would be to model the alignment rank using an ordered logit/probit, which would also allow me to control for other factors. Unfortunately, I am unaware of any existing model in the literature that would be useful for this purpose. Two streams of literaturethe choice of an industry specialist auditor and the choice of a BigN/NonBigN auditor would seemingly be the most relevant for developing such a model. However, modeling the rank using a wide variety of variables f rom these literatures did not yield a model with a statistically significant fit. 20 The expected average rank under the null is: (1 + 2 + 3 + 4) / 4 = 2.5
35 statements, 2.43 (t = 10.86) for the business description, 2.36 (t = 21.42) for MD&A, and 2.29 (t = 19.56) for the footnotes, all of which are significantly less than 2.5. The overall conclusion is that, based on the contents of their disclosures, clients are significantly more likely to be with better fitting auditors than with poorer fitting ones. Likelihood of Auditor Change Although clients tend to be with better aligned auditors, the simple descriptive analysis in the previous section does not address whether this pattern occurs because clients and auditors jointly choose an engagement that already has a higher alignment, or whether clients merely become more similar to their auditor over time as a side effect of the audit process. Therefore, in this section and the next, I undertake several analyses surrounding auditor switches to address these two possible explanations. First, I model the decision to change auditors, using important variables from the existing auditor switching literature, and then augment this model with my auditor client alignment measures. The basic logit model predicts a switch in the subsequent year using current year variables (firm and year subscripts are suppressed): ( 21) The dependent variable, SWITCH is an indicator set to one if the client will change auditors in the subsequent year; all other variables are measured in the current year. Because I expect larger firms to change auditors less frequently, I start by including the natural log of total as sets ( SIZE ) in the model. Following Landsman et al (2009) I include a variety of contr ols for audit and financial risk. As proxies for audit risk, I include growth, inherent risk, the nature of the audit opinion, and auditor tenure.
36 GROWTH in assets is associated with greater litigation risk for the auditor. Inherent audit risk ( IRISK ) is d efined as receivables plus inventories, scaled by total assets. MODOPIN is a dummy set to one for anything other than a clean opinion with no modifying language. I expect all these elements of audit risk to be negatively associated with the probability of an auditor switch. As in Landsman et al (2009), TENURE 10 is the number of years the client has engaged its current auditor, with a maximum value of 10 years. I predict this variable will be positively associated with auditor changes, both because of the g eneral stability of most auditor client relationships and the higher risk accompanying the early learning years of the engagement. To proxy for financial risk, I include both return on assets ( ROA ) and a dummy set to one when ROA is less than zero ( LOSS ). I expect that both of these controls will be positively associated with auditor switches. On the other hand, higher cash and equivalents scaled by total assets ( CASH ) proxies for the relative lack of financial risk for a client.21 Because M&A activity can l ead to an increased likelihood of changing auditors when the previously separate entities engaged different auditors, I include a dummy set equal to one when the current client has engaged in acquisition activity during the prior year that exceeds ten perc ent of total assets ( ACQUIS ).22 The correlations among these variables are presented in Panel A of Table 2 4. The correlations with the similarity scores imply larger ( SIZE ), more profitable ( ROA ), and less risky companies ( IRISK ) are more likely to have a better fit with their auditor. This 21 I omit the leverage variable in Landsman et al (2009) since it is not significant in their study or mine. I also leave out their measure of absolute discretionary accruals because this variable also proxies for the relative unusualness of the client relative to the industry, which is my construct of interest. However, including it does not change my conclusions. 22 I winsorize all controls other than dummies and logtransformed variables at the 1st and 99th percentiles.
37 pattern would arise if these types of companies are more likely to be able to engage their preferred auditor, a potential constraint for peers that may not be able to engage their first best choice if that auditor declines the engagement due to concerns about audit or financial risk. I augment E quation 2 1 with each of my proxies for auditor client fit as measured in the year before the switch and present the results in Table 2 4, Panel B. The controls are generally consist ent with prior literature. All similarity measures except for SIMBUS are negatively related to auditor switches. SIMFS (t = 4.72), SIMMD&A (t = 3.40) and SIMNOTES (t = 3.78) have negative coefficients, supporting the prediction in the second hypothesis that clients having a poorer fit with their auditor are more likely to switch to a new auditor. Holding all the other variables at their means, an interquartile decrease in financial statement similarity is associated with a 9.2 percent higher probability of switching. The increase for MD&A is 11.8 percent and for footnotes is 18.9 percent. This result explains one of the mechanisms through which the patterns in Table 2 3 may occur: clients tend to be with a better fitting auditor because they are more likely to change auditors when the fit is poor. Since the coefficient on SIMBUS is insignificant, the fit observed by examining the discussion of operating results and risks in the MD&A and the detailed accounting disclosures in the footnotes appear to be better predictors of an auditor switch than the general business description text. Auditor Choice Conditional on Decision to Switch Auditors Hypothesis 3 predicts company behavior following the decision to switch auditors, a situation in which the client has alr eady decided the net benefits of a change outweigh the switching costs. Table 2 3, Panel C, summarizes the average rank of the new auditor. Consistent with the earlier results for incumbent auditors, these clients tend to
38 choose a new auditor with better a uditor client fit. To test the statistical significance of this pattern, I compare the average rank of the new auditor to the prediction under the null.23 In Table 2 3, Panel D, the average rank of the financial statements is 1.96 (t = 1.91; p value = 0.028), the business description is 1.94 (t = 2.11; pvalue = 0.018), the MD&A is 1.86 (t = 4.85), and the footnotes is 1.89 (t = 2.19; pvalue = 0.015). Therefore, all similarity measures imply clients are significantly more likely to choose a better fitting a uditor when switching among the Big4. While these tests may seem to overlap with the earlier tests of typical auditor client fit, they provide insight into whether the commonly observed pattern is due to (1) a decision made by clients to choose an auditor with better fit or (2) clients becoming more aligned with their auditors as the auditors preferences gradually affect the clients disclosures over time. Consistent with the tenurerelated increase in auditor client fit demonstrated in Figure 2 1, fit cou ld be solely the result of the engagement rather than a causal factor. However, the patterns in Table 2 3 appear to support the first of these explanations: fit is relevant in the auditor selection process and is not merely an outcome of the audit process itself. Changes in Audit Quality Conditional on Auditor Client Fit To test whether audit quality is affected by auditor client fit, I compare the audit quality of engagements with better fit to those with poorer fit. There are multiple empirical proxies for audit quality; I start the analysis by examining discretionary accruals before considering the existence of Accounting and Auditing Enforcement Releases (AAER s) issued by the SEC. 23 A random choice among auditors would imply a rank of (1 + 2 + 3)/3 = 2.
39 Discretionary accruals Signed discretionary accruals are often used as a proxy for audit quality, with incomedecreasing accruals representing conservatism and incomeincreasing accruals corresponding to aggressive accounting practices. For such a proxy, I use discretionary accruals ( DACC) as defined in DeFond and Subramanyam ( 1998). Of the four similarity measures, SIMFS has a positive univariate correlation of 0.10 with DACC (untabulated), while the others are uncorrelated. I test for differences in discretionary accruals between clients having the best auditor client fit (i.e ., a rank of one in Table 2 3) and those having the worst fit. DACC is lower for better fitting auditors based on the business description (t = 1.50; p value = 0.067) and MD&A (t = 1.73; pvalue = 0.042). Discretionary accruals are insignificantly lower for the financial statement and footnote samples. When focusing on changes that occur in the first year after the switch (rather than levels), there is a significantly larger decrease in discretionary accruals for better fitting auditors than for less compat ible ones. However, this pattern only occurs for the financial statement measure, while the other measures are insignificant. Accounting and Auditing Enforcement Releases AAERs indicate the presence of a severe misstatement, and therefore represents extre me instances of poor audit quality. The advantage of AAERs is that a significant problem truly exists, which may not be the case with extreme levels of continuous proxies such as discretionary accruals. The disadvantage is a relatively small sample size g iven their severe nature. I rely on a highly developed logistic model of AAERs described by Dechow et al (2011) :
40 ( 22) The AAER data is a handcollected dataset provided by the authors of Dechow et al (2011) that covers all AAERs issued starting in 1982. Since the actual AAER is issued later and sometimes much later than the misstatement itself, I only consider fiscal year 2003 and earlier, which is the last year with at least 50 AAERs available. AAER is an indicator set to one if the SEC has issued an AAER that covers a particular fiscal year. The first four independent variables proxy for various aspects of accruals quality. RSST_ACC measures accruals as described in Richardson et al (2005) CH_REC and CH_INV are the change in accounts receivables and change in inventory, respectively, scaled by assets SOFT_ASSETS are scaled total assets after removing fixed assets and cash equivalents. The next two variables are performancerelated. CH_CS measures year over year change in cash sales and CH_ROA is the change in return on assets. Finally, ISSUE is an indicator variable set to one if the firm issued securities during the year.24 The pairwise correlations of these variables are in Table 2 5, Panel A. Using E quation 22 as a starting point, I add each of the auditor client fit measures in turn, with the results of these tests in Panel B of Table 2 5.25 Hypothesis 4 anticipates a crosssectional association between audit quality and auditor client fit, although it does not make a directional prediction. The coefficients on SIMFS (t = 2.83), SIMBUS (t = 24 All variables except ISSUE are winsorized at the 1st and 99th percentiles. 25 The controls for this test are consistent with, but weaker than, those in Dechow et al (2011), a condition that arises because my sample period begins after theirs.
41 2.40; p val ue = 0.016) and SIMMD&A (t = 2.77) are significantly positive, while SIMNOTES is not significant. These results generally support the conclusion that a company is more likely to receive an AAER as it becomes more similar to other clients of the auditor. Ra ther than implying higher audit quality arising from this similarity, the risk of major misstatements actually increases. Such a pattern could be explained by opinion shopping, where companies are looking for well fitting auditors because they are more lik ely to be able to use a preferred accounting treatment. An alternative explanation is that some audit productionrelated efficiencies arising due to increasing overlap among clients is actually decreasing audit quality. Robustness and Sensitivity Analyses Alternative Inputs for Financial Statement Similarity Because it is a new measure, I use a variety of alternative inputs for the financial statement similarity score to judge the sensitivity of SIMFS. Because large companies empirically occur less frequent ly than smaller companies, including SIZE as an input could cause the similarity measure to proxy for large clients rather than similarity more generally. I first remove SIZE from the input variables, leaving the four other inputs in place. As a second alt ernative, I include return volatility as an additional input to the original set to capture risk from a market perspective. Finally, I count the number of nonmissing/nonzero financial statement variables in Compustat as a measure of audit effort and comp lexity, since additional financial statement items are likely to increase the scope and intricacy of the audit.26 None of these changes make a difference in the 26 The number of reporting segments is frequently used to proxy for audit complexity, but is unavailable for many companies in Compustat. Counting the number of variables serves as a broadly available alternative.
42 qualitative results. Based on these modifications and others not reported here, the measure seem s quite stable and insensitive to the exact mix of input variables. Clients Switching from NonBig4 Auditors Conditional on switching auditors, Hypothesis 3 predicts clients will prefer a better fitting auditor. My primary test explicitly excludes clients switching from a non Big4 auditor. I now perform a similar test for clients switching from a nonBigN auditor to a Big4 auditor (i.e., upward switches), ignoring the forced changes by Arthur Andersen clients. Because these clients have four auditors from which to choose, the null would predict an average rank of 2.5. Only the footnote sample mean of 2.25 (t = 2.26; pvalue = 0.01) implies upward switchers are more likely to choose a better fitting auditor. None of the other measures are significant. Howev er, in all cases, the magnitudes of the average auditor ranks are similar to those observed in Panel C of Table 2 3. Overall, there may be some constraints on upward switchers ability to choose the best fitting auditor, but the lack of statistical signifi cance could be due to smaller sample sizes. I also examine former Arthur Andersen clients that switched to a Big4 firm following that auditors collapse. The patterns in this subsample seem the most random of the subsets. Based on my proxies of auditor fit none of the ranks are significantly different than what is expected under a random auditor selection (the MD&A measure is slightly significant, with a p value of 0.09). Because the sample sizes are not dramatically smaller than those in Table 2 3, the fo rmer Andersen clients appear unlikely to be with an auditor with greater compatibility. This result is not surprising because of the capacity constraints induced by the rapid auditor turnover affecting so many large clients at once.
43 Overall, the choices of clients switching among the Big4 most clearly support the idea that companies choose better fitting auditors when making a switch. The patterns for upwardswitching clients are similar, despite more limited formal significance. However, in the resourceco nstrained environment following Arthur Andersens collapse, its former clients seemed unable to choose auditors with high compatibility. Accounting System Comparability While I co nsider a broad notion of client similarity using multiple financial statement variables and narrative disclosure language, De Franco et al (2011) spec ifically examine the comparability of accounting systems between companies. For each company, they regress 16 quarters of earnings (an accounting system output) on returns (the net economic events) to estimate the accounting function for that company. To determine the similarity between any two observations, they use the fitted accounting function to predict earnings for each observation using actual returns. They interpret the difference between the two predicted earnings values as a measure of the diffe rence in accounting systems. Aggregating these differences for all pairs of observations gives a measure of accounting system similarity for each company within an industry year ( C OMPACCT IND). They construct an alternative measure using only earnings by r egressing 16 quarters of earnings of one company on the earnings of another. Aggregating the R2 from each regression also gives a proxy for accounting system similarity ( C OMPACCT R2 ). As a sensitivity test to my primary D2 metric, I calculate these two measures as described in more detail in De Franco et al (2011) as an alternative to SIMFS. The COMPACCT IND variable is uncorrelated with the four primary similarity scores I use in the current study. Nor is it correlated with most of my alternative proxies
44 for client differences (| SIZEDIFF|, |IRISKDIFF |, |TACCDIFF |, |CASHDIFF |, and DACC) with the exception of a 0.06 correlation with | ROADIFF |. In contrast, COMPACCT R2 has correlations of 0.0 7 0.0 8 and 0.04 with SIMBUS, SIMMD&A and SIMNOTES, respectively. As an alternati ve test of auditor changes I separately include the two accounting comparability measures in the base switch E quation 2 1 They are both negative but insignificant. When examining changes in audit quality, I find an increase in the probabilit y of receiving an AAER as COMPACCTIND increases, but only at the 5 percent significance level ; COMPACCTR2 is insignificant. Concluding Remarks I find strong evidence that auditor client compatibility helps predict which auditor a client will choose to engage. When the fit is poorer, clients are more likely to change auditors and choose a compatible auditor from among their remaining options. An interquartile shift in similarity with the current auditors client base can change the probability of switching auditors by as much as 19 percent. Descriptively, auditor client fit is a concave function of auditor tenure; compatibility with the incumbent auditor increases over time, but at a decreasing rate. Based on discretionary accruals, overall audit quality incr eases as auditor client compatibility increases. However, severe audit failures in the form of AAERs appear to increase as fit improves. The similarity measures I introduce have additional applications within accounting research. Chapter 3 of this w ork uses the same measures to look at the effect on audit fees when an auditors clients have different amounts of overlap in their audit processes. Beyond auditing, the measures can be used to isolate firm specific disclosures from disclosures that are similar among a set of companies. Econometrically, it is possible to implement a matchedpair design based on having
45 similar narrative disclosures.27 Another possibility is studying how disclosures propagate throughout an industry by monitoring when a speci fic disclosure becomes similar to other existing disclosures of peer firms. The similarity measures have potential usefulness in any context in which the relationships among a set of companies is of interest. My findings have several implications for both regulators and researchers. With the PCAOB continuing to consider requiring mandatory auditor rotation (PCAOB 2011) requiring an auditor change coul d force auditors and clients into less compatible engagements, which will potentially lead to changes in audit quality. On the other hand, the higher number of SEC enforcement actions as fit increases may imply fewer severe failures if clients switch to le sscompatible auditors. For researchers, prior literature has directly examined the effect of auditor type, such as auditor size and specialization, implicitly assuming auditors are indistinguishable within these groups (e.g., all BigN auditors are essenti ally the same). My findings indicate more heterogeneity among a particular category of auditors than previously thought. Therefore, depending on the nature of the research question, it may be worthwhile to consider the differential effects of specific audi t firms rather than examining them in broad categories. 27 The Mahalanobis measure is already used for this purpose. For example, the psmatch2 Stata library.
46 Figure 21 Trend in n ormalized similarity score over a uditor t enure This figure plots the proxies for auditor client compatibility on the Y axis against the number of years a client has engaged the incumbent auditor (auditor tenure) along the X axis. The normalized similarity score, used only in this graph, is standardized to have a mean of zero and standard deviation of one, and then adjusted to begin at zero in the first year of the engagemen t. The auditor switch occurs when tenure equals one; auditor tenure of zero indicates the year before the switch.
47 Table 21. Descriptive statistics Panel A: Similarity m easures and c omponents Variable Mean Std d ev 25% Median 75% N SIM FS (2.164) 1.23 5 (2.689) (1.978) (1.402) 57,035 SIM BUS 0.002 0.077 (0.053) (0.017) 0.040 33,355 SIM MD&A 0.002 0.087 (0.057) (0.024) 0.038 31,280 SIM NOTES 0.000 0.048 (0.029) (0.013) 0.011 14,439 SIZE 5.743 2.120 4.256 5.657 7.146 59,110 IRISK 0.249 0.195 0.088 0.213 0.367 58,553 TACC (0.066) 0.476 (0.097) (0.048) (0.004) 57,460 CASH 0.210 0.243 0.026 0.104 0.318 59,106 ROA (0.085) 0.365 (0.082) 0.022 0.069 59,067
48 Table 2 1. Continued. Panel B: Model v ariables Variable Mean Std d ev 25% Median 75% N SWITCH 0.070 54,1 49 GROWTH 0.310 1.228 0.057 0.058 0.241 58,799 MODOPIN 0.332 59,110 TENURE 8.847 7.868 3.000 6.000 12.000 59,110 LOSS 0.398 59,110 ACQUIS 0.125 59,110 AAER 0.008 59,110 RSST_ACC 0.034 0.302 0.059 0.026 0.114 53,446 CH_REC 0.012 0. 069 0.010 0.006 0.032 58,431 CH_INV 0.007 0.046 0.002 0.000 0.013 58,402 SOFT_ASSETS 0.502 0.260 0.290 0.523 0.714 59,048 CH_CS 0.220 1.112 0.052 0.080 0.259 54,434 CH_ROA 0.000 0.225 0.047 0.001 0.037 55,279 ISSUE 0.917 0.276 1.000 1.000 1.000 5 9,110 DACC 0.002 0.336 0.049 0.003 0.045 57,460 |DACC| 0.089 0.184 0.021 0.047 0.098 57,410 RETVOL 0.167 0.122 0.093 0.138 0.205 44,011 CSITEMS 150.226 34.960 125.000 146.000 173.000 59,110 Subscripts: FS = Financial Statements, BUS = Business Description, MD&A = Managements Discussion & Analysis, NOTES = Footnotes to Financial Statements. Panel A: SIM = similarity to other clients in the auditor industry year reference group. SIZE = log of total assets. IRISK = receivables plus inventory, scaled b y assets. TACC = total accruals. CASH = cash and equivalents, scaled by assets. ROA = income before extraordinary items, scaled by assets. Panel B: SWITCH = 1 if an auditor change in following year. GROWTH = change in assets, scaled by prior year assets. M ODOPIN = 1 for nonstandard opinion. TENURE = number of years with current auditor. LOSS = 1 if ROA < 0. ACQUIS = 1 if acquisition activity in current year exceeds 10% of assets. AAER = 1 if accounting misstatement in current year. RSST_ACC = accruals as i n Richardson et al (2005). CH_REC = change in receivables. CH_INV = change in inventory. SOFT_ASSETS = assets after removing fixed assets and cash. CH_CS = change in cash sales. CH_ROA = change in return on assets. ISSUE = 1 if securities issued during ye ar. DACC = discretionary accruals from cross sectional modified Jones model. RETVOL = return volatility. CSITEMS = # of nonmissing/nonzero variables in Compustat
49 Table 22. Correlations between similarity and d ifference m easures SIM FS SIM BUS SIM MD&A SIM NOTES SIM BUS 0.10 SIM MD&A 0.11 0.72 SIM NOTES 0.05 0.61 0.66 |SIZEDIFF| 0.25 0.00 0.06 0.02 |IRISKDIFF| 0.27 0.13 0.11 0.13 |TACCDIFF| 0.35 0.06 0.04 0.08 |CASHDIFF| 0.31 0.11 0.04 0.12 |ROADIFF| 0.42 0.08 0.06 0.08 |DA CC| 0.27 0.06 0.05 0.08 Correlations in bold are significant at the 1% level. Those within dashed box are expected to be negative. Based only on observations with valid values for all four similarity scores. |SIZEDIFF|, |IRISKDIFF|, |TACCDIFF|, |CASHDIFF|, |ROADIFF| = absolute difference of the variable from the mean of the auditor industry year reference group. Other variables defined in Table 2 1.
50 Table 23. Auditor s election based on a uditor client compatibility Panel A: Rank of incumbent auditor, based on similarity to each auditor's client base Fin s tmt Bus d esc MD&A Footnotes Freq % Cml Freq % Cml Freq % Cml Freq % Cml 1 (Sim) 13,870 26 26 8,027 27 27 8,352 30 30 3,427 32 32 2 13,833 25 51 7,695 26 5 3 7,041 25 55 2,850 27 59 3 13,622 25 76 7,352 25 77 6,710 24 79 2,331 22 80 4 (Diff) 12,961 24 100 6,751 23 100 5,802 21 100 2,099 20 100 Total o bs 54,286 29,825 27,905 10,707 Panel B: Average rank of incumbent auditor Avg r ank t stat Financial s tatements 2.47 *** 5.67 Business d escription 2.43 *** 10.86 MD&A 2.36 *** 21.42 Footnotes 2.29 *** 19.56 Panel C: Rank of new auditor following a Big4 to Big4 audit or change Fin s tmt Bus d esc MD&A Footnotes Freq % Cml Freq % Cml Freq % Cml Freq % Cml 1 (Sim) 557 36 36 285 36 36 314 41 41 110 41 41 2 510 33 68 266 34 70 245 32 73 79 29 70 3 (Diff) 495 32 100 2 37 30 100 205 27 100 80 30 100 Total o bs 1,562 788 764 269
51 Table 2 3 Continued. Panel D: Average rank of new auditor following a Big4 to Big4 auditor change Avg r ank t stat Financial s tatements 1.96 ** 1.91 Business d escription 1.94 ** 2.11 MD&A 1.86 *** 4.85 Footnotes 1.89 ** 1.99 Panel A: Freq = number of times an auditor of a given rank is engaged by a client. Rank 1 corresponds to the most compatibl e auditor while rank 4 indicates the most incompatible auditor. % = percentage of client years engaging that rank. Cml = cumulative total of the % column. Panel B: Avg Rank = average rank of the auditor engaged by a client. Random choice (null) is 2.5. Pan el C: Similar to Panel A, but only for clients changing from one Big4 auditor to another Big4 auditor in the year of the change. Panel D: Similar to Panel B, but only for clients changing auditors.
52 Table 24. Probability of a uditor c hange. Panel A: Pairwi se Pearson c orrelations SIM FS SIM BUS SIM MD&A SIM NOTES SIZE IRISK GROWTH TENURE10 ROA SIZE 0.22 0.26 0.10 0.19 IRISK 0.04 0.21 0.15 0.14 0.11 GROWTH 0.09 0.03 0.00 0.03 0.01 0.14 TENURE10 0.07 0.04 0.04 0.08 0.22 0.06 0.17 ROA 0.30 0.06 0.04 0.10 0.40 0.16 0.06 0.11 CASH 0.07 0.03 0.10 0.09 0.37 0.39 0.20 0.11 0.28
53 Table 2 4. Continued Panel B: Logit m odel of auditor switch in subsequent year Fin s tmt Bus d esc MD&A Footnotes Exp Coef z stat Coef z sta t Coef z stat Coef z stat (Intercept) 1.810 10.46 *** 0.681 2.54 ** 0.489 1.93 0.440 0.98 SIZE 0.248 18.87 *** 0.390 19.92 *** 0.390 19.40 *** 0.437 13.90 *** IRISK + 0.205 1.66 0.018 0.11 0.045 0.27 0. 217 0.86 GROWTH ? 0.074 3.75 *** 0.053 2.04 ** 0.040 1.59 0.073 1.81 MODOPIN + 0.419 10.14 *** 0.241 4.14 *** 0.243 4.10 *** 0.330 3.70 *** TENURE10 0.039 6.42 *** 0.032 3.81 *** 0.031 3.70 *** 0.024 1.89 ROA 0. 042 0.81 0.154 2.22 ** 0.163 2.32 ** 0.167 1.57 LOSS + 0.345 7.92 *** 0.437 7.38 *** 0.410 6.80 *** 0.361 3.84 *** CASH 0.596 5.76 *** 0.923 6.34 *** 0.975 6.62 *** 1.270 5.53 *** ACQUIS + 0.043 0.74 0.024 0.29 0.021 0.26 0.071 0.60 SIM FS 0.076 4.72 *** SIM BUS 0.430 0.99 SIM MD&A 1.324 3.40 *** SIM NOTES 4.922 3.78 *** Year FE Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Obs 5 2,232 29,778 27,761 12,850 Pseudo R 2 0.07 0.09 0.09 0.11 Panel A: Correlations in bold are significant at the 1% level. TENURE10 = same as TENURE but with a maximum value of 10 years, for compatibility with Landsman et al (2009). Other vari ables defined in Table 2 1. Panel B: The results of a logistic regression with dependent variable of SWITCH, indicating an auditor change in the following year. Variables defined in Table 2 1
54 Table 25. Probability of receiving an AAER Panel A: Pairwise Pearson correlations SIM FS SIM BUS SIM MD&A SIM NOTES RSST_ACC CH_REC CH_INV SOFT_ASSETS CH_CS RSST_ACC 0.07 0.01 0.04 0.01 CH_REC 0.00 0.01 0.03 0.03 0.31 CH_INV 0.01 0.00 0.02 0.01 0.21 0.32 SOFT_ASSETS 0.06 0.20 0.26 0.24 0.00 0.09 0.08 CH_CS 0.02 0.02 0.01 0.03 0.18 0.16 0.12 0.07 CH_ROA 0.10 0.01 0.02 0.00 0.32 0.13 0.07 0.04 0.12
55 Table 2 5. Continued. Panel B: Logit model of future AAER being issued for current year financial statements Fin s tmt Bus d esc MD&A F ootnotes Exp Coef z stat Coef z stat Coef z stat Coef z stat (Intercept) 6.767 14.9 *** 7.287 9.93 *** 6.492 12.08 *** 7.206 6.86 *** RSST_ACC + 0.424 1.97 ** 0.177 0.70 0.273 1.11 0.168 0.41 CH_REC + 1.692 2.21 ** 2.162 2.35 ** 2.630 2.78 *** 2.253 1.53 CH_INV + 1.979 1.89 1.822 1.45 1.180 0.89 1.298 0.63 SOFT_ASSETS + 2.007 7.96 *** 1.747 5.70 *** 1.638 5.25 *** 2.125 4.26 *** CH_CS + 0.031 0.6 0.015 0.23 0.024 0.38 0.082 0.63 CH_ROA 0.605 2.08 ** 0.456 1.40 0.603 1.88 0.190 0.38 ISSUE + 1.456 3.52 *** 2.014 2.83 *** 1.287 2.54 ** 1.716 1.70 SIM FS ? 0.149 2.83 *** SIM BUS ? 2.181 2.40 ** SIM MD&A ? 2.215 2.77 *** SIM NOTES ? 2.943 0.83 Year FE Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Obs 28,676 16,575 15,132 6,495 Pseudo R 2 0.04 0.03 0.03 0.04 Panel A: Correlations in bold are significant at the 1% level. Variables defined in Table 2 1. Panel B: The results of a logistic regression with dependent variable of AAER indicating an SEC enforcement action was released for the current year. Model based on Dechow et al (2011). Variables defined in Table 2 1.
56 CHAPTER 3 SPECI ALIZATION THROUGH CL IENT COMMONALITY AND ITS EFFECT ON AUDIT PRODUCTION COSTS Introduct ory Remarks Each engagement within an auditors portfolio has both idiosyncratic and nonidiosyncratic features deriving from the extent to which the audits have element s in common. A portion of each accounting disclosure of a client is due to the economic and accounting choices of that company, while other portions are the result of common factors such as auditor preferences, industry norms, macroeconomic conditions, and accounting standards. In this paper, I argue the nonidiosyncratic, overlapping components represent opportunities for the auditor to reduce production costs by improving audit technology and reliance on common knowledge spillover, which I refer to collec tively as specialization. I use the similarity of each client to other clients within the same auditor industry year as a proxy for potential opportunities to specialize in that group of companies. Following the approach in Chapter 2, I calculate the commonality between clients based on the similarities of both their financial statement and narrative disclosures contained within the annual report. Using two separate measures derived from different sources and based on different calculation approaches allows for a broader proxy of commonality than either would provide on its own. The financial statement similarity measure is based on the Mahalanobis distance, used in the cluster analysis literature to divide observations into groups based on numeric character istics of each observation.1 The narrative disclosure commonality measure uses the business description, MD&A, 1 I typically use the term similarity in this paper, although the context occasionally calls for the term distance. In the current context, distance is the conceptual inverse of similarity.
57 and footnote items contain in the mandatory annual report as a proxy for how similar the companys disclosure choices are to the choices of peer companies. In my first hypothesis, I argue the degree of commonality among clients of an auditor can affect audit production costs through its effect on both labor and audit technology. Commonality influences labor costs through knowledge spillovers between engagements and changes to the mix of more senior and less experienced labor. Production costs are also a function of audit technology, which is easier to implement and more effective when client overlap is greater. Given the opportunity for reduced production costs, I first predict that a client having more in common with its peer clients has lower audit fees. I find strong evidence of this association for both financial statement and narrative disclosure similarity. The effect is also economically signi ficant: an increase in similarity from the 25th to the 75th percentile is associated with a decrease in audit fees of 4.3 to 8.3 percent. My second hypothesis is that the relationship between client similarity and fees is stronger when the auditor has grea ter financial incentives to take advantage of overlap in its portfolio. The auditor is unlikely to make the necessary investments solely because of the opportunity to do so, but will also consider how economically meaningful the investment might be for overall profitability. In support of this hypothesis, I document an incrementally negative effect for financial statement similarity when an industry provides a higher percentage of the auditors revenues. Finally, having two primary measures for client commonality allows me to examine situations in which the two proxies are inconsistent in their portrayal of similarity. I consider two types of inconsistency between the financial statements and
58 corresponding narrative disclosures: pooled textual disclosures and differentiated textual disclosures. A pooled disclosure occurs when a company has unusual looking financial statements relative to its peers, but the accompanying textual disclosures do not reflect those financial differences. The accompanying text shoul d either reflect the atypical financial statements or explain why the differences are not a true representation of the companys situation. However, the text does not appear to do so, representing an incremental risk factor for the auditor and possibly eroding the production efficiencies predicted to be associated with greater client commonality. I proxy for each type of inconsistency by focusing on firms that are in opposing terciles of similarity for financial statements and narrative disclosures. As predicted, I find that pooled text disclosures are associated with higher audit fees than clients without such inconsistency. The second type of inconsistency a differentiated disclosure occurs when a company has fairly typical financial statements relative to its peers, but the textual disclosures seem to contain more uncommon, possibly firm specific, information. The prediction in this case is less clear than a pooled disclosure since differentiation can be the result of a client who is (1) unjustifiably tryi ng to differentiate itself from its peers or (2) attempting to provide additional, firm specific information that can be useful and risk reducing to both auditors and investors. In contrast to pooled disclosures, I generally find that differentiated disclosures are associated with lower audit fees than clients not having this type of inconsistency. My study makes several contributions to the literature. First, I provide empirical proxies of auditor specialization that have several advantages over existing m easures. The proxies are at the client level, rather than the auditor industry level, which allows a
59 more direct mapping into client level audit fees. This approach also allows for the existence of subgroups within an auditor industry, since auditors do no t necessarily orient their practices around the broad groups provided by third party industry classification systems. Because my proxies rely explicitly on client characteristics, I avoid the use of market share measures that are likely confounded by compe titive pricing strategies and other audit market features, making my measures easier to interpret as proxies for the underlying specialization construct. Although interpreted in an audit context for the current study, the proxies are general purpose measur es of overlap among companies, providing many potential applications outside of the audit setting. My second contribution is to the limited literature on the relation between auditors and clients narrative disclosures. Few audit related studies consider t he role of narrative disclosures in conveying information about the client,2 even as the PCAOB has recently proposed substantially increasing the role of the auditor in reviewing these communications (PCAOB 2011) In this study, I show the usefulness of narrative disclosures in examining the implicat ions of how clients of an auditor relate to one another. In a third contribution, an extensive literature has looked at the relationship between specific client financial statement elements and the audit, without a higher level understanding of what the br oader financial data mean for the auditors client portfolio. The measures I develop allow for a research design that simultaneously considers multiple dimensions of client commonality. I further combine the multiple disclosure channels used by the client to look for inconsistencies, which provides more 2 One exception is Dunn and Mayhew (2004) which finds that clients of industry specialists have higher quality narrative disclosures.
60 nuanced insights than those provided by studies examining only one disclosure mechanism. The rest of the paper proceeds as follows. The next section develops the hypotheses. The section after that describes the rationale and foundation for the similarity measures, followed by a section that explains the sample and calculations of the measures. A description of the results of the empirical tests follows The next section contains alternative similarity measures and other robustness tests, and the final section concludes Hypotheses and Prior Literature Production Costs and Audit Fees Simunic (1980) presents a widely used model of audits in which fees charged to clients are a function of production costs (effor t) and any expected losses due to potential audit failure (risk). Production costs primarily labor in an audit setting are composed of the quantity and unit cost of resources consumed to provide a given level of audit quality. For example, the size of t he client corresponds to a higher quantity of resources required; as such, client assets and sales are positively related to the quantity of labor hours expended (OKeefe, Simunic, et al. 1994) Since there is a nonzero probability that an audit will fail by not detecting or reporting a material financial statement error, the auditor must either charge a higher fee to insure against the possible loss or expend greater effort to reduce the risk. For instance, Hackenbrack and Knechel (1997) show that the labor mix shifts towards more senior, costly auditor employees when audit risks are higher. Overall, prior literature has documented a very
61 strong, positive relation between audit production costs (both effort and risk) and audit fees (Causholli et al. 2010) .3 Specialization and Audit Fees An extensive literature has examined the effect an auditors specialization in a gro up of clients has on audit fees. The group is typically implemented as some category of industry, leading to the customary term industry specialization. Studies in this area variously predict both a decrease and an increase in audit fees due to industry specialization, although the archival evidence has generally supported the latter. Industry specialists are expected to charge lower fees when nonidiosyncratic audit components lead to knowledge sharing and investments in overlapping audit technology that are associated with lower production costs due to having a more efficient and less risky audit. Earlier studies have occasionally acknowledged the possibility of this negative relationship (e.g., Craswell et al. 1995; Willenborg 2002) and some archival results support this prediction. For example, Mayhew and Wilkins (2003) find that auditors who have larger industry market share, but do not dominate the industry, charge lower fees to clients initially going public. Experimental evidence also lends credence to the potential for lower fees (e.g., Owhoso et al. 2002; Low 2004) However, most studies in the area proxy for specialization using industry market sh are and typically find higher fees for specialists (Gramling and Stone 2001) The general interpretation is that the same knowledge sharing and audit technology described under the negative prediction improve audit quality or auditor reputation (e.g., 3 Lower production costs do not necessarily lead to lower audit fees if the auditor is retaining the entire increase in profit margin. However, as long as the audit market is sufficiently competitive, at least some portion of these lower costs will be passed along to the c lient.
62 Ward et al. 1994) Since clients are presumably willing to pay more for higher actual or perceivedquality, specialization should be associated with higher audit fees. Given the two divergent predictions, the choice of proxy for specialization is especially critical. For example, the offsetting effects could lead to no discernable relationship (e.g., Palmrose 1986) On the other hand, if the proxy better captures the quality and reputational effects of specialization, a positive relation will domina te, as appears to be the case when using industry market share. Market sharebased proxies could also be measuring the competitive strategy of an auditor in the audit market for a particular industry rather than specialization per se (Numan and Willekens 2012) Minutti Meza (2011) a rgues that studies documenting a positive relation between industry specialists and audit quality are the result of uncontrolled client characteristics, and finds no improvement in audit quality for specialist clients once fully matching on these attributes. Gramling and Stone (2001) note the link between market share and specialization is typically vague and that existing research offers little justification for applying existing market share and market specialization measures as proxies for industry expertise (p. 14). In the current study, I develop measures that more directly proxy for having a production proces s specialized for a subset of clients so that I can better address the negative relation between specialization and fees. Opportunities to Lower Production Costs Commonality and idiosyncrasies among clients of an auditor can affect audit production costs t hrough their effect on both labor and audit technology. One effect on labor costs is that fewer idiosyncrasies will likely require less planning and oversight due to decreased risk and complexity, thus shifting the labor mix to lower level, less expensive personnel (Hackenbrack and Knechel 1997). There is also the potential for
63 knowledge overlap, which includes familiarity with certain types of clients, rules of thumb, and other relevant onthe job experience (e.g., Beck and Wu 2006) Research in organizational behavior has found that knowledge gained by performing job tasks is transferred within an organization (e.g., Darr et al. 1995) Experimental evidence suggests that specialist auditors are better at detecting errors (Owhoso et al. 2002) and assessing audit risk (Low 2004) However, archival auditing studies have not found strong empirical evidence to support learning by doing or learning over time (Causholli et al. 2010; Davis et al. 1993; OKeefe et al. 1994) possibly due to the specific proxies chosen. Increased client overlap could also affect the auditors ability to develop specialized audit technology. Audit technology is a set of fixed investments by an auditor in innovations such as customized workflow, employee training, specialized software, decision aids, and the formation of inhouse consulting groups (Dowling 2009; Sirois and Simunic 2010) .4 Higher quality audit technology is better at identifying and directing effort to problem areas of individual clients (Blokdijk et al. 2006, 29) A higher degree of client commonality could provide more input into the current audit. For example, analytical procedures have better predictive ability when based on similar peer firms (Minutti Meza 2010) These techniques are likely to be more accurate when based on a larger number of more similar reference clients. Cahan et al (2008) argue that homogenous investment opportunity sets among clients are a specific type of client overlap that creates such an opportunity to invest in audit technology. I extend this line of reasoning to examine client overlap in a more general sense. If there is greater client 4 Note that audit technology is not necessarily implemented using computerized systems.
64 overlap, there will be more common audit components to extract, and thus it will be less costly and more effective to develop common technologies based on those similarities.5 The lack of archival evidence notwithstanding, organizational theory and experimental studies suggest a greater ability to transfer knowledge within the audit firm will lead to lower audit risk and more efficient audits. Both of these outcomes will result in an audit with lower production costs, albeit with potentially higher audit quality. Therefore, I predict in alternative form: H1: Clients having higher overlap with other clients of the auditor pay lower audit fees. Incentives to Lower Production Costs The first hypothesis derives from the opportunities inherent in client overlap, but audit firms and individuals will only invest in additional audit technology and develop common knowledge when there are incentives to do so. Economic incentives are likely to be highest for those cli ents that are relatively more important to the auditors overall profitability. For example, an industry that provides audit fees that are higher than other industries might give the auditor greater incentives to develop audit technology appropriate for th at industry. In contrast, if an industry represents a very small portion of the fee portfolio, the auditor is less likely to make investments in technological improvements for that group of clients, even in the presence of strong opportunities. I expect gr eater incentives to develop specialized audit technology and knowledge will accentuate the relation between opportunities and fees predicted in the first hypothesis. Supporting the significance of stronger portfolio incentives, Knechel, Niemi, and Zerni 5 While some technology and knowledge can be broadly applied, such as audit standards and firm wide policies, I specifically focus on components that are relevant to subgroups of clients to provide adequate crosssectional variation.
65 (2012) find that partner specialization is associated with higher compensation for economically important sectors. Therefore, the next alternative hypothesis is: H2: The negative relation between client overlap and audit fees is stronger in industries that are economically important to the auditor. Inconsistent Signals of Commonality Given multiple consistent signals of the true underlying client overlap, the prior hypotheses make predictions about the relationship between commonality and audit fees. For the purposes of this study, I use the financial statements and the narrative disclosures in the annual report as two broad disclosure channels. Bamber and Cheon (1998) show cross sectional variation in managements choice of channels for disclosing earnings forecasts, along wi th differential investor reaction to those choices. Therefore, an incremental effect beyond the earlier predictions can arise if these disclosure channels are not in agreement with one another regarding the degree of underlying similarity. One type of disc losure inconsistency occurs when the quantitative financial statements seem to represent a company that is relatively unusual for the industry, but the accompanying qualitative narrative disclosures make the client appear very typical. If the financials ar e dissimilar, one would expect that the accompanying text would either reflect these differences or explain why the differences are not a true representation of managements view of the companys position. In either case, the narrative disclosures should appear different from other clients of the auditor. Narrative disclosures give the company greater flexibility and discretion than is usually available in the financial statements. Under this flexible regime, the company is apparently choosing to downplay t he differences in the underlying financials. I call this situation
66 pooled text inconsistency. Inappropriately differentiated disclosures could cause additional risk for the auditor or require more effort to attain the same level of assurance. But even if t he differences are justified, verifying the propriety of the claims will take additional effort by the auditor: H3a: Clients with dissimilar financial statements but similar narrative disclosures (pooled text) pay higher fees than other clients. A second type of inconsistency is when the financial statements indicate a client is relatively similar to other companies, but the narrative disclosures make the client appear more unusual. The client may be attempting to unjustifiably differentiate itself from o ther companies, as might occur before an upcoming equity offering. On the other hand, narrative differences could represent firm specific disclosures that improve the quality of information available about the company. For example, Tasker (1998) shows that managers will use a more flexible disclosure channel when the financial statements are relatively less informative. This improvement in the information environment represents a potentially positive situation for the auditor. I call this type of inconsistency the differentiated text condition. Because there are both beneficial and problematic potential reasons for differentiated narrative disclosures, it is an empirical question as to the relation between this type of inconsistency and audit fees. Stated in alternative form, my final hypothesis is: H3b: Clients with highly similar financial statements but dissimilar narrative disclosures (differentiated text) pay different fees than other clients. S ample Financial Statements As in Chapter 2, I use the Mahalanobis distancesquared ( D2) measure to calculate the commonality of financial statements among clients of an auditor. This
67 proxy is ideally suited to determining the similarity of small sets of variables. By selecting a limited number of key financial statement variables, the measure provides the aggregate distance of one companys financial statement variables from all the other financial statements in the same auditor industry year. To calculate the similarity between one observation and a set of appropriate peers, I define the reference group as the set of client years with the same auditor and industry. I exclude any reference groups that do not have at least five observations; the similarity sc ore is unlikely to be reliable if there are too few observations in the group. Because the reference groups are rarely large enough for nonBig4 auditors, I explicitly limit the sample to Big 4 clients. Finally, I do not allow companies in the reference gr oup in the year that they switch auditors. These restrictions leave 32,412 observations in my financial statement sample. Because there is no theoretical guidance on which variables are appropriate for the financial statement similarity measure, I use fina ncial statement variables having well established relationships in an audit context. Based on the empirical audit fee model components described in Hay et al (2006) I include proxies for audit effort, audit complexity, and client risk. To focus on the clients financial statement similarity, I avoid engagement or auditor specific variables and client related variables that are not included in the financial statements. As a distancebased measure, using unscaled variables would cause the D2 metric to be so heavily influenced by the size of the companies that it would effectively become a proxy for client size. Because client size typically explains a large portion of audit fees and large firms are more uncommon by definition than smaller firms, I do not directly include proxies for size and also scale all
68 variabl es to remove a direct size effect. Correlations with size are normally observed in financial data (e.g., between size and profitability), which ensure client size has an indirect effect on the measure without overwhelming other patterns in the data.6 I gat her the necessary financial statement variables from Compustat, only using observations with assets greater than $1 million, with no fiscal year end change, not in the financial or utility sectors, and having all data fields required to calculate the similarity scores. The sample begins in 2000, when audit fee data is first widely available, and ends in 2009. I count the number of nonmissing/nonzero financial statement variables in Compustat as a measure of audit effort and complexity ( CSITEMS), since additional financial statement items are likely to increase the scope and intricacy of the audit.7 I use long term debt to proxy for the risk due to the clients leverage ( LEV). The combination of inventory and receivables proxies for inherent audit risk ( I RI SK). Audit fee models usually include a measure of profitability, frequently some variant of income or a profit/loss dummy. Departing somewhat from prior literature, I include separate variables for revenues ( REV ) and operating expenses ( EXP) to give the i ncome statement roughly the same representation in the vector as the balance sheet.8 All measures are scaled by total assets. I regress the natural log of audit fees on these 6 Chapter 2 includes SIZE in the set of input variables. Size is excluded in the current paper because of its welldocumented, dominant effect on audit fees, which is the dependent variable in the current context. 7 The number of reporting segments is frequently used to proxy for audit complexity, but is unavailable for many companies in Compustat. Counting the number of variables serves as a broadly available alternative. To my knowledge, this variable has not been used before in the audit fee literature, but is potentially superior to existing alternatives. 8 In an untabulated robustness test, I u se income before extraordinary items ( INC ) in place of REV and EXP, with no change in the qualitative conclusions.
69 scaled variables to verify they are all highly significant in the expected direct ions and consistent with prior literature. Narrative Disclosures I use three important items from the mandatory annual report as separate sources of narrative disclosures with which I proxy for client commonality. I select the business description, MD&A, and footnotes from the annual report because of their relative importance and length within the context of the 10K. Each provides variation in topical coverage, time horizon, and the level of auditor assurance provided. Using three distinct disclosures provides insights beyond using either a single disclosure item or the annual report in its entirety. For example, to the extent that liquidity and results of operations is more strongly related to audit fees than product market competition, I would expect the MD&A similarity measure to have stronger results than the business description. While I make no specific predictions about which narratives have a stronger relationship with audit production costs, I leave the disclosures disaggregated to ensure I can obs erve differences across the various items. To proxy for the commonality among clients, I use the same Vector Space Model (VSM) procedure as in Chapter 2, which is an extension of the approach in Brown and Tucker (2011). The VSM maps documents into numeric vector representations, where each element of the vector is a weighted count of the number of times a particular word occurs in the document. Taking the dot product of any two document vectors yields the cosine of the angle between those vectors, a measure of similarity that ranges from zero (completely dissimilar) to one (identical documents). I calculate this dot product between the observation of interest and every other client in the same auditor industry year. I average the top five most similar client s (i.e., the five with the highest cosine
70 measures) and correct for mechanical biases using the procedure described in Appendix B of Chapter 2.9 For the narrative disclosure sample, I use 10Ks and 10K405s filed electronically via the SECs EDGAR system for fiscal years 2000 through 2009. As in the financial statement sample, the disclosures in the text samples are by Big4 clients having at least five other observations available for comparison within the same auditor industry year reference group. These filters yield 23,146 business description, 22,146 MD&A, and 10,666 footnote observations. There are fewer observations in the narrative disclosure samples than in the financial statement sample. This difference is primarily due to unavailable reports on E DGAR, items included by reference to other locations, and textual idiosyncrasies that lead to problems extracting the 10K items of interest. The substantial drop in the number of footnote observations, as compared to the business description and MD&A samples, is because many companies attach financial statements and footnotes as an exhibit to the report in a variety of unpredictable ways, making their automated extraction difficult. Treating the three narrative disclosure items of the annual report as separate data sets, I calculate the similarity score for each using the same approach described in Chapter 2. The process, summarized in Appendix B, produces three variables SIMBUS, SIMMD&A, and SIMNOTESthat proxy for the amount of commonality between firm i and its five closest peers in the same auditor industry year. Higher similarity scores correspond to greater commonality. 9 Chapter 2 averages all the scores within the auditor industry year, rather than the five most similar. Therefore, when correcting for the mechanical bias in the current chapter, I also include the first three powers of the number of clients in the auditor industry year. In the next draft, I plan to make these two chapters consistent by always using the full set of clients and only using the five most similar in a sensitivity test. Doing so does not change the qualitative results.
71 Analysis of Similarity Audit Fee Model The tests rely on the following base audit fee model developed from the audit fee meta analysis in Hay et al (2006) : I include controls for various c lient attributes, some of which are also used to calculate client similarity. The natural log of client assets is SIZE I proxy for client complexity by counting the number of nonzero/nonmissing items for that client year in Compustat ( CSITEMS). Inherent risk ( IRISK ) is receivables plus inventory, scaled by total assets. LOSS a dummy set to one for negative net income, proxies for financial weakness. Finally, leverage ( LEV) is long term debt scaled by total assets. I also control for auditor and engagement attributes. If the number of days between fiscal year end and the issuance of the 10K is more than 90 days, then DELAY is set to one as a proxy for audit complexity. The log of the dollar amount of nonaudit services is NAS.10 Clients with a December 31 fiscal year end date could lead to increased resource constraints, so BUSY is a dummy set to one for these companies. Audits leading to anything other than a standard opinion might be associated with additional audit effort or risk. Therefore, OPIN is a d ummy set to one for nonstandard opinions, almost always a clean opinion with modified language. I construct a similar measure for internal control, setting ICMW to one if the auditor has noted a material 10 I first add 1 to the nonaudit fees to avoid taking the log of 0.
72 weakness in internal control. TENURE is the number of years the client has been with the current auditor, according to Compustat. To ensure my measures capture a construct distinct from traditional proxies for industry specialization, I include INDSPEC a dummy set to one when the current auditor receives at least 32.5% of the total fees available within the clients GICS industry and year.11 Finally, I control for industry and year fixed effects. All controls are expected to be positive, except IND SPEC which is unpredicted. For these variables, Table 3 1 co ntains descriptive statistics in Panel A and correlations in Panel B. The patterns are consistent with prior audit fee literature. Hypothesis 1 To test the first hypothesis regarding client similarity and audit fees, I augment the base model with one or more of the similarity variables. H1 predicts the coefficients on these similarity variables will be negative. I begin by testing the model with SIMFS to assess the relationship between fees and financial statement similarity, with the results in Table 3 2 .12 The coefficient on SIMFS is significantly negative (t = 9.41), as predicted, so fees are lower as the financial statements of a client are more similar to other clients of its auditor. All control variables are significant and in the expected direction except for LEV, which is insignificant. While size, complexity, and risk are still important determinants of audit fees, it appears that financial statement overlap with other clients is also relevant. 11 Since the literature has not extensively explored GICS industries in a specialization context, I use the samples 75th percentile as a cutoff. I prefer GICS to SIC as the similarity scores are calculated using this categorization. Results are qualitatively unchanged when using a more typical 30% cutoff based on SIC 2 digit codes. 12 All standard errors are heteroscedasticity consistent using a Huber White adjustment.
73 I now turn to the three narrative disclosure similarity measures. For this test, I leave SIMFS in the model as a control for underlying economic similarity and then alternately test the coefficients on SIMBUS, SIMMD&A, and SIMNOTES. Each of the narrative coefficients is significantly negative (t = 9.98, t = 4.77, and t = 2.86, respectively). The results are qualitatively unchanged if SIMFS is excluded from these models. Once again, the control variables are as expected, with the exception of LEV. Overall, there is strong evidence that the similarity of clien t narrative disclosures is negatively related to audit fees, even after controlling for the similarity of the underlying financial statements. As an evaluation of the economic significance of the of the effect, moving from the 25th to the 75th percentile o f SIMFS decreases audit fees by a range of 2.8% in the footnote model to 3.8% in the business description model. Corresponding changes in SIMBUS, SIMMD&A, and SIMNOTES are associated with additional declines in audit fees of 4.5%, 2.1%, and 1.5%, respectiv ely. The largest combined effect is in the business description model where combined interquartile changes in both the financial statements and business description are associated with an 8.3% decrease in fees. Even the smallest economic effect the footnot es is a combined 4.3%. By comparison, the economic effect of being an industry specialist based on market share ( INDSPEC ) increases audit fees by a range of 4.7% to 8.8%, depending on which of the four models in Table 32 is considered. Therefore, the rela tion between client commonality and audit fees is both statistically and economically significant. Hypothesis 2 The second hypothesis predicts t he negative relation in H1 is magnified in industries that are more economically important to the auditor. Chung and Kallapur (2003) measure individual client importance by calculating the clients audit fees scaled
74 by total fees received by the auditor in that year. My tests require a measure of the importance of a group of clients, rather than one specific client.13 Therefore, as a proxy for economic importance I use portfolio shar e ( PORTSHR), the audit fees received from a particular industry year divided by the auditors total fees from all industries in the same year. In keeping with the industry specialization literature, I include in the model the importance of the industry ( IM PIND), which is a dummy set to one if the portfolio share exceeds 2.8% (the upper quartile of PORTSHR). The industry specialization literature has previously used this measure as a proxy for an industrys economic importance to the auditor (Neal and Riley Jr. 2004) To test H2, I expand the model for the first hypothesis by addi ng a main term for IMPIND and its interaction with each similarity score. The interaction coefficients will be negative under H2s prediction that greater economic incentives accentuate the negative relation between similarity and fees. As shown in Table 3 3 the significant IMPIND main effects are positive, consistent with other studies, indicating that higher portfolio share is associated with higher fees. Focusing first on financial statement similarity, the coefficient on SIMFS remains negative (t = 7. 68) as found in the test of H1. The interaction of IMPIND and SIMFS is also negative (t = 3.54), supporting the economic incentive hypothesis. Moving on to the narrative disclosure similarities, I retain SIMFS in the model to control for underlying financ ial statement similarity. I then add each narrative disclosure score and its interaction with IMPIND to the model. The main effects remain significant, as previously found, but none of the interactions are significant. 13 Using their measure directly would bias in favor of a result in my setting since fees would effectively be both a dependent and independent variable.
75 While I can easily reject the H2 null for financial statements, the results for the narrative disclosures are not significant. One ex post interpretation for this outcome is that financial statements are quantitative, which should allow for specific technological improvements that would be more difficult to implement for soft, qualitative disclosures. In other words, while narratives proxy for underlying client similarity, it may be difficult to implement audit technology that specifically leverages this type of overlap. Financial statements a re also likely to be more stable than text, which should more easily allow for technology improvements. In untabulated analysis, I use PORTSHR in place of IMPIND as an alternative measure of industry importance. The conclusions do not change for the financ ial statements, MD&A, or footnotes, but the business description has significantly negative coefficients on both the main and interaction terms. Supporting the importance of stability within the narrative disclosures, the business description is the most s table of the three textual items and it also has the strongest support for H2 in this alternative analysis.14 Overall, I find some empirical evidence that the negative similarity fees relationship is incrementally negative as an industry becomes more import ant to the auditors revenue stream. As portfolio share increases, the auditor may gain more knowledge and streamline its process for these clients. Alternatively, the auditor may not have a different cost structure due to technological investments, but is just more willing or able to charge lower fees to retain these economically important clients. 14 Using the raw year over year disclosure modification score from Brown and Tucker (2011), the business description has an av erage modification score of only 0.09, as compared to much larger annual modification scores of 0.16 for MD&A and 0.14 for the footnotes.
76 Hypothesis 3 The final hypotheses examine the consistency between financial statements and narrative disclosures. To test these hypotheses, I split the financial statement similarity and each of the narrative disclosure similarities into terciles. I am particularly interested in misalignment between the lowest and highest terciles of the financials and text, so I create dummies indicating when such misalignments occur. For each narrative disclosure type, I set the corresponding POOLTEXT variable to one when financial statement similarity is low ( SIMFS is in the bottom tercile) and narrative similarity is high ( SIMBUS, SIMMD&A, or SIMNOTES is in the top tercile). T hese indicators correspond to the riskiest type of disclosure inconsistency pooled text since the financial statements portray a very atypical company for the industry, while its narrative disclosures are very similar to its peers. I then create DIFFTEXT d ummies set to one when financial statement similarity is high ( SIMFS is in the top tercile) and narrative similarity is low ( SIMBUS, SIMMD&A, or SIMNOTES is in the bottom tercile). While still inconsistent, these differentiated text misalignments have potentially benignand potentially beneficial explanations. Examining each of the narrative disclosures in separate models, I expand the base audit fee model to include the respective POOLTEXT and DIFFTEXT dummy for that disclosure type. In each model, I also control for SIMFS and the similarity of the textual disclosure being examined. H3a predicts the coefficient on POOLTEXT is positive and H3b predicts the coefficient on DIFF TEXT is nonzero (although a null result would not be unexpected). Table 3 4 presents the results of the test. Consistent with the earlier tests of H1, all the financial statement and narrative disclosure similarity scores are significantly
77 negative. As predicted by H3a, the coefficients on POOLTEXT for the business description, MD&A, and footnotes are all significantly positive (t = 5.05, t = 4.43, and t = 2.40, respectively). These results support the prediction that inconsistency in the form of pooled text (dissimilar financials, similar text) is associated with higher audit fees. Relative to other clients, the findings are consistent with pooled text clients either (1) representing higher idiosyncratic risk to the auditor or (2) leading to a lower willingness to implement technological improvements to take advantage of client commonality Turning to the test of differentiated narrative disclosures, DIFFTEXTBUS is significantly negative (t = 2.50, pvalue = 0.012), as is DIFFTEXTMD&A (t = 2.73). DIFFTEXTNOTES is negative, but insignificant (t = 1.07), potentially due to the much smaller sample size for the footnotes. The hypothesis makes only weak predictions about these coefficients because it is unclear whether differentiated text increases, decreases, or does not affect the risk and efficiency associated with auditing these inconsiste nt clients. However, the results support the idea that differentiated disclosures reduce risk or increase audit efficiency, even when they are inconsistent with the financial statements. Alternative Measures and Sensitivity Analyses Larger Reference Groups The primary narrative disclosure measures are calculated based on the similarity to the five clients that are most similar to the observation. To test the sensitivity of the results to this choice, I construct textual similarities using all clients in the auditor industry year. These three alternative measures have correlations with the original measures that range from 0.91 to 0.94.
78 Compared with the original test of H1, the negative relation between client similarities and audit fees is qualitatively sim ilar for the business description (t = 6.24), but somewhat weaker for the MD&A (t = 1.97; pvalue = 0.048) and footnotes (t = 2.27; p value = 0.023). The patterns for the second hypothesis are unchanged. For the final hypotheses regarding disclosure consistency, the results are qualitatively unchanged except that DIFF TEXTMD&A is now slightly less significant (t = 2.02; pvalue = 0.044) and DIFFTEXTBUS is no longer significant. Overall, the results are slightly weaker in a few cases as the reference group is expanded to include more dissimilar clients. The changes in significance could be due to additional measurement error in the proxies as less relevant peers affect the calculations. This pattern is also consistent with auditors either explicitly or implicitly taking into account the similarity of more narrowly constructed client subgroups than the GICS industry as defined by Standard & Poors. Minimum Reference Group Size Rather than requiring the auditor industry year to have at least five clients, I a lternatively require at least ten clients from which to choose the five most similar. The results are qualitatively unchanged for the business description and MD&A samples, but weaker for the footnote sample. These changes in significance seem to be attributable to a reduction in sample size from 9,806 observations to only 6,317. Accounting System Comparability As an alternative to my approach, I also examine the accounting system comparability measures from De Franco et al (2011) For each company, they regress 16 quarters of earnings (an accounting system output) on returns (the net economic events) to estimate the accounting function for that company. To determine the
79 similarity between any two observations, they use the fitted accounting function to predict earnings for each observation using actual returns. They interpret the difference between the two predicted earnings values as a measure of the difference in accounting systems. Aggregating these differences for all pairs of observations gives a measure of accounting system similarity for each company within an industry year ( C OMPACCT IND). They construct an alternative measure using only earnings by regressing 16 quarters of earnings of one company on the earnings of another. Aggregating the R2 from each regression also gives a proxy for accounting system similarity ( C O MPACCT R2 ). As a sensitivity test to my primary D2 metric, I calculate these two measures as described in more detail in De Franco et al (2011) as an alternative to SIMFS. As an alternative test of H1, I separately include the two accounting comparability measures in the base audit fee model. They are both significantly negative (t = 5.01 for COMPACCT IND and t = 2.14 for COMPACCT R2 ). These results hold whether or not I include SIMFS in the model, although SIMFS has a much higher economic magnitude and a more negative t statistic in both cases. I find no support for the second hypothesis when using these alternative measures. However, they strongly support H3a and H3b regarding disclosure consistency. Using COMPACCT IND, all of the POOLTEXT and DIFFTEXT coefficients are qualitatively similar to the original tests except that DIFFTEXTNOTES also becomes significantly negative, making it consistent with the business description and MD&A results. Using the relation between earnings and returns, De Franco et al (2011) develop an empirical proxy that is directly related to their theoretical construct. However, even though the statistical significance of their measures are similar to mine, SIMFS is much
80 more strongly related to audit pricing in terms of economic magnitude than their measures of accounting comparability. Therefore, depending on the context, each approach could provide unique insights as proxies for company similarity. The advantages of my approach are that it requires no knowledge about the functi onal form of the relationship, requires less time series data, and can include an arbitrary number of economic dimensions in the similarity score.15 Concluding Remarks I introduce measures of financial statement and narrative disclosure similarity as proxies for audit client overlap. As predicted, I find higher commonality among clients is associated with lower audit fees, which I interpret as reduced production costs arising from increased audit efficiency and reduced risk due to greater potential for improved audit technology and shared knowledge. These patterns are stronger when the auditor has higher financial incentives to profit from the nonidiosyncratic elements of the audit. I also find that inconsistencies between financial statements and narrative disclosures are associated with higher fees when these differences are consistent with the client attempting to reduce its apparent financial differences with peer companies. In contrast, I find lower audit fees when the narrative disclosures differ from f inancial statements in a manner consistent with the client revealing differentiating firm specific information. The measures I develop in this paper have additional potential applications in audit research. For example, client commonality could be relevant to a company that is choosing whether to keep their incumbent auditor or switch to a new one. A company might look for an auditor that already audits similar firms (or dissimilar firms if 15 The DeFranco et al (2011) approach can only be used as an alternative to SIMFS, the financial statement similarity, and not as a proxy for narrative disclosure similarity.
81 knowledge spillover is a competitive concern). Outside of audit res earch, the financial statement and narrative disclosure inconsistency result could have implications for a companys information environment. There could also be econometric applications when the research design calls for a control company that is very sim ilar to the original observation. While somewhat related to other measures of company similarity, such as the accounting comparability measure in De Franco et al (2011), I provide a broader alternative that could be preferable in certain research contexts In addition to the contribution provided by the measures themselves, the proxies allow me to explore topics that were previously difficult to examine empirically. I provide a more direct proxy for the potential of specialization than merely using the prevalent industry market share measures, which can be difficult to interpret. This paper is also one of the few to integrate narrative disclosures as an empirical proxy for elements of the audit.
82 Table 31. Audit fee model variables Panel A: Descriptive statistics Variable Mean Std d ev 25% Median 75% LNFEES 13.384 1.332 12.403 13.367 14.269 AT 3,590.358 13,797.860 110.045 439.236 1,771.330 SIZE 6.119 2 .050 4.701 6.085 7.479 CSITEMS 165.494 33.216 142.000 164.000 188.000 IRISK 0.235 0.185 0.085 0.199 0.341 LOSS 0.383 LEV 0.198 0.302 0.001 0.119 0.300 DELAY 0.253 NAS 11.260 3.575 10.692 12.024 13.190 BUSY 0.710 OPIN 0.474 ICMW 0.030 TENURE 9.712 8.567 3.000 7.000 13.000 INDSHR 0.254 0.111 0.171 0.244 0.325 PORTSHR 0.022 0.015 0.011 0.019 0.028 INDSPEC 0.249 PORTSPEC 0.282 Observations 33,006
83 Table 31. Continu ed. Panel B: Pairwise Pearson correlations of continuous variables in audit fee model LNFEES SIZE CSITEMS IRISK LEV NAS TENURE INDSHR PORTSHR SIM FS 0.02 0.13 0.11 0.10 0.07 0.02 0.05 0.01 0.10 SIM BUS 0.11 0.27 0.06 0.17 0.10 0.05 0.03 0.09 0.09 SIM MD&A 0.02 0.18 0.02 0.16 0.03 0.05 0.03 0.01 0.02 S IM NOTES 0.01 0.20 0.03 0.16 0.08 0.04 0.06 0.04 0.13 LNFEES 0.75 0.77 0.03 0.12 0.28 0.31 0.09 0.03 SIZE 0.65 0.07 0.19 0.38 0.33 0.06 0.10 CSITEMS 0.12 0.16 0.28 0.36 0.04 0.03 IRISK 0.09 0.06 0.10 0.04 0.09 LEV 0.06 0.01 0.0 3 0.07 NAS 0.18 0.02 0.02 TENURE 0.08 0.01 INDSHR 0.32 Panel A: For each client: LNFEES = log of audit fees. AT = total assets. SIZE = log of AT CSITEMS = # of nonmissing/nonzero variables in Compustat. IRISK = receivables plus inventory, scaled by AT LOSS = 1 if net income < 0. LEV = longterm debt, scaled by AT DELAY = 1 if 10K filed > 90 days after fiscal year end. NAS = log of nonaudit fees. BUSY = 1 if 12/31 fiscal year end. OPIN = 1 for nonstandard opinion. IC MW = 1 if material weakness in internal control. TENURE = number of years with the current auditor. INDSHR = % of industrys fees provided to the current auditor. PORTSHR = % of current auditors fees provided by the clients industry year. INDSPEC = 1 if INDSHR >= 32.5%. IMPIND = 1 if PORTSHR >= 2.8% (economically important industry). See Table 21 for descriptive statistics for SIM measures. Panel B: Correlations in bold are significant at the 5% level. SIM = similarity of observation to other clients in the financial statement ( FS ), business description ( BUS), MD&A ( MD&A ), and footnote ( NOTES) reference groups. See Table 2 2 for correlations of SIM measures
84 Table 32. OLS regression of audit fees on client similarity LNFEES LNFEES LNFEES LN FEES Exp Coef t stat Coef t stat Coef t stat Coef t stat (Intercept) 8.169 296.35 *** 8.275 268.65 *** 8.392 268.73 *** 8.426 182.00 *** SIZE + 0.414 135.46 *** 0.391 109.64 *** 0.371 100.40 *** 0.383 73.19 *** CSITEMS + 0.009 44 .70 *** 0.009 42.46 *** 0.009 42.87 *** 0.009 28.30 *** IRISK + 0.534 21.25 *** 0.463 17.14 *** 0.440 16.09 *** 0.446 11.47 *** LOSS + 0.173 21.85 *** 0.170 19.36 *** 0.157 17.68 *** 0.160 12.61 *** LEV + 0.016 0.84 0.034 2.12 ** 0.0 26 1.69 0.005 0.23 DELAY + 0.022 2.40 ** 0.112 7.75 *** 0.117 7.86 *** 0.064 2.89 *** NAS + 0.021 16.14 *** 0.018 12.28 *** 0.018 12.01 *** 0.013 6.50 *** BUSY + 0.071 9.21 *** 0.093 10.98 *** 0.088 10.04 *** 0.100 8.12 *** OPIN + 0.1 05 13.97 *** 0.100 11.74 *** 0.095 10.95 *** 0.079 6.49 *** ICMW + 0.512 24.07 *** 0.445 18.45 *** 0.446 18.15 *** 0.448 13.19 *** TENURE + 0.003 6.50 *** 0.002 5.50 *** 0.002 3.75 *** 0.002 2.39 ** INDSPEC ? 0.084 10.16 *** 0.073 7.79 *** 0.067 7.03 *** 0.046 3.36 *** SIM FS 0.016 9.41 *** 0.015 8.52 *** 0.013 7.41 *** 0.011 4.35 *** SIM BUS 0.339 9.98 *** SIM MD&A 0.158 4.77 *** SIM NOTES 0.245 2.86 *** Years Yes Yes Yes Yes Industries Yes Yes Yes Yes Adj R 2 0.808 0.822 0.810 0.803 Model F 5,729 4,111 3,666 1,632 Obs 32,412 21,450 20,530 9,806 H1 predicts negative coefficients on the SIM measures V ariables defined in Table 3 1. Standard errors are Huber White adjusted. ***, **, and indicate statistical significance at the 1%, 5%, and 10% levels, respectively, in a twotailed test.
85 Table 33. Regression of fees on cli ent similarity conditional on auditor incentives L NFEES LNFEES LNFEES LNFEES Coef t stat Coef t stat Coef t stat Coef t stat (Intercept) 8.169 293.82 *** 8.268 268.16 *** 8.385 268.46 *** 8.416 180.52 *** SIZE 0.414 135.42 *** 0.391 109.62 *** 0.371 100.26 *** 0.383 73.21 *** CSITEMS 0.009 44.81 *** 0.009 42.47 *** 0.009 42.90 *** 0.009 28.34 *** IRISK 0.533 21.24 *** 0.464 17.15 *** 0.440 16.11 *** 0.446 11.47 *** LOSS 0.173 21.79 *** 0.170 19.36 *** 0.157 17.68 *** 0.160 12.62 *** LEV 0.015 0.79 0.034 2.13 ** 0.026 1.68 0.006 0.24 DELAY 0.022 2.45 ** 0.112 7.74 *** 0.117 7.84 *** 0.064 2.89 *** NAS 0.021 16.14 *** 0.018 12.26 *** 0.018 12.00 *** 0.013 6.49 *** BUSY 0.071 9.24 *** 0.093 10.98 *** 0.088 10.06 *** 0.100 8.12 *** OPIN 0.104 13.90 *** 0.099 11.70 *** 0.094 10.90 *** 0.079 6.45 *** ICMW 0.513 24.11 *** 0.445 18.46 *** 0.446 18.16 *** 0.449 13.20 *** TENURE 0.003 6.47 *** 0.002 5.50 *** 0.002 3.74 *** 0.002 2.40 ** INDSPEC 0.077 8.67 *** 0.065 6.53 *** 0 .060 5.91 *** 0.038 2.66 *** ECONIMP 0.010 0.59 0.027 2.19 ** 0.024 1.92 0.027 1.59 SIM FS 0.014 7.68 *** 0.015 8.50 *** 0.013 7.39 *** 0.011 4.31 *** ECONIMP*SIM FS 0.009 3.54 *** SIM BUS 0.337 8.08 *** ECONIMP*SIM BUS 0.006 0.09 SIM MD&A 0.151 4.03 *** ECONIMP*SIM MD&A 0.022 0.33 SIM NOTES 0.317 2.86 *** ECONIMP*SIM NOTES 0.149 0.95
86 Table 3 3. Continued. Years Yes Yes Yes Yes Industries Yes Yes Yes Yes Adj R 2 0.809 0.822 0.810 0.803 Model F 5,270 3,787 3,374 1,502 Obs 32,412 21,450 20,530 9,806 H2 predicts negative coefficients on the interaction terms. IMPIND = 1 if PORTSHR >= 2.8% (economically important in dustry). Other variables defined in Table 3 1. Standard errors are Huber White adjusted. ***, **, and indicate statistical significance at the 1%, 5%, and 10% levels, respectively, in a twotailed test.
87 Table 34. Regression of fees on client similarity conditional on disclosure consistency LNFEES LNFEES LNFEES Exp Coef t stat Coef t stat Coef t stat (Intercept) 8.287 267.89 *** 8.400 267.85 *** 8.427 182.24 *** SIZE + 0.391 109.56 *** 0.370 100.24 *** 0.382 73.14 *** CSIT EMS + 0.009 42.48 *** 0.009 43.06 *** 0.009 28.42 *** IRISK + 0.459 16.99 *** 0.433 15.85 *** 0.442 11.36 *** LOSS + 0.170 19.31 *** 0.157 17.64 *** 0.160 12.61 *** LEV + 0.033 2.05 ** 0.027 1.75 0.005 0.23 DELAY + 0.111 7.72 *** 0 .117 7.87 *** 0.065 2.93 *** NAS + 0.018 12.33 *** 0.018 12.00 *** 0.013 6.49 *** BUSY + 0.094 11.05 *** 0.088 10.01 *** 0.100 8.12 *** OPIN + 0.099 11.68 *** 0.095 10.99 *** 0.079 6.49 *** ICMW + 0.445 18.48 *** 0.446 18.23 *** 0.451 13.2 7 *** TENURE + 0.002 5.45 *** 0.002 3.77 *** 0.002 2.43 ** INDSPEC ? 0.073 7.84 *** 0.069 7.23 *** 0.048 3.49 *** SIM FS 0.012 6.14 *** 0.010 5.46 *** 0.009 3.48 *** SIM BUS 0.415 11.31 *** POOLTEXT BUS + 0.077 5.05 *** DIFFTEXT BUS ? 0.033 2.50 ** SIM MD&A 0.225 6.35 *** POOLTEXT MD&A + 0.064 4.43 *** DIFFTEXT MD&A ? 0.036 2.73 *** SIM NOTES 0.323 3.54 *** POOLTEXT NOTES + 0.050 2.40 ** DIFFTEXT NOTES ? 0.022 1.07
88 T able 3 4. Continued. Years Yes Yes Yes Industries Yes Yes Yes Adj R 2 0.822 0.810 0.803 Model F 3,801 3,384 1,503 Obs 21,450 20,530 9,806 H3a predicts a positive coefficient on POOLTEXT H3b predicts the coeffic ient on DIFFTEXT is nonzero. POOLTEXT (DIFFTEXT) = 1 if observation is in the lowest (highest) tercile of financial statement similarity and the highest (lowest) tercile of narrative disclosure similarity. Other variables defined in Table 3 1. Errors are H uber White adjusted. ***, **, and indicate statistical significance at the 1%, 5%, and 10% levels, respectively, in a twotailed test.
89 CHAPTER 4 CONCLUSION I find broad evidence that clients systematically choose their auditor and that the relationshi p between a client and its auditor has implications for both audit quality and costs. Clients tend to prefer auditors that are more compatible with them, and are even more likely to switch auditors when the compatibility is poor. General measures of audit quality improve when auditor client compatibility increases, even though severe audit failures appear to also increase. When the compatibility is higher, audit fees tend to be lower, which I argue is indicative of lower production costs for the auditor that arise through specialization of the audit process to a specific set of clients. Reinforcing this interpretation, the fees are even lower when the auditor has stronger incentives to develop these specialized processes. As further support, fees are higher in the presence of inconsistencies among the clients disclosure channels, which I interpret as a constraint on auditor specialization. I use the similarity of a client to other clients of the same auditor to proxy for auditor client compatibility and the commonality among a set of clients. I develop two novel measures of how similar one company is to a set of other companies. I take a broad approach by using both financial statements and narrative disclosures as sources of information, and then introduce t wo distinct algorithms appropriate to the structure of each information signal. The Mahalanobis D2 measure has been used to a limited extent in accounting research, but solely for statistical purposes and not as a construct of interest. This measure is appropriate for any setting in which a researcher has a small set of numeric variables. No specific (known) relationship among the variables is necessary, other than them being relevant for the
90 context at hand. For example, no assumptions about an earnings re turn relationship is necessary, as is required in one of the few accounting similarity scores currently available (De Franco et al 2011). The Mahalanobis distance is robust to correlations among the variables and is not affected by the scale of those vari ables. In this study, my basic analysis uses five variables known to be important in an audit context. However, the measure seems quite robust to various input variables, with no changes to my qualitative results when various variables are removed or added to the input set. Overall, the D2 approach seems very useful when constructing a single summary measure of similarity and dissimilarity among a set of financial statements. While the Mahalanobis distance is very powerful and robust, it is not feasibly imp lemented when the set of input variables is extremely large, as is the case when long textual items are used as an information source. In this situation, I provide an alternative that can handle extremely large quantities of information. In fact, the Vector Space Model (VSM) approach I use is implemented by Internet search engines that need to compare many billions of documents. In the current context, I extend the method introduced to the accounting literature in Brown and Tucker (2011), which analyzes onl y two documents, so that I can compare one company to others. The similarity scores I produce are correlated with other known measures of similarity, such as discretionary accruals, simple differences of the company from the mean of the industry, and even the Mahalanobis distance. These correlations with a variety of financial statement constructs are significant despite the fact that the languagebased sources of the similarity measures contain no references to the financial statement variables (they are i ntentionally stripped out of the calculations). My findings
91 demonstrate the VSM approach is continuing to show promise as a measure of narrative disclosure similarity. The two similarity algorithms have additional applications within accounting research. F or example, a common econometric problem is to match a treated observation (e.g., a company receiving an SEC enforcement action) with an untreated observation. A common approach is to match on size and industry, which can produce matches that are not v ery close, other than along the size dimension (and sometimes the observations are not even very similar in size). The Mahalanobis approach is already used to perform a match along multiple dimensions, as in the psmatch2 command within the Stata software package. However, I provide a new alternative based on narrative disclosures that could pair companies who have made very similar disclosure decisions even if they are not that similar financially. Another area of recent focus in accounting research is on t he implications of networks. One example is social networks, which we observe when individuals are connected through board memberships, social clubs, and alma maters. Another example is in studies of how accounting standards and other disclosures become di ffused throughout an economy. These timely topics are concerned with how various entities are connected to one another. Rather than using the traditional econometric approach of assuming independence between observations, they seek to exploit this dependence. My measures provide another means for researchers to study these connections, because they explicitly proxy for the degree of connectedness among observations. For instance, the narrative disclosure similarity score could be used to detect when a speci fic accounting disclosure spreads throughout an industry.
92 While showing promise for future research, my findings also have implications for regulators. One persistent issue raised in the United States is the potential for mandatory rotation of auditors aft er a specific number of years, a rule already in place in other countries. There are obvious potential benefits for audit quality under a mandatory rotation regime because independence in fact and appearance are potentially better preserved. However, a tri ckier issue is assessing the costs of such rotations. My results show that clients do not randomly choose among auditors of a given type (e.g., Big4 auditors). Therefore, a nonvoluntary auditor change has the potential to move a client away from the firs t best auditor for that company; in other words, auditors are not fungible. Auditor switching costs are already documented due to the learning process that takes place in the first years of an audit engagement. However, my studies point to other costs of forcing clients to leave their preferred auditors. These costs take the form of both changes in audit quality and audit fees. Overall, my findings contribute to our understanding of the nature of the auditor client relationship, especially given a relative lack of literature on the mechanisms by which specific clients are connected to specific auditors. These conclusions have implications for both regulators and researchers. In addition, I contribute novel measures of inter company similarity that have broa d potential applications in a variety of studies, even those outside of auditing research.
93 APPENDIX A EXTRACTION OF ANNUAL REPORT ITEMS To gather the business description, MD&A, and footnotes sample, I begin by downloading all 10Ks and 10K405s available on the SECs EDGAR system that meet the following requirements: (1) fiscal years between 1997 and 2009, (2) assets greater than $1 million, (3) no change in fiscal year end, (4) not in the utilities or financial services industries, and (5) engaging a Big4 auditor. As described in Table A1, this initial screen leaves 41,782 annual reports. I next screen out any unusually short annual reports since these typically belong to holding companies, firms that are winding down, and other atypical observations. I use a cutoff of 50,000 characters for this purpose (approximately the 4th percentile of 10K length). This value filters out most of the unwanted observations without losing a substantial number of desired reports. I use characters instead of words because the tables and numbers contained in the report make it difficult to split the document into words at this point in the process. These filters leave 40,149 annual reports. I begin the item extraction process by stripping all HTML formatting and data ta bles as in Li (2008; 2010) I then split each annual report into its component items, keeping only the business description, MD&A, and footnotes (the financial statements are removed when data tables are discarded). I remove any narrative disclosures that contain l anguage indicating the relevant section has been omitted as permitted by regulation. I skip disclosures that are included by reference, either to an external document or an attached exhibit, since the variety of alternate locations dramatically increases t he difficulty in obtaining that data. The footnotes, in particular, are frequently included by reference. I drop any remaining items
94 that do not contain at least 150 characters. Items shorter than this cutoff have typically been omitted or included by reference, but do so using somewhat unusual wording that my initial string search did not recognize. I split each item into words, keeping disclosures with at least 500 words. Items shorter than this length are relatively unusual and are unlikely to provide a meaningful comparison to disclosures by peers in the auditor industry year reference group. Finally, I exclude items exceeding 20,000 words because these frequently indicate problems splitting the 10K into separate items. For example, the extraction process might erroneously treat the entire annual report as the business description due to misspellings and other idiosyncratic document features. Archival studies frequently handle outliers such as these through deletion, winsorization, or robust techniques d uring the empirical analysis. However, doing so in the current study would allow these outliers to be in reference groups and therefore have an undesirable influence on the calculation of the similarity scores. There are fewer observations in the narrative disclosure samples than in the financial statement sample, primarily due to unavailable reports on EDGAR, items included by reference to other locations, and textual idiosyncrasies that lead to problems extracting the 10K items of interest. The substanti al drop in the number of footnote observations, as compared to the business description and MD&A samples, is because many companies attach financial statements and footnotes as an exhibit to the report in a variety of unpredictable ways, making their autom ated extraction difficult.
95 Table A 1. Narrative disclosure sample selection process Reports 10K available on EDGAR; fiscal years 19972009; Compustat assets > $1M; no FYE change; excl. financials and utilities; Big4 auditor 41,78 2 Less: Short reports (<50,000 characters) (1,633) Total annual reports available 40,149 Bus d esc MD&A Footnotes Less: Item not succesfully extracted (1,918) (936) (1,166) Less : Item specifically omitted (10) (41) (117) Less: Item included by reference (23) (2,840) (10,516) Less: Short items (<150 characters) (1,277) (1,173) (3,415) Less: < 500 or > 20,000 words (886) (1,257) (7,227) Less: < 5 other clients in auditor industry year (2,680) (2,622) (3,269) Total items available 33,355 31,280 14,439
96 APPENDI X B CALCULATION OF NARRA TIVE DISCLOSURE SIMI LARITY SCORE As described in Brown and Tucker (2011), the Vector Space Model (VSM) maps a document into a vector, v, with each vector element, wi, representing the weighted frequency of a word in that document. T he weighted frequency is zero if the word does not occur in that document and the length of the vector is n the number of unique words in all documents of the sample: For example, assume there are only two documents in the sample: (1) Earnings have inc reased. and (2) Earnings have decreased. The length of each document vector is four, since there are four unique words in the sample: w1 corresponds to earnings, w2 to have, w3 to increased, and w4 to decreased. The two documents are then repres ented as: The vectors allow for various comparisons between documents in the sample (Manning and Schtze 1999) The cosine of the angle, between any two vectors, vi and vj, is a proxy for the similarity of any two underlying doc uments, SIMDOC,i,j: vivi vj length of vj. SIMDOC ranges from zero (completely dissimilar documents) to one (identical documents). I stem all words using the Porter stemming algorithm to reduce the dimensionality of the data, which in turn limits the computing time and resources
97 required (e.g., earnings, earned, and earn are all converted to earn).1 Consistent with Brown and Tucker (2011) I use the term frequency inverse document frequency (TF IDF) algorithm to decrease the weight on frequently used words and increase the weight on uncommon words.2 Therefore, instead of a raw frequency count, each document vector element is the frequency count of the word multiplied by a weight based on the relative prominence of that word in the entire sample. Because Brown and Tucker (2011) are interested in the differences between just two documents at a time, they only calculate pairwise similarity scores. In contrast, I aggregate these pairwise scores to get a measure of the similarity between one narrative disclosure and the disclosures issued by the client reference group. As with the financial statement similarities, the reference group contains other clients of the same auditor, within the same GICS industry and year. To combine the pairwise SIMDOC,i,j scores between client i and all other clients j in the same auditor, industry, and year, I average the pairwise similarities to get SIMDOC,i for each observation in my sample. I calculate the SIMDOC,i similarity measure for each observation in the business description ( RAWSIMBUS), MD&A (RAWSI MMD&A) and footnote ( RAWS IMNOTES) samples. However, Brown and Tucker (2011) show that these raw scores are positively related to document length because of the mechanics of the calculation, rather than due 1 Even with the reduced dimensions, the calculations take over one week to run on a 2.66 GHz, quadcore machine, while occupying most of the 6 gigabytes of working memory. 2 I do not use a stop word list to remove extremely common (i.e., unimportant) words, such as the and a, from the sample as in Li (2010) These words will receive a weight of zero, or very close to it, via the TF IDF weighting procedure. Brown and Tucker (2011) find no substantial difference in their conclusions between using the TF IDF approach and a simple frequency count combined with a stop word li st. I generate the TF IDF weights independently for each type of narrative disclosure.
98 to any meaningful underlying relation. They control for this relationship by regressing the raw similarity on the first five powers of the number of words in the observation i document ( LENBUS, LENMD&A, and LENNOTES in the current study) in the current study I use the first three powers because the magnit udes of the coefficients rapidly approach zero after this point.3 In order to maximize the sample size for making this adjustment, I use all available observations, including those from nonBig4 auditors; for all other tests in the paper, I use only client s of Big4 auditors. Regressing the raw similarity scores on the first three powers of the document length yields a residual that represents the variation in the raw similarity scores that cannot be explained by these factors. I label these residuals SIMBUS, SIMMD&A, and SIMNOTES, producing the similarity scores I use in my analysis. Descriptive data for these measure components are in Table B1. 3 Hanley and Hoberg (2012) use the VSM to measure the similarity of an IPO prospect us to all the recent IPOs experiencing litigation problems. However, they do not control for document length, making it difficult to ascertain the validity of their measure.
99 Table B 1. Calculation of narrative disclosure similarity measures Variable Mean Std d ev 25% Median 75% Obs SIM BUS 0.002 0.077 (0.053) (0.017) 0.040 33,355 SIM MD&A 0.002 0.087 (0.057) (0.024) 0.038 31,280 SIM NOTES 0.000 0.048 (0.029) (0.013) 0.011 14,439 RAWSIM BUS 0.108 0.085 0.045 0.087 0.148 36,035 RAWSIM MD&A 0.113 0.095 0.044 0.086 0.155 33,902 RAWSIM NOTES 0.054 0.061 0.021 0.037 0.064 17,708 LEN BUS 6,338 3,739 3,602 5,471 8,252 36,253 LEN MD&A 7,054 3,870 3,984 6,437 9,401 34,131 LEN NOTES 8,623 4,018 5,423 7,934 11,178 18,083 Subscripts: BUS = Business Description, MD&A = Managements Discussion & Analysis, NOTES = Footnotes to Financial Statements. Variables: SIM = similarity of observation to other clients in the reference group, adjusted for LEN ; higher va lue indicates more similarity. RAWSIM = SIM before adjustment. LEN = # of words in the observations text.
100 LIST OF REFERENCES Amir, E. 1993. The Market Valuation of Accounting Information: The Case of Postretirement Benefits Other Than Pensions. The Accounting Review 68 (4): 703 724. Balsam, S., J. Krishnan, and J. S. Yang. 2003. Auditor Industry Specialization and Earnings Quality. Auditing 22 (2): 71 97. Bamber, E. M., and V. M. Iyer. 2007. Auditors Identification with Their Clients and Its Effect on Auditors Objectivity. Auditing 26 (2): 1 24. Bamber, L. S., and Y. S. Cheon. 1998. Discretionary Management Earnings Forecast Disclosures: Antecedents and Outcomes Associated with Forecast Venue and Fore cast Specificity Choices. Journal of Accounting Research 36 (2): 167 190. Beck, P. J., and M. G. H. Wu. 2006. Learning by Doing and Audit Quality. Contemporary Accounting Research 23 (1): 1 30. Berry, H., M. F. Guilln, and N. Zhou. 2010. An Institutional Approach to Cross National Distance. Journal of International Business Studies 41 (9): 1460 1480. Blokdijk, H., F. Drieenhuizen, D. A. Simunic, and M. T. Stein. 2006. An Analysis of Cross Sectional Differences in Big and NonBig Public Accounting Firms Au dit Programs. Auditing 25 (1): 27 48. Brown, S. V., and J. W. Tucker. 2011. LargeSample Evidence on Firms Year over Year MD&A Modifications. Journal of Accounting Research 49 (2): 309 346. Cahan, S., J. Godfrey, J. Hamilton, and D. Jeter. 2008. Auditor S pecialization, Auditor Dominance, and Audit Fees: The Role of Investment Opportunities. The Accounting Review 83 (6): 1393 1423. Cairney, T. D., and G. R. Young. 2006. Homogenous Industries and Auditor Specialization: An Indication of Production Economies. Auditing 25 (1): 49 67. Campbell, J. L., H. Chen, D. S. Dhaliwal, H. M. Lu, and L. Steele. 2010. The Information Content of Mandatory Risk Factor Disclosures in Corporate Filings. Working Paper Causholli, M., M. De Martinis, D. Hay, and W. R. Knechel. 20 10. Audit Markets, Fees and Production: Towards an Integrated View of Empirical Audit Research. Journal of Accounting Literature 29: 167 215. Chaney, P. K., D. C. Jeter, and L. Shivakumar. 2004. Self Selection of Auditors and Audit Pricing in Private Firms The Accounting Review 79 (1): 51 72.
101 Chung, H., and S. Kallapur. 2003. Client Importance, Nonaudit Services, and Abnormal Accruals. The Accounting Review 78 (4): 931 955. Craswell, A. T., J. R. Francis, and S. L. Taylor. 1995. Auditor brand name reputati ons and industry specializations. Journal of Accounting and Economics 20 (3): 297 322. Darr, E. D., L. Argote, and D. Epple. 1995. The Acquisition, Transfer, and Depreciation of Knowledge in Service Organizations: Productivity in Franchises. Management Sci ence 41 (11): 1750 1762. Davis, L. R., D. N. Ricchiute, and G. Trompeter. 1993. Audit Effort, Audit Fees, and the Provision of Nonaudit Services to Audit Clients. The Accounting Review 68 (1): 135 150. Dechow, P., W. Ge, C. Larson, and R. Sloan. 2011. Predicting Material Accounting Misstatements. Contemporary Accounting Research 28 (1): 17 82. DeFond, M. L., and J. Jiambalvo. 1994. Debt Covenant Violation and Manipulation of Accruals. Journal of Accounting and Economics 17 (12): 145 176. DeFond, M. L., and K. R. Subramanyam. 1998. Auditor changes and discretionary accruals. Journal of Accounting and Economics 25 (1): 35 67. Dowling, C. 2009. Appropriate Audit Support System Use: The Influence of Auditor, Audit Team, and Firm Factors. The Accounting Review 8 4 (3): 771 810. Dunn, K. A., and B. W. Mayhew. 2004. Audit Firm Industry Specialization and Client Disclosure Quality. Review of Accounting Studies 9 (1): 35 58. Dye, R. A. 1991. Informationally motivated auditor replacement. Journal of Accounting and Econ omics 14 (4): 347 374. Feldman, R., S. Govindaraj, J. Livnat, and B. Segal. 2010. Managements tone change, post earnings announcement drift and accruals. Review of Accounting Studies 15 (4): 915 953. De Franco, G., S. P. Kothari, and R. S. Verdi. 2011. Th e Benefits of Financial Statement Comparability. Journal of Accounting Research 49 (4): 895 931. Fried, D., and A. Schiff. 1981. CPA Switches and Associated Market Reactions. The Accounting Review 56 (2): 326 341. Gramling, A. A., and D. N. Stone. 2001. Au dit firm industry expertise: A review and synthesis of the archival literature. Journal of Accounting Literature 20: 1 29.
102 Guilding, C., and L. McManus. 2002. The Incidence, Perceived Merit and Antecedents of Customer Accounting: An Exploratory Note. Accou nting, Organizations and Society 27 (12): 45 59. Gul, F. A., S. Y. K. Fung, and B. Jaggi. 2009. Earnings quality: Some evidence on the role of auditor tenure and auditors industry expertise. Journal of Accounting and Economics 47 (3): 265 287. Hackenbrac k, K., and W. R. Knechel. 1997. Resource allocation decisions in audit engagements. Contemporary Accounting Research 14 (3): 481 499. Hair, J. F., W. C. Black, B. J. Babin, R. E. Anderson, and R. L. Tatham. 2006. Multivariate Data Analysis 6th ed. Upper S addle River, NJ: Pearson Prentice Hall. Hammersley, J. S. 2006. Pattern Identification and Industry Specialist Auditors. The Accounting Review 81 (2): 309 366. Hanley, K. W., and G. Hoberg. 2012. Litigation risk, strategic disclosure and the underpricing of initial public offerings. Journal of Financial Economics 103 (2): 235 254. Hay, D. C., W. R. Knechel, and N. Wong. 2006. Audit Fees: A Metaanalysis of the Effect of Supply and Demand Attributes. Contemporary Accounting Research 23 (1): 141 191. Iyer, V. M. 1998. Characteristics of Accounting Firm Alumni Who Benefit Their Former Firm. Accounting Horizons 12 (1): 18 30. Jaffe, A. B. 1986. Technological Opportunity and Spillovers of R&D: Evidence from Firms Patents, Profits, and Market Value. The American Economic Review 76 (5): 984 1001. voluntary auditor changes. Journal of Accounting and Economics 12 (13): 281 308. Johnstone, K. M., C. Li, and S. Luo. 2011. Client Supply Chain Relationships, Audit Firm Selection, and Implications for Audit Quality and Pricing. Working Paper Knechel, W. R., V. Naiker, and G. Pacheco. 2007. Does Auditor Industry Specialization Matter? Evidence from Market Reaction to Auditor Switches. Auditing 26 (1): 19 45. Knechel, W. R., L. Niemi, and M. Zerni. 2012. Empirical Evidence on the Implicit Determinants of Compensation in Big four Audit Partnerships. Working Paper
103 Kravet, T. D., and V. Muslu. 2010. Informativeness of Risk Disclosures in Corporate Annual Reports. Working Paper Krishnan, G. V. 2003. Does big 6 auditor industry expertise constrain earnings management? Accounting Horizons 17: 1 16. Krishnan, J. 1994. Auditor Switching and Conservatism. The Accounting Review 69 (1): 200 215. Landsman, W., K. Nelson, and B. Rountree. 2009. Auditor Switches in the Preand Post Enron Eras: Risk or Realignment? The Accounting Review 84 (2): 531 558. Lawrence, A., M. Minutti Meza, and P. Zhang. 2011. Can Big 4 versus NonBig 4 Differences in Audit Quality Proxies Be Attributed to Client Characteristics? The Accounting Review 86 (1): 259 286. Lennox, C. 2005. Audit quality and executive officers affiliations with CPA firms. Journal of Accounting and Economics 39 (2): 201 231. Lennox, C., and C. Park. 2007. Audit Firm Appointments, Audit Firm Alumni, and Audit Committee Independence. Contemporary Accounting Research 24 (1): 235 258. Li, F. 2008. Annual Report Readability, Current Earnings, and Earnings Persistence. Journal of Accounting and Economics 45 (23) : 221 247. 2010. The Information Content of ForwardLooking Statements in Corporate Filings A Nave Bayesian Machine Learning Approach. Journal of Accounting Research 48 (5): 1049 1102. Lim, C. Y., and H.T. Tan. 2008. Nonaudit Service Fees and Audit Quality: The Impact of Auditor Specialization. Journal of Accounting Research 46 (1): 199 246. Low, K. Y. 2004. The Effects of Industry Specialization on Audit Risk Assessments and Audit Planning Decisions. The Accounting Review 79 (1): 201 219. De Maessc halck, R., D. Jouan Rimbaud, and D. L. Massart. 2000. The Mahalanobis Distance. Chemometrics and Intelligent Laboratory Systems 50 (1): 1 18. Mahalanobis, P. C. 1936. On the Generalized Distance in Statistics. Proceedings of the National Institute of Sciences of India 2 (1): 49 55. Manning, C. D., and H. Schtze. 1999. Foundations of Statistical Natural Language Processing Cambridge, Mass: MIT Press.
104 Mayhew, B. W., and M. S. Wilkins. 2003. Audit Firm Industry Specialization as a Differentiation Strategy: E vidence from Fees Charged to Firms Going Public. Auditing 22 (2): 33 52. Menon, K., and D. D. Williams. 2004. Former Audit Partners and Abnormal Accruals. The Accounting Review 79 (4): 1095 1118. Mimmack, G. M., S. J. Mason, and J. S. Galpin. 2001. Choice of Distance Matrices in Cluster Analysis: Defining Regions. Journal of Climate 14 (12): 2790 2797. Minutti Meza, M. 2010. Does Auditor Industry Specialization Improve Audit Quality? Evidence from Comparable Clients. Working Paper Neal, T. L., and R. R. Riley Jr. 2004. Auditor Industry Specialist Research Design. Auditing 23 (2): 169 177. Nichols, D. R., and D. B. Smith. 1983. Auditor Credibility and Auditor Changes. Journal of Accounting Research 21 (2): 534 544. Numan, W., and M. Willekens. 2012. An empir ical test of spatial competition in the audit market. Journal of Accounting and Economics 53 (1 2): 450 465. OKeefe, T. B., R. D. King, and K. M. Gaver. 1994. Audit Fees, Industry Specialization, and Compliance with GAAS Reporting Standards. Auditing 13 ( 2): 41 55. OKeefe, T. B., D. A. Simunic, and M. T. Stein. 1994. The Production of Audit Services: Evidence from a Major Public Accounting Firm. Journal of Accounting Research 32 (2): 241 261. Owhoso, V. E., W. F. Messier Jr., and J. G. Lynch Jr. 2002. Err or Detection by Industry Specialized Teams during Sequential Audit Review. Journal of Accounting Research 40 (3): 883 900. Palmrose, Z. V. 1986. Audit Fees and Auditor Size: Further Evidence. Journal of Accounting Research 24 (1): 97 110. PCAOB. 2011. Conc ept Release on Auditor Independence and Audit Firm Rotation. PCAOB Release No. 2011006 Public Company Accounting Oversight Board. 2011. Concept Release on Possible Revisions to Pcaob Standards Related to Reports on Audited Financial Statements PCAOB Release No. 2011003 Rege, U. P. 1984. Accounting Ratios to Locate TakeOver Targets. Journal of Business Finance & Accounting 11 (3): 301 311.
105 Richardson, S. A., R. G. Sloan, M. T. Soliman, and I. Tuna. 2005. Accrual reliability, earnings persistence and st ock prices. Journal of Accounting and Economics 39 (3): 437 485. Riedl, E., and S. Srinivasan. 2010. Signaling Firm Performance Through Financial Statement Presentation:: An Analysis Using Special Items. Contemporary Accounting Research 27 (1): 289 332. Sa lton, G., A. Wong, and C. S. Yang. 1975. A Vector Space Model for Automatic Indexing. Communications of the ACM 18 (11): 613 620. Securities and Exchange Commission. 1987. Concept Release on Managements Discussion and Analysis of Financial Condition and O perations Release No. 336711. 1989. Managements Discussion and Analysis of Financial Condition and Results of Operations; Certain Investment Company Disclosures Release No. 336835. 2003. Commission Guidance Regarding Managements Discussion and Analysis of Financial Condition and Results of Operations Release No. 338350 Shevlin, T. 1991. The Valuation of R&D Firms with R&D Limited Partnerships. The Accounting Review 66 (1): 1 21. Shu, S. Z. 2000. Auditor resignations: clientele effects and legal liability. Journal of Accounting and Economics 29 (2): 173 205. Singhal, A. 2001. Modern Information Retrieval: A Brief Overview. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering 24 (4): 35 43. Sirois, L. P., and D. A. S imunic. 2010. Auditor Size and Audit Quality Revisited: The Importance of Audit Technology. Working Paper Sun, Y. 2010. Do MD&A Disclosures Help Users Interpret Disproportionate Inventory Increases? The Accounting Review 85 (4): 1411 1440. Tasker, S. C. 1 998. Bridging the Information Gap: Quarterly Conference Calls as a Medium for Voluntary Disclosure. Review of Accounting Studies 3 (1): 137 167. Wahlen, J. M. 1994. The Nature of Information in Commercial Bank Loan Loss Disclosures. The Accounting Review 6 9 (3): 455 478. Ward, D. D., R. J. Elder, and S. C. Kattelus. 1994. Further Evidence on the Determinants of Municipal Audit Fees. The Accounting Review 69 (2): 399 411.
106 Willenborg, M. 2002. Discussion of Brand Name Audit Pricing, Industry Specialization, and Leadership Premiums post Big 8 and Big 6 Mergers. Contemporary Accounting Research 19 (1): 111 116.
107 BIOGRAPHICAL SKETCH Stephen V. Brown has a Bachelor of Science in computer science from the University of North Florida in Jacksonville, Florida, where he also earned a Master of Business Administration with a concentration in a ccounting. He was a computer analyst and programmer before becoming licensed as a Certified Public Accountant in the State of Florida. His Ph.D. from the University of Flori da, is in b usiness a dministration with a concentration in a ccounting.