The Effect of Office-Level Factors on Audit Quality

Transcription

1 The Effect of Office-Level Factors on Audit Quality William Floyd 1 Maureen McNichols 2 Patricia C. O Brien 3 Rimmy E. Tomy 4 February 2017 Preliminary draft: Please do not circulate without permission of the authors Abstract We aim to provide new evidence on factors that affect audit quality, among U.S. clients of Big 4 audit offices during A key contribution of this study is our method of controlling for clients misreporting risk, which helps us to separate client risk from auditor traits and behaviors associated with audit failure. We create, for each audit office, a synthetic office of similarly risky clients of different auditors, and use the synthetic office as a benchmark. We then use this method to study how the auditor s willingness to report unfavorable information, via additional language in the audit opinion, affects restatement rates and rates of receiving SEC comment letters. Our validity tests show that the synthetic office is an effective control. Our results do not, however, support the conjecture that audit offices exhibiting greater willingness to voice concerns provide higher quality audits. We gratefully acknowledge the helpful comments of seminar participants at Stanford University. All remaining errors are our own. 1 Pluribus Labs. william.floyd@pluribuslabs.com 2 Stanford University. fmcnich@stanford.edu 3 University of Waterloo. pobrien@uwaterloo.ca 4 University of Chicago. rtomy@uchicago.edu 0

2 1. Introduction Assessing audit quality entails understanding how auditors carry out the watchdog role entrusted to them by investors, audit committees, regulators and other stakeholders of financial reporting. Important elements of the audit process include the auditor s initial decision to accept a client, the auditor s expertise and effort in designing and carrying out an appropriate work plan, and the auditor s willingness to ensure correction of problems and report unfavorable findings when necessary. The factors that influence these decisions, and their consequences for audit quality are of great relevance to audit firms, regulators and financial statement users, and as a consequence, they are the focus of a large literature. Our research aims to provide new evidence on the role of office-level factors in audit quality. We address five questions about audit quality at the office level. First, do audit outcomes vary significantly across offices of Big 4 firms? Second, does this variation remain after controlling for the risk of the office s clients? Third, do audit offices vary significantly in the willingness of auditors to reveal unfavorable findings? Fourth, does this variation remain after controlling for client risk? Lastly, do offices where auditors are more willing to issue unfavorable findings have higher quality outcomes? Studying audit quality at the office level allows us to test hypotheses about audit quality at a level closer to the engagement team than the region or the audit firm. This has the potential to allow more refined inferences about audit quality than studies that focus on differences between Big N and non-big N auditors. 1 Aggregate quality at the audit firm level may mask variation in quality across offices that could pinpoint specific influences on audit quality, such as office culture, expertise or incentives. Finding variation in audit quality across offices may 1 For discussion of this literature, see Lawrence, Minutti-Meza and Zhang (2011), Defond and Zhang (2014) and Defond, Erkens and Zhang (2014). 1

3 highlight locations that warrant further focus by audit firms and regulators. Furthermore, because the office level of an audit is publicly available, investors in U.S. listed companies can condition on this information. A key contribution of this study is our method of controlling for the risk that clients misreport. In broad terms, we estimate a model of client-year misreporting risk, and use it to calculate a risk score for each client-year. Then, for each audit office-year, we construct a synthetic office to act as the control for that office-year. For each client audited in a given Big 4 office, we identify the firm with the closest estimated risk of misreporting from the same industry and year, audited by a different audit firm. The synthetic office is the set of matches for a given office s actual clients with the closest possible risk of misreporting. The control for clients risk of misreporting is important for the policy implications of this work. If the variation in audit failures is solely due to differences in client risk, and no new approach is available to improve audit quality for riskier clients, then improving audit quality is a zero-sum game; no societal benefit can be gained by shifting risky clients around. On the other hand, if audit outcomes differ after controlling for client risk, then improvements in audit practices at the office level can be beneficial. Our tests use restatement frequency and SEC comment letter frequency as measures of low audit quality, motivated by recent literature such as Defond and Zhang (2014) and Hribar, Kravet and Wilson (2014). For a restatement to occur, the issuer must have initially committed an unintentional or intentional error that results in materially misstated financial statements. In addition, the auditor must either have failed to identify the reporting problem, or identified the problem but failed to require the firm to make a timely correction. The SEC s Division of Corporate Finance (DCF) issues a comment letter, pursuant to SOX section 408, when their 2

4 review of the firm s filed financial statements raises questions. Gao, Lawrence and Smith (2010) note the DCF can be viewed as a monitor of last resort because the financial statements have previously been audited and reviewed by other monitors. As such, the frequency of comment letters at the audit office level can reflect the quality of their audits (Johnston and Petacchi, 2015). With the synthetic office design described above to control for the client risk of misreporting, we can focus more closely on the auditor s failure to detect or report errors. In addition to measures of audit quality based on outcomes, we study additional language in the audit opinion to assess whether auditors willingness to issue unfavorable findings is associated with higher audit quality. The literature has studied going concern opinions as one measure of auditors willingness to issue unfavorable reports concerning their clients. We focus on additional language in the audit opinion following the approach of Czerney, Schmidt and Thompson, (2014), to capture a broader measure of disclosure of unfavorable audit findings. 2 Their study examines a number of modifications and additions to the audit opinion, including emphasis of matter, reference to consistency with prior financial statements due to accounting principles or restatements, audit-related matters such as division of responsibility or scope, and references to going concern or supplemental information. Because issuing anything other than the standard audit opinion can potentially impair the auditor s relationship with the client, we believe additional language credibly reflects the auditor s view. Thus, after controlling for client risk of misreporting, an office with a higher rate of additional language is likely an office where auditors report the issues they have identified. Relatedly, audit offices with lower rates of additional language for a given degree of client risk 2 We plan to incorporate auditor identification of internal control weaknesses (Rice and Weber, 2012) as an additional measure of willingness to issue unfavorable findings in the next draft of this paper. 3

5 are likely offices where auditors identify fewer issues or are reluctant to report the issues they have identified. We hypothesize that the willingness to identify and disclose unfavorable findings will be associated with higher audit quality. Therefore, we expect that, among clients whose audit opinions do not include additional language, those audited by offices with higher rates of additional language will experience fewer adverse outcomes such as restatements or SEC comment letters than those audited by offices with lower rates of additional language. Our research is subject to certain limitations. First, we focus on Big 4 auditors to obtain a relatively large and economically important sample. The factors and outcomes we consider are relevant to any audit firm, but it is possible that our results do not generalize to smaller audit firms. For example, if issues such as access to central resources create substantial differences in outcomes between small and large audit firms, then our sample will not represent the experience at smaller firms. Second, our time period, corresponds to the period of data availability, which encompasses the SOX reforms, the beginning of PCAOB regulation and inspections, and the global financial crisis. We urge caution in extrapolating results to any other time period. The layout of the paper is as follows. Section 2 discusses the motivation for our hypotheses, section 3 presents the methodology, and section 4 presents our findings. Section 5 concludes with an overview of our findings and implications for audit firms, regulators and future research. 2. Hypothesis development We build on a large literature of research on audit quality and its determinants. DeFond and Zhang (2014) provide a thorough review of the audit quality literature, in which they call 4

6 specifically for more research on the role of auditor and client competency in driving audit quality. Our proposed contributions to this literature are: first, to focus on the audit office as the unit of analysis; second, to provide a novel and strong control for the audit risk inherent in an office s client portfolio; and third, to focus on factors that can be controlled by regulators or audit firms, leading to policy implications from our work. Our research question concerns the audit office. Typically, research studying audit office traits uses office characteristics such as number of clients, client influence or office location as test variables, but the unit of analysis in their tests is the individual client. 3 This creates an implicit weighting scheme, in which offices with many public clients receive more weight than those with few public clients. If clients of large offices differ systematically from those of smaller offices in ways not captured in the research design, then their predominance in the sample could create a false impression that office size, and not client traits, determines the outcome. An important distinction of our work is that our unit of analysis is the audit office, allowing us to make clearer inferences about the role of office traits, after controlling for the office s client mix. Our measures of audit failure are outcome-based. The first outcome, a financial statement restatement, occurs when a client s financial statements contain misstatements that the auditor either does not detect, or detects but allows to go uncorrected and unreported. 4 The second measure, a comment letter from the SEC, arises when the SEC has questions about the financial statements as filed and requests additional information from the client. Client characteristics, 3 Examples include Reynolds and Francis (2001), Li (2009), Francis and Yu (2009), Choi, Kim, Kim and Zang (2010), Francis and Michas (2013), Francis, Michas and Yu (2013), and Whitworth and Lambert In future work, we would like to incorporate process-level measures of audit failure, such as inspection findings based on audit documentation or procedures. 5

7 such as trustworthiness of its managers and systems, along with client incentives to misreport, play a role in the likelihood of each of these outcomes. The issue of auditor detection introduces factors relating to the auditor s competence and diligence, while the issue of correcting or reporting a detected misstatement introduces the auditors own incentives to ensure or breach integrity in financial reporting. Controlling for the risk that a client will misreport is not straightforward, in principle or in practice. Audit planning should assess client risk, and auditors should then implement procedures to provide adequate assurance given the assessed risk. Hence, finding that audit failures stem from client characteristics does not exonerate the auditor, but rather suggests that deficiencies in audit planning or evaluation of audit evidence contributed to the failure. (PCAOB, 2015) The findings in the literature confirm the importance of controlling for risk. Lawrence, Minutti-Meza and Zhang (2011) propose that differences in client risk of misreporting could account completely for observed differences in outcomes between Big N and non-big N auditors, and conclude that they do. In contrast, using an alternative methodology, Defond, Erkens and Zhang (2014) provide evidence that significant differences in audit quality remain after controlling for client risk. Relatedly, Seidel (2014) finds that accounts with identified weaknesses in the company s internal control procedures are more likely than other accounts to experience restatements, suggesting inadequate audit planning to protect against the risk of misstatement. However, if some auditors insulate themselves by selecting only low-risk clients, then it would be incorrect to credit them with higher quality audits, when they could have lowerquality audits of lower-risk clients. 6

8 We begin by examining whether audit quality varies across Big 4 offices. 5 We focus on Big 4 auditors, for three reasons. First, substantial prior research finds that Big N firms generally have higher audit quality than non-big N firms (e.g. DeAngelo 1981; Louis 2005; Defond et al. 2014; but see Lennox and Pittman 2010; and Lawrence et al for contrary evidence). We wish to avoid the potential confound of differences among audit firms, to focus on traits at the level of the individual audit office. Second, the Big 4 firms audit roughly 80% of public issuers in our data, giving them economic importance. Third, each of these audit firms has between 50 and 90 U.S. offices in our sample period, allowing us a reasonable sample size for our analyses. Thus, Big 4 firms provide us with a large and economically important sample, with relative homogeneity at the audit firm level. We initially test for variation across offices without controlling for the client s risk of misreporting, with the rationale that it is the auditor s responsibility to assess misreporting risk and reflect this in the engagement plan. Thus, following auditing standards should result in failure rates across offices that do not vary with the client s risk of misreporting. Our first hypothesis, stated in null form, is: H1: Offices of Big 4 audit firms do not differ in audit quality. Our second hypothesis incorporates the client risk of misreporting. If audit plans do not completely compensate for the risk, audit quality will still vary. This hypothesis examines whether different offices conduct audits of similar quality, after controlling for the risk of misreporting inherent in the office s portfolio of clients. H2: Offices of Big 4 audit firms do not differ in audit quality, after controlling for clients risk of misreporting. 5 Because Arthur Andersen is included for only two years in the Audit Analytics database, we exclude them from our analysis. 7

9 Our second set of hypotheses focus on auditors willingness to issue unfavorable audit findings. We expect that if incentives and pressures to maintain clients vary across offices, then the willingness to disclose client weaknesses will vary across offices. Because auditor offices may vary in their expertise, it is also possible that variation in use of additional language arises from the ability to detect problems rather than willingness to disclose problems. Stated in null form, we hypothesize: H3: Offices of Big 4 audit firms do not differ in the frequency of their use of additional language in audit opinions. H4: Offices of Big 4 audit firms do not differ in the frequency of their use of additional language in audit opinions, after controlling for client risk of misreporting. Lastly, we expect that offices with a culture that leads to greater disclosure of unfavorable findings are likely offices with auditors who are better able to identify and deter misreporting by their clients, and better able to require correction of misreporting when detected. We therefore hypothesize that clients who do not receive additional language and are audited by an auditor from an office with greater rates of additional language will have lower rates of restatements than clients who do not receive additional language and are audited by offices with lower rates of additional language. The intuition for the prediction is that clients in offices with low rates of additional language are more likely to have an auditor who was unwilling to issue unfavorable findings when detected. The rationale for the test is in the spirit of examining whether the A students of teachers who are tougher graders are stronger students than the A students of teachers who are easier graders. H5: The audit quality outcomes of Big 4 audit offices do not vary with the frequency the office uses additional language. 8

10 3. Methodology In this section, we first describe our method of measuring and controlling for client risk, by constructing a synthetic office for comparison. We then describe our model for hypothesis tests Measuring client risk of misreporting For non-financial firms, we identify a client s risk of misreporting based on the wellknown construct of abnormal accruals. If we restricted attention to clients for whom we can use the accruals-based measure of misreporting risk, the resulting client lists would be substantially incomplete for many audit offices, because accruals models are unsuitable for financial companies. We therefore employ alternative measures in the financial sector, based on external risk assessments, as discussed further in section For all firms, financial and non-financial, our approach is to model the risk using a dependent variable that does not directly measure audit failure, but is associated with the risk of client misreporting. We model this risk using factors that act as ex ante observable signals of or incentives for misreporting, avoiding factors that may be direct indicators of misreporting Client risk for non-financial firms For non-financial client firms, we first model the likelihood of misreporting using abnormal accruals, a standard measure of unexpected, often suspicious financial reporting (Jones 1991; Dechow et al. 1995; Dechow and Dichev 2002, McNichols and Stubben 2014). For this first stage, our dependent variable relates to the client s reporting quality, but it does not measure audit failure directly, to separate these two constructs. Following McNichols and Stubben (2014), we estimate the following accruals model across firms, within Fama-French12 (FF12) 6 Appendix 1 provides details of our variables, their measurement and our data sources. 9

11 industry and year. 7 The model residuals define abnormal or discretionary accruals (industry subscripts on coefficients suppressed): Acc!" = α!! + α!! ΔSales!" + a!! PPE!" + α!! CF!"!! + α!! CF!" + α!! CF!"!! + ε!!", (1) and DA!" = ε!!" = Acc!" (α!! + α!! ΔSales!" + a!! PPE!" +α!! CF!"!! + α!! CF!" + α!! CF!"!! ). (2) Appendix 1 contains definitions of the variables used in the above equations. Dechow, Richardson and Tuna (2003) find that high abnormal accruals can detect extreme earnings management, defined by SEC enforcement releases. We define abnormal accruals in the top three deciles for that industry and year as high-risk. We next estimate a logit model over two-year rolling windows of industrial client firmyears, with an indicator variable for suspicious abnormal accruals, HighRisk, as the dependent variable, and independent variables that are observable ex ante, and that suggest the client firm has an incentive or opportunity to misreport (industry subscripts on coefficients suppressed): Pr HighRisk!" = Λ(β! + β! Ret!"!! + β! Lev!"!! + β! Herf!"!! + β! CVEmpl!"!! +β! ΔEmpl!"!! + β! AQI!"!! + β! StdRet!"!! + β! CVSales!"!!!!"! +β! LTG!"!! + β!" Size!"!! + β!! ROA!"!! + β!" IntCov!"!! +β!" IntCov!"#!"!! + β!" IntCov!"#!"!! + β!" RankDA!"!! +ε!!" ) (3) Specific definitions of the variables in (3) appear in Appendix 1, and our rationales for including them appear below. 7 We exclude FF12 industry number 11 - Money from the accruals analysis, and handle financial firms separately. We thank Kenneth R. French for sharing the industry classification data at: 10

12 Dechow, Sloan and Sweeney (1995) and Dechow, Ge, Larson and Sloan (2011) stress the importance of entity performance as a determinant of earnings management. We therefore include lagged stock returns, Ret, and accounting return, ROA, in our model. We include AQI, a measure of intangible asset intensity, based on Beneish s (1999) idea that intangibles present more opportunities to manipulate than other assets. Beneish (1999) and Dechow, Ge, Larson and Sloan (2011) further identify the importance of unusual growth patterns. Our variables ΔEmpl and CVEmpl capture growth and variation in employees, and LTG captures analysts growth expectations. These prior studies discuss financing concerns as an incentive to misreport; we capture these concerns using leverage, Lev, and three indicator variables for different levels of interest coverage, IntCov. Bentley, Omer and Sharp (2013) cite the emphasis on incentives in SAS No. 99 to motivate including competition in the client s industry in a model of client risk; we capture competition using the 3-digit industry Herfindahl index, Herf, and also incorporate a measure of within-industry variation in sales, CVSales. We follow Hribar and Nichols (2007) in including return volatility, StdRet, as a measure of stock market incentives influencing earnings management. We include RankDA, the lagged value of the rank of the decile of discretionary accruals. We measure all firm-specific traits on the right-hand-side of (3) with a one-year lag relative to HighRisk, and industry traits, which we expect to be relatively stable through time, in the same year as HighRisk. To improve the fit of our client risk model, we follow a step-wise procedure described in Imbens and Rubin (2015) to include higher order terms in our basic model specified in Equation (3). Based on the 15 variables included in the basic model for industrial firms, we create an additional 120 variables (105 interaction terms and 15 squared terms). In the first iteration, we include one variable at a time to the basic model and run 120 regressions. For each of these 11

13 regressions, we calculate the likelihood ratio statistic for the hypothesis that the coefficient on the additional term is zero. We include the covariate with the highest likelihood ratio, provided that it is greater than some pre-set threshold (p-value (LR) < 0.01). We then repeat the process with the remaining 119 variables. The LR statistic converges over successive iterations. As a result, we augment our basic model for industrial firms with 18 variables selected by the algorithm, which increases the mean pseudo R-squared of the risk model for industrials from 10.9 to 12.9%. We follow the same procedure for our models for banks and non-banking financial institutions, which we describe in the next section. The procedure adds seven and three non-linear terms, respectively to these models. The mean pseudo R-square for banks increases from 15.4% to 37.8%, and for non-banking financial firms from 37.2% to 72.5% Client risk for financial firms We retain banks (2-digit SIC code = 60) and other non-bank financial companies (2-digit SIC codes from 61 to 67) in our audit client samples by using alternative measures of HighRisk, as described below. We also modify (3) for these financial firms, omitting from the right-handside: the intangibility measure AQI, the IntCov indicators, and the industry variables Herf and CVSales. The first two variables are not meaningful for financial firms, while the industry variables have no variation within a single industry. Finally, for banks, we add Capital Ratio to control for regulatory risk. To construct the HighRisk variable for banks, we use the estimated CAMELS score from Veribanc, a private company that provides bank safety ratings. The CAMELS rating is a confidential score assigned by banking regulators, ranging from 1 (safe) to 5 (high risk), with scores of 3 or greater considered risky. Veribanc, sourced from SNL Financial databases, claims 12

14 that its estimated CAMELS rating is close to the original. 8 We create an indicator variable that takes the value one if the bank s Veribanc rating is 3 or greater, and zero otherwise, which classifies roughly ten percent of our sample banks as HighRisk. We repeat the methodology for non-bank financial institutions, using long-term S&P issuer ratings, sourced from Compustat, to measure client risk. Our HighRisk indicator for these firms takes the value one if the issuer rating is B+ or below, and zero otherwise. This classifies roughly ten percent of non-bank financials as high-risk clients, about the same proportion of high-risk as in banking. As mentioned above, we modify model (3) for non-bank financial firms, omitting AQI, the IntCov indicators, Herf and CVSales. 3.2 Creating the synthetic office The estimated client risk, or fitted value from (3), is our match variable for creating the synthetic office. 9 We use the fitted value because it is the portion of client risk that relates to ex ante observable information. For each client of a given office-year, we search within FF12 industry (for financials: within banks or within non-bank financials) and year for the closest match on client risk, provided the matched client is audited by a different audit firm. To qualify as a match, the two client risk scores, which generally lie between zero and one, must differ by 0.1 or less. We lose fewer than 0.5% of client firm-years due to non-matched clients Tests of H1 and H2 We look at restatements and comment letters as individual indicators of audit failure (i.e., inverse audit quality). Our first analyses examine the distribution of audit quality across audit 8 See: [accessed 9 Mar 2015]. 9 The method is similar to propensity-score matching, but PSM measures the unit s propensity to be a treatment firm in the second stage. In our design, the first-stage likelihood provides us with a comparison group that is matched on a second-stage control variable, rather than on the treatment variable. 13

15 offices, before and after controlling for the audit quality of the synthetic office. We examine whether and how controlling for the client propensity via synthetic offices affects the crosssectional distribution of audit quality. We estimate a fixed effects model of audit quality outcomes for audit office-years to identify variation in audit quality across offices, with restatement rates and comment letters as separate proxies for audit quality. AQ!" = γ! + γ! + ε!!", (4) where j indexes audit offices and t indexes years. The individual office effects provide an estimate of the office s average restatement rate or comment letter rate, after controlling for the year effects in the sample. The base test for H1 using (4) is simply that the office effects add explanatory power to the model. We estimate a similar model that includes the restatement rate (comment letter rate) for the corresponding synthetic office as a regressor, to control for the expected misreporting rate of the office based on its client portfolio: AQ!" = δ! + δ! + a Synthetic_AQ!" + ε!!". (5) The base test for H2 using (5) is that the office effects add explanatory power to the model. If client misreporting risk is a material contribution to audit quality outcomes at the office level, we expect to find a significant coefficient on Synthetic_AQ, the synthetic office restatement (comment letter) rate. In principle, the adjustment could eliminate differences in these rates across audit offices. We perform validity tests to assess the explanatory power of the client risk model, described in section 5. We next test the null hypothesis H3 that audit offices do not vary in their use of additional language, motivated by the argument that culture across offices may influence auditor 14

16 willingness to report unfavorable findings. Our analysis parallels our test of H1, where the fixed effects model offers a test of differences across offices. AL!" = γ! + γ! + ε!!", (6) where j indexes audit offices and t indexes years. The individual office effects provide an estimate of the office s average additional language rate, after controlling for the year effects in the sample. The base test for H3 using (6) is that the office effects add explanatory power to the model. Our test of H4 includes a control for the use of additional language by including the rate of additional language used by the synthetic office of clients matched on misreporting risk. AL!" = δ! + δ! + a Synthetic_AL!" + ε!!". (7) The base test for H4 is that δ!, the office effects, remain significant after including Synthetic_AL!" to the model. 10 Lastly, we examine whether differences across offices in willingness to issue unfavorable findings are a signal of higher audit quality. We conjecture that this could be true if auditors who more frequently issue unfavorable findings are more diligent or competent in identifying areas of concern, or have greater integrity in reporting concerns that arise in the audit. Our test of H5 partitions offices into five groups, based on the magnitude of the office mean rate of additional language from equation (7), after controlling for Synthetic_AL. We then examine the restatement (comment letter) rates of client-years within each quintile that did and did not receive additional language. We hypothesize that if greater use of additional language is associated with higher audit quality, the clients in offices with higher rates of additional language 10 Our tests to date estimate the match rate for additional language based on the synthetic office s rate of additional language, where the synthetic office is matched on client risk of misreporting. Because additional language aims to draw investor attention to areas that could cause greater risk of misreporting, we believe this is a reasonable approach. 15

17 who do not themselves receive additional language are less likely to restate or receive a comment letter than clients in offices with low rates of additional language. We focus on clients who do not receive additional language in the current client-year to avoid confounding our tests with the consequences associated with additional language in a given year, such as higher restatement rates. We aim to identify tougher audit offices through their use of additional language, and to test whether the clients of these audit offices have better audit outcomes. 4. Sample and Data Description Our primary unit of analysis is an individual Big 4 audit office in a particular year or short span of years. We have viable data on audit offices only dating from the early 2000s, so our study is limited to the period from 2000 onward. Table 1 Panel A shows that we start by gathering data on potential clients from Compustat and CRSP. We require that the client firmyear has traded common equity, as well as certain baseline financial data, yielding 67,885 firmyears for 8,943 firms within Restricting attention to firm-years that Audit Analytics (AA) identifies as clients of Big 4 firms reduces our sample to 36,551 firm-years and 5,234 firms, audited by 357 Big 4 offices. We can measure client risk and generate matches for 36,433 firm-years, representing 5,231 firms audited by 357 Big 4 offices. Of these matched observations we find comment letters for 9,612 firm-years and 3,095 firms, between 2004 and We find additional language in the audit opinion for 19,208 firm-years and 3,755 firms, after excluding additional language that relates to accounting principles changes, following Czerney, Schmidt and Thompson (2014). We find restatements for 3,833 firm-years and 2,169 firms. Table 1 Panel B describes our further selection of the restatement observations. To focus on restatements of audited amounts, following Hennes et al. (2014), we remove restatements 16

18 related to SAB 108, the SEC's 2005 letter regarding leases, and retrospective adjustments for consistency. 11 Finally, to avoid double-counting, we exclude subsequent restatements of a previously restated year, leaving us with 3,336 restatements by 2,114 Big 4-client firms with 6,391 firm-years. Table 2 Panel A reports on the distribution of restatement rates across Big 4 audit offices, by year. Across all years, the average office-year had restatements by 17.5% of its clients and the median office had restatements by 12.5% of its clients. Average restatement rates per year increase from 11.8% in 2000 to 20.3% in 2004, decline back to 13.5% in 2007 and climb to 24.1% by The data for the last years in our sample period may in fact understate the ultimate rate of restatement, because restatements can be announced years after the report date to which the restatement applies. The median rate is generally lower than the mean, and in fact is zero in The rightmost two columns report the proportion of offices in a given year with no restatements, and the proportion that maintain that status cumulatively from the start of our sample period. On average 40.34% of offices do not have a restatement in a given year. Over the entire sample period, 15.82% of offices experience no client restatements. Table 2 Panel B reports similarly on the distribution of comment letters across Big 4 audit offices, by year. Not surprisingly, these are substantially more frequent than restatements, which indicate material misstatements of the financial statements. Across all years, the SEC issued comment letters to 38.9% of clients for the average audit office and 37.5% of clients in the median office. Average comment letters per year increase from 25.9% in 2004, the first year in our data, to 49.3% in 2009, then decline to 25.9% by Median rates generally are similar to the means, indicating that a sizable fraction of clients receive comment letters. The rightmost 11 FIN 48 restatements, while in the database, are removed by an initial requirement for restatement start and end dates. 17

19 two columns report the proportion of offices in a given year with no comment letters, and the proportion that maintain that status cumulatively from the start of our sample period, and the proportion with that status in the prior three years. On average only 18.5% of offices do not have a client receiving a comment letter in a given year, and over the full sample period, only 5.28% of offices have no clients who have received a comment letter. Table 2 Panel C reports on the rates of additional language across Big 4 audit offices, by year. 12 Across all years, the average office rate (median office rate) of additional language in opinions is 49.7% (50%). The average percent of clients with additional language ranges from 41.8% in 2000 to 54.5% in Median rates generally are similar to the means, indicating that a sizable fraction of clients receive additional language. The rightmost two columns report the proportion of offices in a given year that do not use additional language, and the proportion that maintain that status cumulatively from the start of our sample period. On average only 15.7% of offices do not use additional language for any of their public clients in a given year. Over the entire sample period, only 7.76% of offices did not use additional language for any of their clients. Thus, most offices use additional language for at least some of their clients in some years. 5. Estimation and Results In this section, we summarize the stages of estimating client risk of misreporting, leading to our formation of synthetic offices of matched clients. We then provide an analysis of the validity of our risk model, and finally assess our five hypotheses. We provide descriptions of variables relevant to each stage when we reach that stage. 5.1 The accruals model for non-financial firms 12 For more detail about the types of additional language reflected in audit opinions, see Appendix 2. 18

20 For non-financial client firms, we measure client risk based on high abnormal accruals, as defined in equations (1) and (2). Table 3 Panel A displays information about the distributions of variables in the accruals model (1), along with DA, the residual from the model. We estimate (1) within industry and year. Table 3 Panel B summarizes the mean coefficient estimates across years, by industry. Panel B shows that the industries vary in number of firms available for estimation from a low of 1,335 across all years in industry 5 Chemicals, to a high of 11,373 in industry 6 Business Equipment. The model s explanatory power is roughly consistent with prior work, with average adjusted R 2 ranging from 14% in industry 6, to 61% in industry 8 - Utilities. Overall, the model behaves as expected and documents similar coefficients to those in prior literature, such as Dechow, Ge, Larson and Sloan (2011) and Bentley, Sharp and Sharp (2013). 5.2 The client risk model Table 4 provides information about our client risk model (3), which we estimate by year to ensure our risk controls do not contain information that was not observable at the time of the audit. For non-financial firms, we define the indicator variable HighRisk using the top three deciles by industry of DA, the residual from the accruals model discussed in section 5.1. To retain financial clients in our analysis, we use alternative HighRisk measures: for banks, an estimated CAMELS ratings of 3 or greater, and for other financial firms, an S&P issuer rating of B+ or lower. Panel A of Table 4 describes the input variables to (3) separately for industrials, banks and other financial firms. The distribution of interest coverage, IntCov, is both highly skewed and highly variable. For this reason, we use three categorical variables in our analyses: IntCov high = 1 if IntCov is at least 2.0; IntCov low = 1 if IntCov falls between 0.0 and 2.0; and 19

21 IntCov neg = 1 if IntCov is negative. These variables, along with the intangibility measure AQI and the lagged decile of discretionary accruals, RankDA, are relevant only for industrial firms. Table 4, Panel B shows the logit estimation of the client risk model (3) for industrial, bank and non-bank financial firms, where the dependent variable is HighRisk. We estimate the model using the prior year s (t-1) data to predict the risk for year t. The coefficients for this model are then applied to year t data to calculate the year t risk measure. Although the estimates for financial firms are not directly comparable to those for industrials because of different dependent variables, the model fit appears reasonable in all cases. Panel C of Table 4 shows year-by-year distributions of our client risk measure, the fitted values from model (3). Our intention is to capture ex ante client risk of misreporting, hence we use the portion of HighRisk that can be explained by ex ante observable traits. The fitted variables give us a continuous measure to use as a match variable for constructing the synthetic office. One point of interest is the rise in risk scores for banks, and to a lesser extent for other financial firms during the financial crisis, along with increased cross-sectional variation in scores. Because the industrials measure of high risk is defined by the top 3 deciles of abnormal accruals for the industry and year, these average risk measures do not vary across years or industries. 5.3 Controlling for client risk by constructing a matched synthetic office We use the scores developed in section 5.2 as the match variable to construct a synthetic office. Specifically, for each audit office, we form a synthetic office portfolio of clients matched by industry, year and risk score, but audited by a different audit firm. This allows us to control for client risk, to test whether variation in quality across offices remains after that control, and to examine the role of other factors in explaining that variation. 20

22 Table 5 presents evidence of the explanatory power of the synthetic office for audit office outcomes. Panel A presents the estimation results for the models with dependent variables of restatement rates, comment letter rates and additional language rates. The findings indicate that the synthetic office controls created by matching clients based on the risk model are significant at the 0.06, 0.04 and levels respectively. Each of the models has substantial explanatory power for office level rates of audit outcomes, ranging from 35.44% for restatement rates to 63.62% for additional language rates. The explanatory power of the models and the significance of the risk-based synthetic office rates support their use in controlling for client risk. Panel B of Table 5 examines the efficacy of assembling the synthetic office based on the risk measure, relative to a simple industry-year match or a random match. For brevity, this panel reports only the statistics for the synthetic office variable, although in all cases a full model similar to that in Panel A underlies the statistics. We present the significance of the Type III (regression) sums of squares for the synthetic office rates using alternative match variables. For ease of comparison across risk measures, in the first row we repeat the statistics for the model in Panel A. For each dependent variable, namely rates of restatement, comment letters and additional language, the synthetic office based on the risk model is significant at conventional levels, though more so for additional language than restatements. The F-values for the synthetic office based on industry and year matches are lower, and only significant for additional language. The F-values for the synthetic office based on random matches are also less significant than those based on the risk model match. Panel C presents the significance of the Type I (sequential) sums of squares for the synthetic office rate variables, when synthetic office is first in the model. The risk-based matches dominate the industry and year matches, except for restatements. The random match is 21

23 inferior to both the risk-based match and the industry-year match, as expected. These findings support the use of the synthetic office to control for client risk, and the risk model as the method for selecting client matches for the synthetic office. Table 6 presents the estimation results underlying the tests of our first four hypotheses. The F statistic testing whether restatement rates differ significantly across offices when we control for year effects but not for client risk, appears in the first row and column. The F value 4.69 is significant with a probability value less than , and indicates that offices differ in restatement rates after controlling only for differences in restatement rates across years. The second row documents that significant differences across offices remain after controlling for client risk using the risk model. In fact, the F value barely changes, indicating that the synthetic office-year rate is largely uncorrelated with the means across offices. The third and fourth rows present the office-effect F value for the model based on industry and year matches and on random matches, respectively. The difference across offices remains significant, highlighting that the mean rates across offices are largely uncorrelated with the synthetic officeyear rates. The findings indicate we can reject the hypothesis that offices do not vary in restatement rates, both before and after controlling for client misreporting risk. Our findings in Tables 5 and 6 for the comment letter proxy for audit quality parallel those for restatements. Panel A of Table 5 shows both that the synthetic office rate based on risk-model matches significantly explains comment letter rates across offices, and that significant audit office effects remain after controlling for this match. Panel B documents that the synthetic office based on risk model matches has more explanatory power than the two other synthetic office matches. Panel C shows that both the risk match and industry and year match provide some explanatory power when evaluated first, but the client risk model explains more variation. 22

24 Table 6 documents that average comment letter rates differ significantly across offices, again with little variation in the magnitude of the F statistics across models. The findings support rejection of H1 and H2 and document that before or after controlling for client risk, significant differences across offices remain in comment letter rates, our second proxy for audit quality. The third columns of Table 5 and Table 6 present estimation results and statistics for equations (6) and (7), for which the dependent variable is additional language in the audit opinion. The findings parallel our findings for our two audit quality measures: the synthetic office rate has a significant positive association with use of additional language. Even so, we reject that the rates of additional language across offices are equal, both before and after controlling for client misreporting risk. Finally, we present our tests of H5, that offices with greater use of additional language are less likely to provide low audit quality. We separate client-years that did not receive additional language from those that did within each year. We assess audit office propensity to issue unfavorable reports through the rate it provides additional language to its clients, controlling for year effects and the additional language rate of its synthetic office. We use clientyears without additional language to gauge whether that propensity is reflected in overall higher quality audits. Partitioning on quintiles of office mean effects for additional language estimated in equation (7), Table 7 Panel A shows that restatement rates are significantly lower for offices with high additional language for client-years that receive additional language and are significantly higher for those without additional language. While the findings for client-years receiving additional language are consistent with our expectation, those for client-years without additional language are not. If lower rates of additional language from the auditor reflect either less willingness to disclose unfavorable findings or less competence in identifying contexts 23

25 where it is needed, then we would expect higher restatements for the clients of these offices. The findings therefore do not allow us to reject H5 that additional language is positively associated with audit quality. We find evidence instead that higher use of additional language is associated with higher restatement rates, our proxy for lower audit quality. Table 7 Panel B shows the comment letter rates across offices ranked as in Panel A, by quintiles of the office mean effects for additional language estimated in equation (7). The findings indicate that comment letter rates do not vary across offices with low to high additional language, but rather are fairly close to 40% for all client-years. Our next steps will be to further refine the tests of H5 and provide further context for how to interpret the findings. The finding that restatement rates are lower in lower additional language offices could reflect many possibilities. First, our proxy for additional language may be too noisy to capture circumstances where auditors show greater competence or willingness to reveal negative findings. A possible step is to refine the additional language measure to include more extreme cases of negative findings. Second, the lower restatement rates in offices with less additional language may reflect underreporting of restatements. An assumption in using restatement rates as a proxy for audit quality is that it reflects weakness in auditing. If weaker auditors are less willing to require a restatement than stronger auditors, then we are less likely to see our hypothesized higher restatement rate for weaker audit offices. Third, the lack of variation in comment letter rates across offices with more vs. less additional language may indicate the SEC does not focus on misreporting risk. If this is true, then comment letter rates are not likely to be a powerful proxy for audit quality. Alternatively, our use of all comment letters may be too noisy, and a more precise measure may be constructed by focusing on the types of comments made in the comment letters. Lastly, the findings may reflect a disclaimer 24

26 effect, that by highlighting a risk to investors, the auditor or investors perceive that the auditor has less responsibility, perhaps due to reduced litigation risk. This argument is consistent with Aobdia s (2015) finding that the association between going concern opinions and Part I inspection findings is positive, which does not support the argument that going concern opinions are indicative of higher quality audits. 6. Summary and conclusions This study examines variation across offices in audit quality, as measured by restatement rates and comment letter rates. We control for the misreporting risk of clients, which varies across offices. An important contribution of our study is the development of a method to control for clients misreporting risk by constructing a synthetic office of clients. Specifically, each office is matched with a synthetic office comprised of the clients of other audit firms, each of which has misreporting risk closest to one of the office s clients. This method allows a control for the office s client characteristics so that we can focus on audit office characteristics. It may also be adapted for analyses of audit firms or partners, and may also be applied to match on other client characteristics. The findings indicate that offices differ significantly in audit quality, controlling for year effects, suggesting that office-level factors that influence audit quality are persistent. Such factors could include office-level expertise, incentives, workload or culture. We find these differences for two different outcome-based measures of audit quality, restatement rates and comment letter rates. We also find these differences both before and after controlling for the rates of synthetic offices comprised of clients with similar misreporting risk. This evidence does not support a conclusion that differences in audit outcomes across offices are attributable to differences in client risk. If confirmed in further investigation, this would suggest that the 25