The Positive Effects of Negative Assurance. Brad A. Badertscher University of Notre Dame. Jaewoo Kim University of Rochester

Transcription

1 ACCOUNTING WORKSHOP The Positive Effects of Negative Assurance By Brad A. Badertscher University of Notre Dame Jaewoo Kim University of Rochester William Kinney University of Texas at Austin Edward Owens* Emory University Thursday, May 25 th, :20 2:50 p.m. Room C06 *Speaker Paper Available in Room 447

2 The Positive Effects of Negative Assurance Brad A. Badertscher University of Notre Dame Jaewoo Kim University of Rochester William Kinney University of Texas at Austin Edward Owens Emory University May 2017 Abstract What is the comparative cost effectiveness of financial statement audit, review, and compilation services of expert accountants hired by the statement preparer to improve financial reporting quality? The answer is important because audit-level services may not be the most cost effective in an era when confirmatory data analytics and inquiries are improving and new standards for compilations may economically substitute as some regulators relax audit mandates. We examine this question using private firm data that varies the level of CPA firm association and assurance for financial statements (company-prepared, compilation, review, and audit). For more than 41,000 U.S. private companies, we examine expected fees for each service level, and the association of service level with financial reporting quality metrics and a measure of lenders perceptions of quality. We find that although reporting quality is generally increasing with assurance levels, financial statement review is a particularly cost-effective service that yields quality metrics that are close to that of an audit. However, we find that audit is associated with a material reduction in cost of debt. Taken together, these results suggest that either cost of debt reflects value dimensions of audits that output-based quality measures do not capture, or lenders may have an inflated perception of the quality-related benefits of financial statement audit, particularly relative to financial statement review. Keywords: financial reporting quality, auditing, cost of debt, Benford s law JEL Classification: G21, M41, M42, M48 We gratefully acknowledge comments from Jerry Zimmerman and workshop participants at the University of Arkansas. We would like to thank PricewaterhouseCoopers and the Mendoza College of Business for financial support.

3 1. Introduction The choice of financial statement assurance levels (and the use of GAAP basis accounting) by private firms is receiving attention from scholars, standards setters, and regulators. For example, in a 2011 report to the FAF, members of the Blue Ribbon Panel stated [m]any Panel members believe that within the U.S. marketplace, significant, unnecessary cost is being incurred for GAAP financial statement preparation and audit, review, or compilation services. 1 Because financial statement assurance (and the use of GAAP) is a voluntary choice by U.S. private companies, a private firm s assurance level will reflect its individual cost-benefit tradeoff (e.g., Dedman et al. 2014). 2 Recent survey evidence suggests that regulators around the world view the benefits of an audit as outweighing its cost more than companies themselves do, with companies perceiving little external benefits associated with an audit (Minnis and Shroff 2017). The Securities Act of 1933 and the Sarbanes-Oxley Act of 2002 assume that financial statement audits (and after 2002, internal controls over financial reporting) lead to higher quality financial statements. However, absent is empirical evidence on the relative informational benefits and costs across the four alternative levels of CPA firm financial statement service and assurance available to private firms: company-prepared (no CPA involvement), compilation (CPA reads for format, but no verification or assurance), review (CPA applies analytical procedures and makes relevant inquiries for moderate assurance, also known as negative assurance ), and audit (CPA assesses misstatement risk and obtains sufficient evidence for high assurance, also known as positive assurance ). Understanding these dynamics is important, 1 A recent working paper notes that almost two-thirds of medium to large private firms studied do not produce audited GAAP-based financial statements (Lisowsky and Minnis 2016). 2 Regarding public firm issuers of securities, the PCAOB recently founded the Center for Economic Analysis to advance understanding of the economic role of auditing in capital markets. 1

4 particularly given the increasing scrutiny placed upon audit cost. In this study, we provide evidence on the associations among assurance levels, financial reporting quality, and service fees. Estimating the link between attestation level and financial reporting quality (and cost) is inherently challenging because U.S. public firms face mandatory audits for annual GAAP-based financial statements. A recent working paper (Botosan et al. 2016) uses public company data to estimate quality differences between each firm s quarterly financial statements (required to be reviewed, but not audited) and its corresponding audited annual financial statements. However, it is unclear whether any resulting differences reflect the public company audit vs. review scrutiny, rather than other factors such as less stringent accounting and internal control requirements for quarterly vs. annual filings. Moreover, because the audit firm reviewing a public firm s quarterly filings also audits the annual filings, it is not clear that a quarterly review is as rigorous as a review of a private firm s annual financial statements. Because of the lack of variation in assurance levels in public company data, several studies use private firm samples to explore assurance level dynamics. There is some evidence using private firm data that audited annual financial statements are of higher quality than nonaudited financial statements (i.e., reviewed, compiled, and perhaps preparation assistance are generally not differentiated). For example, using a sample of U.S. private firms, Minnis (2011) finds that cash flow predictability is higher for audited firms relative to non-audited firms. Also, using a sample of U.K. private firms, Clatworthy and Peel (2013) find that audited financial statements are less likely to be subsequently amended than are non-audited financial statements. To differentiate among and between CPA firm service levels and financial reporting quality, we use Sageworks, which provides private firms annual financial statement data 2

5 classified into four levels of CPA firm service no accounting firm involvement, compilation, review, and audit. 3 We compute two financial statement output-based measures of financial reporting quality: (1) the deviation of a firm-year s financial statement numbers from Benford s Law (e.g., Durtschi et al. 2004; Amiram et al. 2015) (said to indicate fraud), and (2) an estimate of each firm-year s unsigned abnormal accruals (a widely used measure of financial reporting quality). In addition, as a measure of lender-perceived financial reporting quality, we consider a firm s cost of debt for a subsample with relevant data (e.g., Blackwell et al. 1998; Minnis 2011). Any evidence on the magnitude of reporting quality benefits associated with increasing levels of assurance is incomplete without some indication of the relative costs across assurance levels. Sageworks does not provide data on CPA firm fees, so we obtain confidential assurance fee data from a middle-market accounting firm for a sample of clients comparable in size to firms in Sageworks. We then regress audit, review, and compilation fees on client firm size and use the estimated coefficients to compute expected fees for each observation in the Sageworks sample. We find that the cost of compilation, review, and audit for the average-sized sample firm is approximately $9,286, $17,826, and $43,168, respectively. Further, our models suggest that audit cost is most sensitive to firm size, while review cost is relatively insensitive to firm size. We find directionally consistent evidence across our two output-based quality measures that financial reporting quality generally increases with assurance level. However, the biggest quality difference is between compilation and review, with relatively modest additional benefits between review and audit. Specifically, our results imply that reviews have approximately twice the difference in quality compared to compilations, whereas an audit yields quality about 5% greater than that of a review. Because audits are about 2.5 times as expensive as reviews, it is 3 Based on our discussions with Sageworks, the private firms in our sample apply U.S. GAAP. 3

6 plausible that a financial statement review of annual financial statements may be a particularly cost-effective alternative assurance service. As to cost of debt (as a measure of lender-perceived quality) and assurance levels, we find that cost of debt is decreasing with assurance level, where audited financial statements are associated with significantly lower cost of debt than reviewed financial statements. 4 Taken together with results from our output-based measures, this result suggests that either the cost of debt reflects benefits of audits that the two output-based quality measures do not capture (e.g., implicit assurance about internal controls or management attitudes) or lenders have inflated perceptions of the quality-related benefits of financial statement audits relative to reviews. Identification is a key empirical concern, particularly in that firms may choose CPA firm services in response to their extant financial reporting quality. For example, firms with good internal control and high accounting quality within the firm may choose to have financial statements audited. We try to address this concern using two approaches. First, we estimate propensity score weighted regressions for all analyses to help control for selection effects involving multiple treatments. Second, we estimate service level change effects, and find evidence that provides broad support for the inferences in our levels analyses. While the available identification approaches are limited by the scope of the private firm data, these analyses give us some basis to draw directional conclusions. Our study contributes to the literature by documenting how the various levels of assurance (across the entire CPA firm service spectrum) are associated with output-based measures of financial reporting quality. We provide novel evidence that reviews and audits are associated with similar financial reporting quality, and that reviews are associated with 4 We also find evidence that reviewed financial statements are associated with lower cost of debt than compiled financial statements, which contrasts with findings in both Blackwell et al. (1998) and Minnis (2011), who note no difference in cost of debt between reviews and compilations. 4

7 significantly greater quality than compilations. These results further suggest that reporting quality differences between audited and non-audited financial statements documented in prior literature are likely driven by differences between audits and compilations (or company-prepared statements) rather than differences between audits and reviews, both of which provide some assurance that the numbers are about right. Financial statement review seems very cost effective in providing financial reporting quality approaching that of an audit. At the social level, our results have implications for independent audit or review mandates as well as the value of expert accountants compiling financial statements to comply with a mandated financial reporting framework, such as U.S. GAAP. Finally, our results provide a benchmark to evaluate future alternatives that will develop over time, such as reviews that apply confirmatory data analytics as analytical procedures, which may change the precision and economics of the audit/review/compilation markets. The remainder of our study proceeds as follows. Section 2 discusses background literature and technical details on the various levels of financial statement assurance. Section 3 describes our research design. Section 4 outlines our data, sample selection, and descriptive statistics. Section 5 presents our empirical analyses, and Section 6 concludes. 2. Background and Motivation 2.1. Assurance and other financial statement services for private firms In the absence of audited financial reporting mandates for private firms, the American Institute of Certified Public Accountants (AICPA) (the U.S. trade association of CPAs) maintains criteria and standards for less intensive verification services by its members. The standards, referred to as SSARSs (Statements on Standards for Accounting and Review Services), provide guidance and a floor on service quality by its members and also a basis for 5

8 their clients and third-party users of financial information to differentiate reliability of financial statements by service level. Financial statement audits based on auditing standards developed by the AICPA are known since the early 20 th century and, late in the century, AICPA standards were established for financial statement reviews and compilations. 5 These services differ by verification work effort required, level and form of assurance provided, and whether the CPA must be independent of the client firm (e.g., not hold equity in the client). 6 Indeed, the SEC has recently piggybacked off of the AICPA SSARS review standards in implementing its rules under Regulation CF (under provisions of the JOBS Act of 2012), which permits issuers of securities via crowdfunding to provide reviewed, rather than audited, financial statements for certain levels of capital raising. As summarized in Appendix B, CPAs conducting private company audits must be independent and are required to evaluate the company s risks of material misstatement based on understanding the client firm s industry and accounting practices as well as its internal controls, and to evaluate and verify the resulting financial statements using overall confirmatory data analytics applied to account balances as well as selected verification of account balance details when appropriate. The work effort is intended to provide high, but not absolute assurance that the statements are free of material misstatement in applying GAAP. Further, the auditor must be independent and express positively the finding that the financial statements present fairly in applying GAAP. Reviews require independence, understanding the client s industry and accounting, and like an audit, application of confirmatory data analytics and inquires of management. The reviewer does not perform tests verifying details of particular accounts, and thus there is higher 5 AICPA standards also exist for financial statement preparation and basic accounting and bookkeeping, but these services have no reports and thus cannot be tracked on a systematic basis. 6 We denote the practitioner a CPA to avoid confusing calling a reviewer or compiler an auditor. 6

9 risk of failure to detect and material misstatements. The reviewer expresses negative assurance about the findings in the form I am not aware of any modifications needed to make the financial statement comply with GAAP in all material respects. Compilations do not require the CPA to be independent (lack of independence must be disclosed) and the only verification-related work effort required is the auditor must read the financial statements in light of the stated financial reporting framework [e.g., GAAP] and consider whether the financial statements are in the proper form and free from obvious material misstatement. 7 The compilation report will note a lack of independence, and states that the financial statements have been compiled, but not verified and a statement that no assurance is expressed. Thus, there is substantial variation in the work effort, assurance attained and expressed, and independence of the audit firm. It is also apparent that the likelihood of material misstatement depends on both the integrity and quality of internal control of the private company as well as the level of audit firm assurance chosen by the firm and the quality of the audit firm s work whether as auditor, reviewer or compiler Motivation A related line of research documents private firms accounting and audit service choices from the users side, for example, what lenders request from private firms when seeking information and monitoring loans, and how private firms respond (e.g., Minnis and Sutherland 2017; Berger et al. 2016; Lisowsky et al. 2017). This research reveals financial statement-related information and attestation that is desired by lenders, and the resulting supply in a market 7 Guide to Financial Statement Services: Compilation, Review, and Audit, AICPA oadabledocuments/financial-statement-services-guide.pdf 8 We do not consider how variation in the quality of procedures across accounting firms within each attestation service contributes to observed financial reporting quality (e.g., Gaynor et al. 2016). 7

10 without regulation. However, these papers consider neither the costs of attestation nor objective measures of financial reporting quality. Several recent studies find that audited financial statements are associated with higher financial reporting quality than non-audited financial statements (see DeFond and Zhang 2014 for a recent review). Little, however, is known about the impact of negative assurance reviews based on only analytical procedures and inquiries of management on the quality of financial reporting. For U.S. private firms, Minnis (2011) investigates and finds that the predictability of accruals for future cash flows is higher for audited financial statements than non-audited counterparts (i.e., reviews plus compilations). He also reports in a footnote that the predictability of accruals for future cash flows does not appear to vary between reviews and compilations. However, Minnis (2011) uses a relatively narrow sample, and focuses on cost of debt rather than financial reporting quality. Botosan et al. (2016) attempt to examine the effect of financial statement audits versus reviews on the quality of financial reporting by comparing public firms audited annual and reviewed interim (quarterly) financial statements. Using the Financial Statement Divergence (FSD) score based on Benford s law (e.g., Amiram et al. 2015), they find that audited annual financial statements exhibit less error than reviewed interim financial statements. Although Botosan et al. (2016) provide evidence consistent with negative assurance reviews being associated with lower financial reporting quality, it is difficult to generalize their findings to private firms because they compare annual and interim financial statements of public firms. In the U.S., public and private firms face significantly different regulatory and capital market environments (Securities Act of 1933 and the Securities Exchange Act of 1934). Moreover, because the audit firm reviewing a public firm s quarterly filings also audits the annual filings, it 8

11 is not clear that a quarterly review is comparable the review assurance service that is applied to a private firm s annual financial statements as an audit alternative. Foster et al. (2016) investigate the effect of different assurance types (e.g., PCAOB-audit, GAAS-audit, and Review) on abnormal production (a proxy for real earnings management). For a set of both public and private manufacturing firms in the U.S., they find that public firms tend to engage in income-decreasing real earnings management, whereas privately-held firms that receive GAAS-audits tend to engage in income-increasing real earnings management. 9 Reviewed private firms do not seem to engage in real earnings management. Whereas Foster et al. (2016) examine the effects of assurance level on real activity management, we are interested in the effects of assurance level on financial reporting quality. Clatworthy and Peel (2013) examine the impact of audits and governance on quality of accounting information of U.K. private firms, and find that unaudited financial statements are more likely to be amended subsequently than audited financial statements. However, that study leaves unexplored the effects of less than audit assurance levels. More broadly, our paper is related to the literature that investigates financial reporting quality of private firms, in general. Hope et al. (2013) find that relative to public firms, private firms show lower accrual quality and less conservatism. 10 Givoly et al. (2010) focus on firms that do not have public equity, but have public bonds, and find that private equity firms have higher accrual quality and lower reporting conservatism. Both papers leave unexplored the impact of different levels of assurance on financial reporting quality within private firms, which is precisely the focus of our study. 9 GAAS-audit refers to an audit following AICPA standards for private companies rather than use of PCAOB auditing standards, which are required for public issuers. 10 Hope et al. (2013) conduct a sensitivity analysis where they remove all unaudited firms (i.e., company-prepared. compilations, and reviews) from the private sample but they do not analyze the impact of specific attestation levels (company-prepared, compilations, reviews, and audits) on the financial reporting quality for private firms. 9

12 Understanding the role of alternative assurance levels in general, and reviews in particular, is important because reviews may be a cost effective assurance alternative to full audits for both public and private firms. As pointed out above, under Regulation CF the SEC already permits public firms in certain circumstances to file reviewed, rather than audited, financial statements. Further, audits are becoming costlier due to new regulations and the increasing complexity of U.S. GAAP, and the cost effectiveness of reviews will likely be strengthened by advances in data analytics and software (Murphy and Tysiac 2015) Research Design 3.1. Primary empirical specification To examine the association between assurance levels and financial reporting quality, we estimate the following regression model using OLS, with standard errors clustered by firm: FRQ Compile Review Audit Controls it, 0 1 it, 2 it, 3 it, k it, IndustryFE YearFE it,, (1) where FRQ represents one of three alternative measures of financial reporting quality described in the following section. Compile (Review, Audit) is an indicator that equals one if firm i s year t financial statements were compiled (reviewed, audited), and equals zero otherwise, where the omitted category (captured by the intercept) reflects company-prepared financial statement (i.e., no CPA firm involvement). Following prior work (e.g., Minnis 2011; Hope et al. 2013), we include the following set of control variables that are determinants of financial reporting quality for firm i in year t: natural log of total assets (LnAssets); percentage change in sales squared to capture non-linearities in the relation between financial reporting quality and sales growth (Collins et al. 2017) 11 Badertscher et al. (2014) provide some evidence on the relative cost of private firm audits in a study of public debt issuers with private equity. 10

13 (SalesGrowthSq); debt-to-assets ratio (Leverage); the ratio of current assets to current liabilities (CurrRatio); an indicator variable that equals one if the firm is incorporated under Subchapter C of the Internal Revenue Code and equals zero otherwise (CCorp); property plant and equipment deflated by total assets (PPEta); an indicator variable that equals one if total shareholders equity is negative (NegEquity); interest coverage ratio (IntCoverage). In Eq. (1), 1 ( 2, 3 ) measures the effect on financial reporting quality of a compilation (review, audit) relative to companyprepared financial statements Measures of financial reporting quality There is no universally accepted measure of financial reporting quality (e.g., Dechow et al. 2010; DeFond and Zhang 2014). This is perhaps even more the case for private firms in the U.S (e.g., Hope et al. 2013). Therefore, we consider several candidate measures of reporting quality used in extant literature. First, we follow Amiram et al. (2015) and construct the Financial Statement Divergence Score (hereafter referred to as FSD Score). FSD Score is based on Benford s Law, which describes the theoretical distribution of all the first digits that appear in a given population of numbers, and has been shown to apply to a wide variety of data sets, including financial statement data (e.g., Amiram et al. 2015). FSD Score measures the deviation between the realized distribution of first digits in annual financial accounting data (across all accounts) and the theoretical distribution of first digits implied by Benford s Law. Specifically, following Amiram et al. (2015) we measure FSD Score using the mean absolute deviation statistic: K FSDScoreit, ADkit,, TDk / K, (2) k 1 11

14 where AD is the actual distribution of each leading digit in firm i s year t financial statements (i.e., the proportion of numbers in the financial statements that begin with that particular leading digit), TD is the theoretical distribution of each leading digit under Benford s Law, and K is the number of distinct leading digits that appear in firm i s year t financial statements. 12 Lower values of FSDScore are associated with better financial reporting quality. FSDScore has many merits relative to accrual-based measures of financial reporting quality. A key advantage is that this measure is designed to capture both intentional and unintentional financial statement errors, whereas accrual-based measures are more focused on detection of intentional earnings management. Thus, FSD Score particularly suitable for our setting of private firms choice of assurance, as earnings management incentives are likely less acute in private firms relative to public firms, and we are interested in the association between assurance levels and general financial reporting quality (i.e., our interest is not merely in intentional errors). Indeed, several studies highlight the application of Benford s law to auditing (e.g., Nigrini and Mittermaier 1997; Durtschi et al. 2004; Nigrini and Miller 2009). In particular, Nigrini and Mittermaier (1997) show the usefulness of Benford s law in assisting auditors analytical procedures. Also, there is no clear reason why a measure based only on general properties of numerical distributions would be ex-ante correlated with firm characteristics. A disadvantage of FSD Score is that it is admittedly devoid of accounting intuition that facilitates economic interpretation. Accordingly, we also consider a traditional accrual-based measure of financial reporting quality based on estimation of a variant of the modified crosssectional Jones (1991) model. Although accrual-based measures of financial reporting quality have been widely used in the audit literature, their shortcomings do not go unnoticed (e.g., 12 We include all financial statement data items that appear in Sageworks for a given year, which include data from the income statement and balance sheet. The average number of data items for a firm-year is

15 Dechow et al. 2010; Ball 2013; DeFond and Zhang 2014; Owens et al. 2017). For example, it is well known that discretionary accruals estimated using a Jones-model-based approach are correlated with firm economic fundamentals. We acknowledge these limitations, but nonetheless consider the accrual-based measure to provide complementary evidence vis-à-vis FSDScore. Specifically, we estimate the following regression annually for each two-digit NAICS industry with at least ten observations: Accruals 1 it, it, it, it, Assetsit, 1 Assetsit, 1 Assetsit, 1 ROA SalesGrowth 4 i, t 5 i, t i, t Sales Receivables PPE, (3) where Accruals is total accruals, Assets is total assets, ΔSales is annual change in sales in year t, ΔReceivables is the change in trade receivables in year t, PPE equals net property, plant, and equipment in year t, ROA is return on assets in year t, and SalesGrowth is the percentage change in sales from year t-1 to year t. All variables are further defined in Appendix A. The residual from Eq. (3) is a proxy for discretionary, or abnormal accruals (AbAccruals), and is generally interpreted as a measure of directional earnings management. We follow prior literature and consider the unsigned residual ( AbAccruals ) as a general measure of financial reporting quality in our analyses (e.g., Dechow and Dichev, 2002). Whereas both FSDScore and AbAccruals are statistically generated measures of financial reporting quality, we next examine a market-based measure of perceived financial reporting quality as it relates to assurance level. Minnis (2011) provides evidence that private firms with audited financial statements (relative to lower levels of attestation) enjoy lower cost of debt. Further, Minnis (2011) provides evidence that lenders use audited financial statements more intensively in the debt contracting process, relative to non-audited financial statements. Both findings imply that lenders perceive audited financial statements to be of superior quality 13

16 for assessing borrowers. Accordingly, we use a firm s cost of debt (CostOfDebti,t+1) as a measure of perceived financial reporting quality, where we follow Minnis (2011) and measure CostOfDebt as interest expense in year t+1 divided by average debt (Debt) outstanding in year t+1 (where Debt equals total liabilities). As in Minnis (2011), we truncate CostOfDebt at the upper and lower 5% levels, and code CostOfDebt as missing for any firm years where Debt more than doubles or reduces by half relative to the prior year. This leads to a reduced sample size for our CostOfDebt analyses Identification A major concern with interpreting the results from estimating Eq. (1) is endogeneity. Specifically, firms may choose their level of assurance in response to their extant financial reporting quality. For example, firms with good internal controls may choose to have financial statements audited, consistent with the bonding hypothesis (e.g., Watts and Zimmerman 1986). We address this concern using two approaches. First, we estimate propensity score weighted regressions (e.g., McCaffrey et al. 2013; Austin et al. 2015; Guo and Fraser 2015). A popular method of addressing self-selection is propensity score matching, which is particularly convenient for studies with a dichotomous treatment variable. As we have four levels of treatment, we rely on propensity score weighted regressions that account for multiple treatments, which has the added advantage of preserving all observations in the sample for use in the regression estimation. The estimation is conducted in two steps. First, we estimate the following multinomial logistic model, which regresses a variable that captures the level of assurance on the same set of determinants used in Eq. (1): Assur LnAssets SalesGrowthSq Leverage CurrRatio it, 0 1 it, 2 it, 3 it, 4 it, CCorp PPEta NegEquity IntCoverage 5 it, 6 it, 7 it, 8 it, IndustryFE YearFE it,, (4) 14

17 where Assur equals 1, 2, 3, or 4 when firm i s year t financial statements are company-prepared, compiled, reviewed, or audited, respectively, and all other variables are as previously defined. This estimation yields a generalized propensity score corresponding to each of the four levels for each sample observation (i.e., each observation receives a score that indicates the likelihood that the observation has company prepared, compiled, reviewed, and audited financial statements, based on the firm characteristics included in the model). 13 Then, for each observation, we retain only the generalized propensity score that correspond to the actual treatment. For example, for an observation that received an audit, we keep that observation s generalized propensity score for audit. In the second step, we estimate Eq. (1) using the inverse of each observation s generalized propensity score as a sampling weight to mitigate the potential selection bias. Further, to mitigate the impact of outlier weights (i.e., observations with a very high or low generalized propensity score), we normalize each observation s inverse propensity scores by multiplying it times the expected value of being in the respective treatment (a technique referred to as stabilization see Harder et al. 2010). 14 The intuition underlying this approach is that observations for which the choice of assurance level is well-predicted (difficult to predict) based on firm characteristics receive lower (higher) weight. 15 Although this approach helps mitigate the 13 Because there are four levels of Assur, the multinomial logistic model estimates three models (with one Assur level used as the referent group), which leads to 24 estimated coefficients, not including fixed effects (i.e., 8 predictors times 3 models). In this untabulated estimation, 22 of the 24 coefficient estimates are significant at the p<0.01 level, and the pseudo-r-squared is 34.58%. 14 For a sensitivity test, we also trim observations with very high and low inverse propensity scores. 15 To help understand the intuition underlying this inverse propensity-score weighting, consider the case of an experiment with random assignment (i.e., assignment is independent of a firm s propensity to select a particular assurance level). Considering two assurance levels (review and audit), in the experiment a firm is equally likely to be assigned to audit as review, regardless of their likelihood of choosing audit vs. review (i.e., they do not get to choose). This is not the case in our sample. Instead, firms that are innately likely to choose an audit (review) tend to select an audit (review). As a result, our sample is likely overpopulated by observations that self-select an audit (review). We are interested in what would happen to financial reporting quality if firms that are likely to choose an audit (a review) are forced to choose a review (an audit). Accordingly, giving more weight to observations where attestation choice was poorly predicted (based on firm characteristics) gets us closer to random assignment. That is, we really want to place great weight on an observation where firm characteristics predict audit but the firm chooses review instead. 15

18 potential confounding effect of self-selection, it is subject to the same concerns and assumptions inherent in propensity score matching (e.g., Minnis 2011). Specifically, inferences are susceptible to selection on unobservables if the observable variables (included in the model) do not account for all differences in treatment. Second, we examine the association between changes in assurance level and changes in financial reporting quality, which should further help mitigate concerns about reverse causality. Specifically, we estimate the following regression model, with standard errors clustered by firm: FRQ AssurDnAtoR AssurDnOther AssurUpRtoA it, 0 1 it, 2 it, 3 it, AssurUpOther Controls IndustryFE YearFE 4 i, t k i, t i, t (5) where AssurDnAtoR is an indicator that equals one if the firm decreases its assurance from audit in year t 1 to review in year t, AssurDnOther is an indicator that equals one if the firm decreases its assurance in any other combination from year t 1 to t (audit to compilation, review to compilation, compilation to company-prepared, etc.), AssurUpRtoA is an indicator that equals one if the firm increases its assurance from review in year t 1 to audit in year t, and AssurUpOther is an indicator that equals one if the firm increases its assurance in any other combination from year t 1 to t (compilation to review, compilation to audit, company-prepared to review, etc.). Accordingly, 14 - capture the effect of changing levels of assurance on changes in financial reporting quality, relative to those observations with no changes in assurance (as captured by the intercept). We further note that this changes specification mitigates potential confounding effects of time invariant firm characteristics. 4. Data and Descriptive Statistics 4.1. Sample selection We obtain balance sheet, income statement, and assurance level data for private firms 16

19 from Sageworks Inc. 16 Sageworks contains both income statement and balance sheet items (similar to the corresponding data structure in Compustat for public firms), but does not have items from the statements of cash flows. Additionally, basic demographic information such as geographic location and North American Industry Classification System (NAICS) industry codes are provided. Finally, Sageworks provides the level of assurance service chosen by each firm (e.g., audit, review, compilation, company-prepared) for each firm-year of accounting data. The unique combination of financial statement data and levels of assurance allows us to examine the association between financial reporting quality and assurance level. The private firm data span ten fiscal years from 2001 through We have a total of 103,114 firm-years and 41,280 firms with all necessary variables used in our analyses (hereafter referred to as the full sample ). In addition to examining the full sample, we examine a sample comprised of firm-year observations with assets in the five middle asset-based deciles from the full sample (i.e., assets ranging from $1.94M to $7.5M) (hereafter referred to as the sizerepresentative sample ). These five deciles maximize balance in sample coverage of various assurance levels, which helps ensure that our results are not driven by firm size differences. An additional motivation for use of the size-representative sample is that we obtain separate fee data (described below) for audits, reviews, and compilation, and the overwhelming majority of fee data observations likewise falls in this five-decile size range. The size-representative sample consists of 51,557 firm-year observations across 22,782 distinct firms. While we report results 16 Sageworks collects private firm financial data from banks and accounting firms, including large national accounting firms and smaller regional firms. Sageworks collects confidential financial statement information of nonlisted clients of large and regional accounting firms, and sells the data, aggregated by industry and region, with financial tools to its clients who are the accounting firms, banks and other financial institutions. For research purposes only, Sageworks granted some researchers confidential access to the non-aggregated data with firm names removed and unique firm identifiers substituted. For the purposes of this study, Sageworks granted us confidential access to their data. This dataset has recently been used in numerous accounting and finance studies (e.g., Minnis 2011; Badertscher et al. 2013; Badertscher et al. 2014; Hope et al. 2013; Lisowsky and Minnis 2016). 17

20 from our analyses using both the full sample and size-representative sample, we focus most of our discussion on the size-representative sample for brevity. 17 As alluded to above, although we focus on the association between financial reporting quality and attestation level, we would like to also offer some perspective from data on variation in fees across attestation levels. Unfortunately, Sageworks does not include assurance fee data. Therefore, to shed some light on cost, we obtain assurance service fee data from a mid-market accounting firm for a sample of 125 private-company engagements comprised of approximately equal numbers of audit, review, and compilation services. The clients span multiple industries and range in assets from approximately $1 million to $20 million. These fee data, all from year 2015, do not include any financial information for the client firms other than total assets, and do not provide firm identifiers. Because the data are from one audit firm and market area, the fees are presumed to reflect market conditions and decisions of private company client management with respect to the costs and value of the various CPA firm services. Because there is substantial overlap in firm size between our fee data and our size-representative Sageworks sample, comparisons that combine fee evidence are not likely affected by differential size characteristics across the two samples. Table 1 details our Sageworks sample selection process. We follow Minnis (2011) and Badertscher et al. (2013) and exclude Canadian firms, as well as observations with missing financial data or for which the accounting numbers fail to satisfy basic accounting identities (which suggests data errors). We follow Badertscher et al. (2013) and exclude financial firms (finance and insurance industries, NAICS code 52) and utilities (NAICS code 22) because of 17 In addition to examining the full sample, we also examine a constant sample (20,324 firm-year observations) comprised of 5,081 firms that are in the full sample in all years from to ensure that our results are not driven by firms that exit the sample. We choose the constant sample period because that period coincides with the broadest coverage available within the Sageworks data. All results are similar when examining the constant sample. 18

21 their different business models. Finally, we exclude firms with total assets less than $1 million (because we utilize assets as a scalar, values less than $1 million could introduce distortions). Finally, we Winsorize all continuous variables at 1% and 99% Assurance levels by size and industry To better understand how service choice varies by firm size, we partition our full sample into deciles based on total assets. Specifically, Panel A of Table 2 reports the frequency of an audit, review, compilation, and company-prepared financial statements for each asset decile. For instance, in the smallest asset decile (decile 10), which has assets ranging from $1 million to $1.47 million, the percentage of audit, review, compilation, and company-prepared statements is 6.0%, 40.6%, 51.6%, and 1.8% respectively. Moving to decile 5 (firm-years with assets ranging from $4.2 million to $5.5 million), the percentage of audit, review, compilation, and companyprepared statements is 19.5%, 51.1%, 28.3%, and 1.1% respectively. In contrast, in the largest asset decile (decile 1), the percentage of audit, review, compilation, and company-prepared statements is 60.9%, 20.0%, 18.0%, and 1.1% respectively. It is noteworthy that, although size is associated with assurance level choice, there are a significant number of very large firms that obtain reviews and compilations, as well as a non-trivial number of very small firms that obtain an audit. In order to put our sample of private firms into perspective, we obtain a comparative sample of public firms from Compustat that meet the same criteria as our private sample. The median (10 th percentile) audited public firm has total assets of $198 ($3.7) million. Accordingly, although our private sample is considerably smaller in size than a typical public firm sample, our private firm sample contains approximately 23,869 firms (56,072 firm-years) that would be at or above the 10 th percentile of assets in the public firm distribution. 19

22 We next examine assurance level choice by industry, both for our private firm sample and the comparative Compustat public firm sample. As reported in Panel B of Table 2, the largest percentages of private firms are in the Construction and Construction Materials industry, followed by Wholesale. The distribution of public and private firms across industries is generally quite similar. For example, 3.9% of private firms and 2.4% of public firms are in the Food Products industry. Only five industries have a significant difference in the percentage of private and public firms in the industry Healthcare, Pharmaceutical Products, Construction and Construction Materials, Business Equipment, Wholesale, and Retail. Examining the percentage of each assurance level across industries also reveals some interesting results. The average percentage of firm years receiving audit assurance across all industries is 22.6%, yet there exists considerable variation among industries. For example, Retail has only 15.0% of firm-years receiving an audit, while Tobacco Products has 64.7%. Similarly, the average percentage of firm years receiving a review is 44.3%, yet Construction and Construction Materials has 54.5% while Petroleum and Natural Gas has only 9.0%. As another example of industry-level variation, the average percentage of firm years receiving a compilation is 31.8%, yet Restaurants, Hotels, Motels has 47.6% while Tobacco Products has only 17.6%. To our knowledge, this is the first study to document details on the different assurance levels for a large sample of U.S. private firms. The data indicate that, although audit is the focus of most academic research, the percentage of private firms that do not get audited is almost 80%, which highlights the importance of understanding the benefits, costs, and tradeoffs of different assurance levels. 20

23 4.3. Descriptive statistics Panel A and Panel B of Table 3 present descriptive statistics for the variables used in our study for the full sample and size-representative samples as a whole, and within each assurance category. By construction, the size-representative sample has lower mean Assets than the full sample (5.369 vs ). Descriptive statistics of all other variables are of the same general magnitude across samples, which highlights that the basic relations in the data are not driven by firms in the tails of the size distribution (which are excluded from the size-representative sample). Accordingly, we will focus our discussion of the descriptive statistics on the full sample for brevity. The univariate statistics reveal a fairly systematic pattern in the data. Specifically, mean FSDScore increases monotonically from the audited statements (0.052) to the company-prepared statements (0.056). For AbAccruals, the mean and median across all observations is and 0.069, respectively, which is similar in magnitude to corresponding data from public firm studies (e.g., Owens et al. 2017) and other private firm studies (e.g., Hope et al. 2013). Audit firm years have the lowest mean AbAccruals (0.096), while firm-years with no assurance (Company) have the highest (0.117). Cost of debt (CostOfDebt) exhibits a similar pattern, in that it is lowest for the Audit firm years (0.053) and highest for Compile firm years (0.147). The overall mean and median CostOfDebt is comparable to descriptive statistics in Minnis (2011). Turning to the control variables, we first note that the firms in our sample are of sizeable magnitudes and are not mom-and-pop firms, with mean (median) assets of $9.6 ($4.1) million. Second, as expected, firms that are audited are the largest, on average, among the four assurance levels. Specifically, mean Assets across audited, reviewed, compiled, and company-prepared firm years is $19.8, $6.4, $6.8, and $9.0 million, respectively. Aside from size, we do not see any 21

24 particular patterns in control variables across assurance categories, again reinforcing the particular need to control for firm size in our multivariate regressions. The average amount of interest bearing debt across all observations is $3.47 million, with audited firm years having the highest levels ($6.33 million). The mean liabilities to assets ratio (Leverage) is 28%, while the mean current ratio (CurrRatio) is 2.4. Overall, the control variables have similar distributional properties as those reported in Minnis (2011) and Badertscher et al. (2014). Panel C of Table 3 displays the Pearson and Spearman correlations for the full sample. Consistent with the descriptive statistics in Panels A and B, higher levels of assurance are generally associated with higher financial reporting quality metrics. For example, Audit is negatively correlated with FSDScore, AbAccruals, and CostOfDebt, where the correlation with CostOfDebt is particularly strong ( 0.42) relative to the other quality metrics. On the other hand, Compile is positively associated with FSDScore, AbAccruals, and CostOfDebt. Audit is positively correlated with Assets, SalesGrowth, and Debt; while Compile and Review are negatively correlated with these variables. These findings suggest that firms that are larger and have more growth opportunities tend to obtain higher assurance. Audit is negatively associated with CurrRatio and PPEta, consistent with firms having higher levels of liquid and tangible assets being less likely to receive higher assurance. Overall, the correlation patterns are consistent with prior work (e.g., Minnis 2011; Badertscher et al. 2014). 5. Empirical Results 5.1. Assurance fees Before examining the associations between financial reporting quality and assurance levels, we examine our limited data on fees, as doing so better allows us to put our financial reporting quality analyses into perspective. While we are not able to obtain assurance fee data for 22

25 our Sageworks sample, we are able to shed some light on the relative costs of an audit, review, and compilation using confidential audit fee data we obtained from a mid-market accounting firm that focuses on providing assurance services for private firms similar to those in Sageworks. 18 Although the fee data are limited in scope, they span all assurance levels across a wide range of firm sizes. The private firms for which we have audit fee data range in size from $500,000 to $20 million in assets, and include 33 compilations, 44 reviews, and 48 audits. Figure 1 provides a graphical representation of how fees vary by firm size (quintile) for each level of attestation. Specifically, Figure 1 provides the average asset amount (right hand scale) and fees (left hand scale) for each asset quintile, where quintiles are formed within each assurance level subsample. A salient feature of Figure 1 is that audit fees increase materially with size. For example, for audit firms in size quintile 3 (average assets of $14,293,000) average fees are $48,347, while for audit firms in quintile 5 (average assets of $94,553,000) average fees are $82,586. In contrast to audit fees, review fees demonstrate less variation across asset quintiles, suggesting a fixed review fee component, with incremental fees that do not vary much with client firm size. For example, for review firms in size quintile 3 (average assets of $6,767,000) average fees are $19,026, while for review firms in quintile 5 (average assets of $23,387,000) average fees are $21,660. Finally, compilation fees appear to be more sensitive to firm size than review fees, but less so than audit fees. In order to better understand the relative costs of assurance level after controlling for the effect of firm size, we estimate the following model using the 125 observations of available fee data: 18 The audit fee data come from fiscal year 2015, so it provides a recent view of the fees for various attestation levels. 23

26 AssurFees Review Audit Assets i 0 1 i 2 i 3 i Review* Assets Audit * Assets, 4 i 5 i i (6) where AssurFee is the fee associated with the client engagement, Compile is the omitted assurance category (captured by the intercept), and other variables are defined as previously. As we note above, we do not have fee data for our Sageworks sample. Therefore, we also use the estimation results from Eq. (6) to compute implied fees for our Sageworks observations, which allows us to provide an integrated discussion of the costs and benefits of different levels of assurance. Panel A of Table 4 presents the results from estimating Eq. (6). Turning first to Column (1), the coefficient on Review is $9,622 while the coefficient on Audit is $30,847, where an F-test reveals that Review and Audit are statistically different from each other at the 1% level. Together, these results indicate that, after controlling for the effect of firm size on assurance fees, financial statement reviews are $9,622 more expensive than compilations, and audits are $21,225 more expensive than reviews, on average. Column (2) presents results from the full Eq. (5) specification, which captures any interactive effects of size on the cost of each assurance service. The coefficients on Review and Audit are of the same magnitude as reported in Column (1). The interaction between Review and Assets is significantly negative, indicating that relative to compilation, the cost of a financial statement review is less sensitive to firm size, consistent with Figure 1. The interaction between Audit and Assets is negative but insignificant, suggesting that the relation between audit fees and firm size is of similar sensitivity to compilation and firm size, which is again consistent with the observed pattern in Figure 1. In Panel B of Table 4, we report average implied assurance costs for our Sageworks sample, using estimated coefficients from Eq. (6) applied to the Sageworks observation data on firm size and attestation level. The results imply that the average costs across assurance levels in 24

27 our Sageworks data are $9,286 for a compilation, $17,826 for a review, and $43,168 for an audit. On average then, implied audit fees are about 150% higher than review fees, while review fees are about 90% higher than compilation fees. This provides some initial perspective on the relative costs of each attestation level. Although these estimates are formed using underlying data from only one accounting firm for one year, we are unaware of any other studies that provide insight into the relative costs of attestation levels for private firms. This evidence therefore should help scholars, standards setters, and regulators better assess the costs relative to the benefits of each service level. In the sections below, we turn our attention to the foundational benefits that ostensibly follow from higher levels of attestation financial reporting quality Assurance levels and financial reporting quality Tables 5 and 6 present results from estimating Eq. (1) separately for both the full sample and the size-representative sample using the financial statement deviation score (FSDScore) and unsigned abnormal accruals ( AbAccruals ) as measures of financial reporting quality, respectively. Within each table, we report results from both OLS regression (Columns 1 and 3) and propensity-score-weighted regression (Columns 2 and 4). We generally focus our discussion of the results on the OLS full sample results (i.e., Column 1), as the results are remarkably similar across estimation approaches and samples. We note that the coefficients on Compile, Review, and Audit capture incremental effects of the respective assurance level relative to company-prepared financial statements, which is captured in the intercept (included in the estimation but unreported, as our use of fixed effects makes the intercept difficult to interpret on a stand-alone basis). As reported in Table 5 (where financial reporting quality is proxied by FSDScore), the coefficients on Compile, Review, and Audit are all significantly negative (p-value < 0.01), with a 25

28 monotonically increasing coefficient magnitude as assurance level increases. This provides evidence that CPA firm involvement of any kind is positively associated with financial reporting quality, after controlling for numerous alternative determinants of financial reporting quality. Turning to a pairwise comparison of assurance levels, the coefficients on Compile and Review are and , respectively, where an F-test indicates that the two coefficients are significantly different from one another (p-value < 0.01). This suggests that a financial statement review (i.e., moderate assurance) is associated with better financial reporting quality than a financial statement compilation (i.e., no assurance), as FSDScore is decreasing in financial reporting quality. Next, turning to the comparison between a financial statement review and an audit, an F-test of the coefficients on the two variables ( vs , respectively) indicates that the coefficient estimates on Review and Audit are significantly different from one another, which indicates that a financial statement audit is associated with better financial reporting quality than a review. Because Compile, Review and Audit are indicator variables, simple ratios of the coefficient estimates provides a sense of the magnitude difference in the relative effects of assurance levels on financial reporting quality. Specifically, these estimates suggest that, controlling for other determinants, the effect of a review on financial reporting quality is twice that of a compilation (i.e., vs ), whereas the effect of an audit on financial reporting quality is merely 4.9% greater than that of a review (i.e., [ ]/0.0040). Column (2) reports remarkably similar results using the propensity score weighted regression approach, which helps mitigate reverse causality concerns. Thus, even after addressing the potential self-selection problem, the benefits of an audit relative to a review seem quite modest. Inferences in the remaining columns (using the size-representative sample) are likewise 26

29 consistent, where all specifications suggest that the effect of a review is at least twice that of a compilation, but the effect of an audit on financial reporting quality is less than 10% greater than that of a review. Table 6 presents results using unsigned abnormal accruals ( AbAccruals ) as the measure of financial reporting quality. Consistent with Table 5, the coefficients on Compile, Review, and Audit are all significantly negative (p-value < 0.01), which again provides evidence that CPA firm involvement of any kind is positively associated with financial reporting quality. However, whereas F-tests indicate again that review leads to significantly better financial reporting quality than does compilation, F-tests indicate that there is no difference in financial reporting quality between financial statement reviews and audits (which is somewhat consistent with the economically small but significant difference in financial reporting quality effects between reviews and audits when examining FSDScore). In the previous section, we infer that audits are on average about 2.5 times as expensive as a review. It therefore seems reasonable to interpret the evidence in tandem as suggesting that a financial statement review is a particularly costeffective service that yields quality that is not too far removed from quality achieved by a financial statement audit Assurance levels and cost of debt The preceding output-based measures of financial reporting quality are designed to capture the aspect of financial reporting quality that are targeted by confirmatory data analytics and inquires, and may not capture other dimensions of quality that flow from financial statement audits, such as those that come from improvements in internal control procedures. Therefore, we next examine the effect of assurance level on the cost of debt as a proxy for lenders perceptions 27

30 of financial reporting quality, although data requirements for this analysis substantially reduce our sample size. Table 7 presents the results of estimating Eq. (1) using the cost of debt (CostOfDebt) in year t+1 as a proxy for (lender perceived) financial reporting quality. We begin in Column (1) by estimating the effect of financial statement audit on cost of debt, relative to all less-than-audit assurance levels combined. As reported, a financial statement audit is associated with a 49 basis point reduction in interest rate, relative to non-audited firm years. This finding is consistent with Minnis (2011), who found that the effects of an audit (versus all non-audit cases) ranged between 25 and 105 basis points, depending on specification. In Columns (2)-(5), there are negative and significant coefficients on Review and Audit (consistent with an increase in perceived financial reporting quality relative to company-prepared statements), whereas Compile is consistently insignificant. Thus, there appears to be relatively little effect on cost of debt from obtaining a financial statement compilation (i.e., no assurance). Further, the effect on cost of debt from financial statement audit appears to be substantially larger than that from financial statement review on the order of 50% larger, depending on sample and estimation method. For example, the coefficient estimates on Review and Audit in Column (2) are and , respectively, suggesting that, relative to no CPA firm involvement, a review (audit) is associated with a 66 (92) basis point reduction in interest rates. In summary, the cost of debt analysis is consistent with the analyses in Tables 5 and 6 in one regard financial statement review has a much larger effect than does financial statement compilation. However, there is a significant difference in the comparison of audit versus review. Relative to a financial statement review, whereas audit has little effect on output-based measures of financial reporting quality, audit has a material effect on cost of debt. However, conclusions 28

31 are difficult to draw. It may be the case that cost of debt reflects quality dimensions of audits that the other two measures do not capture, or it may be the case that lenders have an inflated perception of the benefits of financial statement audit particularly relative to review Changes in assurance level and financial reporting quality Although our propensity-score-weighted analyses help mitigate concerns that our results are attributable to self-selection or other endogeneity concerns, to provide further comfort we next present results from an analysis of firm-level changes in assurance level from one year to the next. Before presenting results from estimation of Eq. (5), in Panel A of Table 8 we provide some descriptive data on the frequency with which our sample firms move up or down zero, one, two, or three levels of assurance from year t-1 to t. For instance, a firm that changes assurance level from an audit to a review from one year to the next is labeled 1, a firm that changes from a compilation to an audit is labeled +2, and a firm that changes from an audit to company-prepared (i.e., no assurance service) is labeled 3. Well over 95% of full-sample firm-years exhibit no change in assurance level relative to the previous year (i.e., 99,632/103,114 = 96.6%). Among firm-years with a change in assurance level, nearly the same amount of firm years move up two levels and down two levels (477 and 442, respectively). While there are very few firm-years that move three levels of assurance, it is more common for firms to move from audit to company-prepared (101 firm years) than vice versa (20 firm years). Panel B of Table 8 presents results from estimating Eq. (5) using both output-based measures of financial reporting quality and cost of debt. To summarize, there is broad support for the inferences drawn in our previous analyses. First, there is no significant change in either output-based quality measure (FSDScore or AbAccruals ) from changes in assurance levels between audit and review (either in moving up from review to audit, or in moving down from 29

32 audit to review). In contrast, there is a significant increase in cost of debt when assurance level decreases from audit to review (e.g., full sample coefficient estimate of ), and a significant decrease in cost of debt when assurance level increases from review to audit (e.g., full sample coefficient estimate of ). Second, there are generally significant changes in the outputbased quality measures (in the expected directions) when assurance levels change across other combinations (i.e., changes other than from audit to review or review to audit), which are not always accompanied by changes in cost of debt. Overall, this evidence helps triangulate our main results, and provides additional evidence of the link between the level of assurance and financial reporting quality. In particular, this evidence is consistent with reviews and audits providing comparable levels of financial reporting quality, which leaves as an open question the rationale for lenders providing a substantial reduction in cost of debt for audits relative to reviews. 6. Conclusion The relative costs and benefits financial statement assurance levels is the subject of increasing attention. However, absent from this discussion is empirical evidence on the relative financial reporting quality benefits across the four alternative assurance levels available to private firms. In this study, we address this gap in the literature by providing evidence on the association between varying assurance levels and financial reporting quality using private firm data from Sageworks. Further, we use confidential attestation fee data to examine approximate relative costs for the various attestation levels, which provides needed perspective for interpreting our reporting quality results. We find consistent evidence that financial reporting quality is higher for higher assurance levels. However, the biggest quality increase appears to come from the movement from 30

33 compilation to review, with relatively modest additional benefits in moving from review to audit. Because audits are on average about 2.5 times as expensive as a review, our evidence suggests that a financial statement review is a particularly cost-effective service that yields quality that is comparable to audit. However, we document that audited financial statements are associated with significantly lower cost of debt relative to reviewed financial statements. Taken together, these results suggest that either the cost of debt reflects quality dimensions of audits that the other two output-based measures do not capture, or lenders have an inflated perception of the qualityrelated benefits of financial statement audit. Our study is subject to several important caveats. First, our private firm data permits computation only of relatively noisy empirical measures of financial reporting quality, and our cost data represent approximations. Second, our analysis does not represent a complete analysis of all costs or benefits involved in a firm s choice of attestation level (for example, Kausar et al provide evidence that private firm audit choice conveys information to capital providers, which reduces financing frictions). Finally, although we attempt to mitigate the potential selfselection problem via propensity-score-weighted regression, this approach does not fully address selection bias. The maintained assumption is that unobservable (or omitted) variables not included the first-stage model of assurance level choice do not systematically affect financial reporting quality. Despite substantial evidence on the determinants of the choice of audit firm for public companies, the literature offers little guidance on the determinants of the choice of assurance level for private firms. This lack of evidence renders it difficult to directly assess the potential confounding effect of unmodeled first-stage variables. Thus, caution should be exercised when drawing causal inferences from our findings. 31

34 Appendix A - Variable Definitions AbAccruals i,t AbAccruals i,t Accruals i,t Assets i,t Assur i,t AssurDnAtoR i,t AssurDnOther i,t AssurFees i,t AssurUpOther i,t AssurUpRtoA i,t Audit i,t CCorp i,t Company i,t Compile i,t CostOfDebt i,t+1 Firm i s abnormal accruals in year t, computed as the residual from estimation of a variant of the modified cross-sectional Jones (1991) model, as outlined in Eq. (3). We estimate the model by industry-year using two digit NAICS industry codes and require that at least 10 observations be available for each industry-year estimation. (data from Sageworks) Firm i s absolute abnormal accruals in year t, computed as the unsigned residual from AbAccruals. (data from Sageworks) Firm i s year t total accruals, measured as the change in non-cash current assets minus the change in current non-interest-bearing liabilities, minus depreciation and amortization expense. (data from Sageworks) Firm i s year t total assets (millions). (data from Sageworks) A multinomial variable that equals 1, 2, 3, or 4 if firm i s year t financial statements were company-prepared, compiled, reviewed, or audited, respectively. (data from Sageworks) An indicator that equals one if firm i decreases its level of assurance from audit in year t 1 to review in year t, and equals zero otherwise. (data from Sageworks) An indicator that equals one if firm i decreases its level of assurance in any combination other than from audit to review from year t 1 to t, and equals zero otherwise. (data from Sageworks) Dollar cost of firm i s year t assurance services. (data obtained from a Midwest regional accounting firm for 125 client observations) An indicator that equals one if firm i increases its level of assurance in any combination other than from review to audit from year t 1 to t, and equals zero otherwise. (data from Sageworks) An indicator that equals one if firm i increases its level of assurance from review in year t 1 to audit in year t, and equals zero otherwise. (data from Sageworks) An indicator that equals one if firm i receives a financial statement audit in year t, and equals zero otherwise. (data from Sageworks) An indicator that equals one if the firm is incorporated under Subchapter C of the Internal Revenue Code, and equals zero otherwise. (data from Sageworks) An indicator that equals one if firm i prepares its own financial statements without engaging the services of a CPA firm in year t, and equals zero otherwise. (data from Sageworks) An indicator that equals one if firm i receives a financial statement compilation in year t, and equals zero otherwise. (data from Sageworks) Firm i s year t+1 interest expense divided by average debt outstanding in year t+1. (data from Sageworks) 32

35 CurrRatio i,t Debt i,t FSDScore i,t IntCoverage i,t Leverage i,t NegEquity i,t PPE i,t PPEta i,t Firm i s year t ending current assets divided by ending current liabilities. (data from Sageworks) Firm i s year t ending total liabilities. (data from Sageworks) Firm i s year t financial statement deviation score, which is a measure based on Benford s law that is decreasing in financial reporting quality. FSDScore is computed as in Eq. (2), and measures the extent to which the numbers reported in a given set of financial statements deviate from the theoretical distribution of numbers implied by Benford s law. (data from Sageworks) Firm i s year t earnings before interest, taxes, depreciation, and amortization expenses divided by interest expense. (data from Sageworks) Firm i s year t ending total liabilities divided by ending total assets. (data from Sageworks) An indicator that equals one if firm i s year t ending total liabilities are greater than ending total assets, and equals zero otherwise. (data from Sageworks) Firm i s year t ending net value of property, plant, and equipment. (data from Sageworks) Firm i s year t ending net value of property, plant, and equipment divided by beginning total assets. (data from Sageworks) Receivables i,t Firm i s year t ending trade receivables. (data from Sageworks) Review i,t An indicator that equals one if firm i receives a financial statement review in year t, and equals zero otherwise. (data from Sageworks) ROA i,t SalesGrowth i,t Firm i s year t return on assets. (data from Sageworks) The percentage change in sales from year t 1 to year t. (data from Sageworks) SalesGrowthSq i,t SalesGrowth squared. (data from Sageworks) 33

36 Appendix B Descriptions of Different CPA Firm Assurance Services (source: AICPA) 34