Do Fees Paid to Auditors Increase a Company’s Likelihood of Meeting Analysts’ Earnings Forecasts?
by
Jere R. Francis School of Accountancy University of Missouri-Columbia Columbia, MO 65211 Tel: 573-882-5156 Fax: 573-882-2437 Email: francis@missouri.edu
and
Bin Ke Department of Accounting The Smeal College of Business Administration The Pennsylvania State University University Park, PA 16802 Tel: 814-865-0572 Fax: 814-863-8393 E- mail: bxk127@psu.edu
First draft: November 2001 Current draft: May 21, 2003 Version 3.9
Comments are welcome. Please do not quote without permission.
We the editor and two anonymous referees for their many useful comments. We also thank Mark Dirsmith, George Drymiotes, and participants at the 2002 American Accounting Association annual meeting and George Washington University for their comments, and Jonah Gruber and Santhosh Gowda for able research assistance.
Do Fees Paid to Auditors Increase a Company’s Likelihood of Meeting Analysts’ Earnings Forecasts?
Abstract The definition of the dependent variable (and sample) is misspecified in the analysts’ forecast benchmark test in Frankel et al. (2002) and therefore cannot be used to infer the effect of fees on auditor independence and the earnings management behavior of clients that pay their auditors large fees. We can replicate the significant result in Frankel et al. (2002) using their definition of the dependent variable and a sample of quarterly earnings surprises. However, when we use a more defensible definition of the dependent variable based on the distribution anomalies documented in Degeorge et al. (1999), there is no association between fees paid to auditors and a firm’s likelihood of meeting or just beating analysts’ earnings forecasts. These findings are robust to alternative measures of fee dependence and we conclude there is no systematic evidence that fees paid to auditors for either nonaudit services or for total services compromise auditor independence, at least with respect to allowing clients discretion to meet analysts’ earnings forecasts.
Key Words: Auditor independence; Nonaudit services; Auditor fees; Earnings management
Do Fees Paid to Auditors Increase a Company’s Likelihood of Meeting Analysts’ Earnings Forecasts? 1. INTRODUCTION The objective of this study is to further investigate if fees paid to auditors affect their clients’ likelihood of meeting analysts’ earnings forecasts. In a seminal study, Frankel, Johnson and Nelson (2002), hereafter FJN, report evidence that firms which pay their auditors higher levels of nonaudit fees are more likely to meet (or just beat) analysts’ forecasts. Using the same basic research design but a slightly smaller sample, Ashbaugh, LaFond and Mayhew (2003), hereafter ALM, cannot replicate the result in FJN. The contradictory evidence in FJN and ALM is unsettling because of the significant regulatory issues at stake for the accounting profession. The results in FJN indicate that auditor independence may be compromised by nonaudit services, and that auditors allow clients who purchase high levels of nonaudit services greater discretion to manage earnings and meet benchmark earnings targets. 1 Their findings lend support to the Securities and Exchange Commission which proposed an outright ban on nonaud it services (SEC 2000a, 2000b, 2000c), and the mandated restrictions on some nonaudit services in the SarbanesOxley Act of 2002 (Public Law 107-204, July 30, 2002). Alternatively, if the results in ALM are correct then prohibitions on nonaudit services may be unnecessary and could represent an unwarranted regulatory intrusion into the private sector (AICPA, 1997; Copeland, 2000). Our paper makes three significant contributions to this line of research. First, we show that the dependent variable is miscoded in FJN and ALM by the inclusion of observations with large negative earnings surprises and this results in inappropriate samples for testing. For this reason neither study can provide unambiguous evidence on the association between fees paid to
1
The analyst benchmark test is a potentially powerful test of earnings management because prior research documents a strong managerial incentive to manage earnings to meet or just beat analysts’ earnings forecasts (Degeorge et al. 1999; Brown 2001; Payne and Robb 2000; Matsumoto 2002; Brown and Caylor 2003).
auditors and earnings management by firms to meet or beat analysts’ earnings forecasts. Second, we use a sample of quarterly earnings for the year 2000 to test if there is an association between fees paid to auditors and the likelihood of firms meeting or just beating quarterly earnings forecasts. FJN and ALM report contradictory evidence using annual earnings surprises for year 2000. However, the advantage of using quarterly earnings surprises is that it provides a larger and potentially more powerful sample because quarterly earnings are only subject to limited review rather than full audit and are therefore more subject to earnings management. Third, we test six measures of the economic dependence created by fees paid to auditors, in addition to FEERATIO (the ratio of nonaudit to total fees) which is the primary test variable in FJN. Our results are as follows. We are able to replicate FJN in our sample of quarterly earnings when we define the dependent variable in the same manner as they do, and find a positive association between FEERATIO and the likelihood of meeting or just beating analysts’ forecasts. However, when we use a coding rule (and sample) that is more defensible and consistent with the evidence in Degeorge et al. (1999), the results on FEERATIO become insignificant at conventional levels. It turns out that the replication of FJN is driven by the inappropriate inclusion of firms with large negative earnings surprises, and when we delete these observations FEERATIO becomes insignificant. These results are robust to alternative measures of fees paid to auditors including the dollar magnitudes of nonaudit fees and total fees. Except for the replication of FJN based on the miscoding of the dependent variable, we find no systematic evidence that companies which pay their auditors higher levels of fees are more likely to manage earnings to meet analysts’ forecasts, and therefore conclude that fees paid to auditors do not necessarily have an adverse effect on auditor independence. 2 The remainder of
2
FJN also report that abnormal accruals are larger for firms paying their auditors high levels of nonaudit services which implies that accruals are used to achieve earnings management objectives. Abnormal accruals are not
2
the paper is organized as follows. The next section describes our research design choices, the sample and descriptive statistics. Section III presents a replication of the result in FJN using our sample of quarterly earnings surprises and their coding rule for the dependent variable. Section IV analyses alternative definitions of the dependent variable and alternative proxies for auditor independence. Section V concludes the paper.
II. RESEARCH DESIGN AND SAMPLE Research Design Choices The dependent variable and sample are incorrectly defined by FJN and ALM with respect to the evidence in Table 6 of Degeorge et al. (1999) that is used to motivate the coding rule in both studies. FJN and ALM code the dependent variable “one” for firms that meet or beat analysts’ forecasts by +1 cent and this is consistent with Degeorge et al. (1999). However, they miscode the zero condition and use inappropriate samples for their tests because they include all other firms in the zero coding condition. This is inappropriate because there is no evidence in Degeorge et al. (1999) to support the inclusion of observations that have large negative earnings surprises in the zero coding condition, and the inclusion of such observations is unnecessary (at best) and may introduce unkno wn biases (at worst). Degeorge et al. (1999) document distributional anomalies for analysts’ forecast errors (actual earnings minus forecasted earnings). They show that zero and small positive earnings surprises (firms that beat the forecast by +1 cent per share) are over-represented in the
investigated in our analysis because the association between nonaudit fees and abnormal accruals reported in Frankel et al. (2002) is the primary focus of Ashbaugh et al. (2003) who are able to replicate the Frankel et al. result using the same research design, but they document that the statistical association is driven by firms with incomedecreasing accruals and further show that the association does not hold after controlling for the endogenous effect of firm performance on abnormal accruals (Kothari et al. 2002). Chung and Kallapur (2003) use a different research design, but they also find no association between abnormal accruals and the level of fees paid to auditors, after controlling for industry effects on accruals. They test both nonaudit and total fees measured as percentages of the audit firm’s total fees from its national clientele, and alternatively as percentages of total fees of the engagement office administering the audit, along the lines suggested by the office-level analysis in Reynolds and Francis (2000).
3
distribution, while small negative earnings surprises of -1 cent per share and large positive earnings surprises in excess of +1 cent per share are under-represented. Based on this they conclude there is evidence of earnings management in which some firms meet or just beat analyst forecasts by either (1) increasing earnings to avoid a small negative surprise, or (2) decreasing earnings when firms would otherwise have a large positive earnings surprise. In other words, earnings management occurs by shifting earnings upward from a near miss (small negative surprise) or “reigning in” large positive earnings surprises through so-called cookie jar reserves. Although Degeorge et al. (1999) conjecture there is possible earnings management for firms with large negative earnings surprises (i.e., big bath), there is no evidence in their study to suggest that the excessive number of small positive earnings surprises has been shifted from the large negative earnings surprise interval. Therefore, based on Degeorge et al. (1999), the appropriate definition of the dependent variable for the benchmark test is to code as “one” those firms in the over-represented region as more likely to have managed earnings (those that exactly meet or just beat the forecast by +1 cent); and to code as “zero” those firms in the underrepresented regions that are less likely to have managed earnings (those with small negative earnings surprises of -1 cent per share or those firms with large positive earnings surprises of greater than +1 cent per share). 3 Our second design choice is the use of quarterly earnings surprises rather than annual earnings surprises as used by FJN and ALM. There are several advantages in testing quarterly earnings. Quarterly data provide a potentially more powerful test due to a larger number of observations than if annual data are used, and this is an important design consideration given the contradictory evidence in FJN and ALM using annual data. Prior research also documents the
3
As a sensitivity analysis we also test cutoffs at -2 cents negative surprises and +2 cents positive earnings surprises and the results are qualitatively the same.
4
active management of quarterly earnings to meet analysts’ forecasts, which means that earnings are actively managed throughout the year and not just at year end (Ettredge et al. 2000; Mendenhall et al. 1988). In addition, interim quarterly earnings are only subject to a timely review rather than a full audit, which means that firms have greater discretion (at least in the short term) to actively manage quarterly earnings (Manry et al. 2003). What constrains this behavior is that the sum of quarterly earnings filed in 10-Qs must equal annual audited earnings. This means that fourth quarter earnings will implicitly include adjustments for misestimations in the interim quarters, and a firm with an independent and uncompromising auditor is unlikely to risk such a large auditor-imposed adjustment. On the other hand, if fees do compromise auditor independence, then firms that pay their auditors large fees may more aggressively manage quarterly earnings because the risk of such a fourth quarter adjustment is small. For all of these reasons, we believe quarterly earnings provide a better setting than annual earnings to test the association of auditor fee dependence and potential earnings management behavior by clients. A number of variables are used to measure fees paid to auditors. FEERATIO (the ratio of nonaudit fees to total fees) is the initial test variable because it is the primary metric used for the analyst benchmark test in FJN. However, six alternative measures of the potential effect of fees on auditor independence are examined as a robustness check. FEERATIO has been
criticized for failing to measure the magnitude of the economic bond created by fees paid to auditors, and the variable literally measures nothing more than the proportion of total fees derived from nonaudit services (Kinney and Libby 2003). 4 With this criticism in mind, we create
4
While FEERATIO has been used in prior studies, it is difficult to interpret this variable beyond the simple information it conveys which is the relative proportion of fees paid for auditing versus other services. To illustrate why the variable may be misleading, Ashbaugh et al. (2003) report that two firms in their sample that have an identical fee ratio of 73% (nonaudit fees are 73% of total fees paid to the auditor); yet in one case the total fees are only $71,000 and in the other case total fees are $5.7 million. It is difficult to imagine any scenario in which total fees of $71,000 would create an economic bond that threatens auditor independence even if virtually all of the fees were for nonaudit services.
5
FEERATIO1 and FEERATIO2 as two alternative indicator variables that identify smaller subsets of observations with more extreme values of nonaudit fees in terms of both absolute size and their size relative to audit fees. FEERATIO1 is coded one for firms whose FEERATIO is greater than 0.50, and whose nonaudit fees are greater in magnitude than the 75th percentile value for Big 5 or non-Big 5 auditors, as appropriate, with all other observations coded zero. FEERATIO2 is the same as FEERATIO1 except that the 90th percentile value is used as the cutoff for the magnitude of nonaudit fees. In addition, Kinney and Libby (2003), Chung and Kallapur (2003) and Ashbaugh et al. (2003) suggest that the dollar magnitude of fees (either nonaudit fees or total fees) is a potentially better measure of the economic bond that fees create and which could lead to the impairment of auditor independence. So the next two proxies for auditor independence represent direct measures of the dollar magnitude of fees: NONAUDITFEE which is the dollar amount of nonaudit fees paid to the auditor, and TOTALFEE which is the dollar amount of total fees (nonaudit and audit) paid to the auditor. The final two measures are based on rankings represented by the dollar amounts of the nonaudit fees and total fees for an observation relative to all other companies in the sample audited by the same auditor. The idea here is to calculate a measure of economic fee dependence for a
particular client based on the size of client fees relative to the auditor’s fees from all clients. The variable NONAUDITFEERANK is defined as the percentile ranking (1 to 100) of a client’s nonaudit fees relative to the nonaudit fees of all clients of the auditor, and TOTALFEERANK is defined as the percentile ranking (1 to 100) of a client’s total fees (nonaudit and audit) relative to the total fees of all clients of the auditor. Since non- Big 5 auditors audit a small number of clients, both NONAUDITFEERANK and TOTALFEERANK are measured and tested only for the Big 5 auditors.
6
Following FJN and ALM we use year 2000 data for the benchmark tests because this is the fiscal year related to the initial fee disclosures. However, the SEC’s attempt in early 2000 to proscribe all nonaudit services and the intense publicity surrounding the nonaudit services controversy could create a bias with respect to year 2000 data. Auditors may have been
especially careful with clients in year 2000 to avoid claims they were favoring clients who paid large fees. Therefore, as a sensitivity analysis, we test the two prior years on the grounds that they may better represent an unbiased test period. The test results (not tabulated) are consistent using both 2000 data and pooled 1998-99 data, indicating no evidence of bias in year 2000 data. 5 Sample and Descriptive Statistics The sample is drawn from the 2000 Quarterly Compustat files. Regulated firms in utilities (SIC codes 4000-4999) and financial institutions (SIC codes 6000-6999) are excluded because their economic environments are different from firms in other industries. In addition, observations are deleted with missing quarterly earnings per share before extraordinary items, and missing values on the following variables used in the models: book value of common equity, market value of common equity, cash flows from operations, total assets, and auditor identification (Big 5 or non-Big 5). We also require non-missing analysts’ forecast data from the IBES database, and disclosures of fees for audits and for nonaudit services in proxy statements filed electronically with the SEC through October 31, 2001. Actual earnings and analysts’ earnings forecasts come directly from the IBES database. These screens result in a sample of 1,588 unique firms having one or more quarters of useable data, and 5,208 firm-quarter
5
Fscal 2000 fees are likely to be a good proxy for prior years’ fees because such fees are highly correlated from year to year. Nonaudit fees are publicly disclosed in Australia, and the evidence there indicates that nonaudit fees have very high year-to-year correlation. For a sample of Australian public companies, the correlation of year 2000 nonaudit fees with 1999 nonaudit fees is 0.86 (n=411); the correlation of year 2000 nonaudit fees with nonaudit fees from two years earlier (1998) is 0.88 (n=331). Yearly correlations of audit fees with prior year fees are even higher. We thank Professor Allen Craswell, University of Queensland, for providing data for these calculations.
7
observations. In comparison, FJN (Table 5) have a sample of 2,012 unique firms for their analyst benchmark test, while ALM (Table 7) have a sample of 1,666 unique firms. Table 1 provides descriptive statistics of year 2000 audit fees and nonaudit fees for the sample of 1,588 unique firms in the study. The SEC disclosure rule requires nonaudit fees to be disaggregated into financial information systems design and implementation fees, and all other nonaudit fees. Nonaudit fees are greater than audit fees for just over 50% of the unique firms in the sample. Of the total nonaudit fees, most are classified as “other fees.” Financial information systems design and implementation fees are minimal for most firms, although they tend to be large when they do exist. The sample mean (median) ratio of nonaudit fees to total fees is 0.519 (0.539), which is similar to the mean (median) of 0.49 (0.51) in Frankel et al. (2002, Table 2). The interquartile range is 0.357 to 0.686 and is comparable to the interquartile range of 0.30 to 0.69 in FJN (Table 2). Thus the profile of our sample fee data corresponds to the sample in FJN. [Insert Table 1 Here] Descriptive statistics are reported in Table 2 for the full sample of all firm-quarter observations. The dependent variable is denoted SURPRISE and is initially defined in the same manner as FJN and ALM. Earnings surprises are coded “one” for firms with earnings that exactly meet or beat analysts’ earnings forecasts by one cent per share, and “zero” for all other observations. This coding rule results in 36% of total firm-quarter observations (1,895 quarterly observations) with earnings surprises that are zero or +1 cent per share, and 64% (3,313 quarterly observations) for all other observations. 6 These other observations are classified as follows: small negative earnings surprises of -1 cent per share are 268 observations (5% of sample); and large positive earnings surprises greater than +1 cent share are 2,157 observations (41% of
6
There are 1,096 observations with zero earnings surprises, and 799 observations with +1 cent surprises. The tests are robust to coding the dependent variable one for each of these groups separately.
8
sample). The smaller number of observations in the small negative earnings surprise interval relative to the zero/small positive earnings surprise interval is consistent with Degeorge et al. (1999). Finally, the interval of large negative earnings surprises (in excess of -1 cent per share) contains 888 quarterly observation (17% of sample), and we show later in the study how the inappropriate inclusion of these observations and their coding as zero (for the dependent variable) drives our replication of FJN using quarterly data. [Insert Table 2 Here] The variables which proxy for auditor fee dependence are reported next in Table 2. The variable FEERATIO has a median value of 0.558 which indicates nonaudit fees are greater than audit fees for over half of the firm-quarter observations. FEERATIO1 and FEERATIO2 are indicator variables and represent 33.3% and 15.6% of total firm-quarter observations, respectively. NONAUDTFEE and TOTALFEE are the dollar amounts of nonaudit fees and total audit fees paid to auditors, and these values are logged for the benchmark tests.
7
NONAUDITFEERANK and TOTALFEERANK are the percentile values for an observation’s nonaudit fees and total fees relative to the auditor’s total clientele in the sample. The remaining variables in Table 2 are control variables used in FJN. 8 FJN use BM, LOSS, and LITIRISK to proxy for the firm’s earnings management incentives. FJN argue that firms are more likely to have managed earnings to meet benchmark targets if they have low BM ratios (growth firms), have avoided a loss (LOSS), or are in industries with high litigation (LITIRISK). Additional control variables are firm size (MV), operating cash flows (CF),
7 8
Since nonaudit fees are zero for some firms, we added 1 to nonaudit fees before taking the logarithm.
Our study excludes three control variables in FJN: the number of years that the incumbent auditor has audited the firm’s financial statements, the contemporaneous abnormal stock return, and percentage of institutional ownership. We exclude the first two variables because they are insignificant in FJN. For institutional ownership, FJN report a significantly positive coefficient, but we are unable to include this variable as we do not have access to institutional ownership data. However, omission of this variable does not affect our ability to successfully replicate their statistical results for the test variable FEERATIO with our sample of quarterly earnings surprises.
9
profitability or return on assets (ROA), and type of auditor (BIG5). FJN include MV as a control for the general effect of firm size, CF to control for the possibility that firms with high cash flow may be more likely to beat analyst forecasts, and ROA because it is a determinant of non-audit services in other studies. Finally, the variable BIG5 controls for systematic differences between Big 5 and non-Big 5 auditors.
III. REPLICATION OF FRANKEL ET AL. (2002) We begin with a replication of Frankel et al. (2002) using their definition of the dependent variable and our sample of quarterly data, and test for evidence of an association between FEERATIO and the likelihood of meeting the benchmark earnings target. The following probit model is adapted from Frankel et al. (2002): SURPRISEit = a + ß1 FEERATIO i + ß2 BMiT + ß3 MV it + ß4 CFit + ß5 ROAit + ß5 LOSSit + ß6 LITIRISKit + ß7 BIG5i + eit ; where i t SURPRISE = firm index; = quarter index; = 1 if quarterly earnings per share minus the latest analyst consensus (median) forecast before the actual earnings announcement is 0 or +1 cent, and 0 for all other firms , i.e., positive surprises greater than +1 cent and all negative earnings surprises; FEERATIO BM = ratio of nonaudit fees to total fees paid to the incumbent auditor in year 2000; = ratio of book value to market value of common equity at the beginning of quarter t; MV CF = natural log of market value of common equity at the beginning of quarter t; = quarterly operating cash flow scaled by lag-one total assets. (1)
10
ROA LOSS LITIRISK
= quarterly earnings per share scaled by lag-one total assets per share. = 1 if quarterly earnings per share is negative, and zero otherwise. = 1 if a firm is in high litigation industries (i.e., SIC 2833-2836, 3570-3577, 73707374, 3600-3674, 5200-5961), and zero otherwise.
BIG5
= 1 if a firm is audited by one of the Big- five accounting firms, and zero otherwise. The dependent variable SURPRISE is initially defined as in FJN and ALM, but later
analyses redefine SURPRISE to determine the sensitivity of the results to the observations included in the zero coding condition. The initial variable to test the potential effect of the auditor’s fee dependence is FEERATIO. As in FJN and ALM, p-values on FEERATIO are reported as two-tail p- values since the directional relation between fees and earnings surprises is unknown even though a positive association is the primary direction of interest. Since
FEERATIO does not directly measure the magnitude of fees paid to auditors, a sensitivity analysis is reported later in the paper to test other measures of auditor fee dependence. [Insert Table3 Here] The estimation of equation (1) is reported in column (1) of Table 3 using all firm-quarter observations for year 2000. 9 The results in column (1) are consistent with Frankel et al. (2002, Table 5) and indicate a strong positive association between FEERATIO and meeting the benchmark earnings target (p=0.023). Thus we are able to replicate FJN using our sample of quarterly earnings surprise and their coding rule for the dependent variable. 10
9
Since each firm has multiple observations in our sample, we compute all regression variables’ t-statistics based on the method of Rogers (1993), which allows heteroskedasticity and any type of correlation for observations of the same firm but assumes independence for observations of different firms.
10
The results in column (1) of Table 3 are robust to estimation using firm-quarter observations from 1998 and 1999, before non-audit services became such a high-profile issue. For all 1998-99 firm-quarter observations (n=11,861) FEERATIO is positive and significant at p=0.018.
11
IV. ALTERNATIVE MODEL SPECIFICATIONS AND FEE VARIABLES We now demonstrate how the significant result on FEERATIO in column (1) of Table 3 is driven by the miscoding of the dependent variable in the benchmark test and the inappropriate inclusion of observations having large negative earnings surprises. In addition, we show that the results on FEERATIO are not robust to other measurements that more directly capture the economic magnitude of fees paid to auditors. Definition of the Dependent Variable and Sample FJN classify all firms other than those that meet or just beat the benchmark by +1 cent per share as the zero coding condition. As explained earlier in the paper, this coding is not supported by the distributional analysis in Degeorge et al. (1999). Specifically, there is no evidence that the zero coding condition should include firms with large negative earnings surprises. We report four sets of results in columns (2) to (5) of Table 3 to illustrate the sensitivity of the result in column (1) of Table 3 to the coding rule used in FJN and ALM. The models in columns (2) to (5) of Table 3 differ only in terms of the observations that are included in the sample with respect to the zero coding condition for the dependent variable. Column (2) uses only large negative earnings surprises (in excess of -1 cent per share) in the zero coding condition; column (3) uses only large positive earnings surprises (in excess of +1 cent per share) in the zero coding condition; and column (4) uses only firms with small negative earnings surprises (-1 cent per share) in the zero coding condition. Finally, column (5) is a sample (and coding rule) that is most consistent with the analysis in Degeorge et al. (1999), using only those firms with small negative earnings surprise ( 1 cent per share) and large positive earnings surprises (in excess of +1 cent per share) in the zero coding condition.
12
The analysis in columns (2) to (4) of Table 3 shows very clearly that the statistical significance of FEERATIO is driven by the inappropriate inclusion of firms with large negative earnings surprises. The test variable FEERATIO is significant at conventional levels (p<0.05) only in column (2) in which the zero coding condition uses only observations with large negative earnings surprises. FEERATIO is not significant at even the 0.10 level in columns (3) or (4) which test large positive earnings surprises and small negative earnings surprises, respectively. Column (5) uses a sample (and coding rule) that is most consistent with Degeorge et al. (1999), and FEERATIO is not significant at the conventional 0.05 level. We conclude from Table 3 that the statistical significance on FEERATIO in column (1) of Table 3 is sensitive to the definition and coding rule for the dependent variable used in Frankel et al. (2002) and that the significance for FEERATIO in our sample is driven by the inappropriate inclusion of observations coded zero that have large negative earnings surprises. When these observations are deleted FEERATIO is not significant at conventional levels. 11 Additional diagnostics are now reported to better understand exactly how firms with large negative earnings surprises drive the statistical significance of FEERATIO. The mean va lue of FEERATIO for observations with large negative earnings surprises is 0.48, and these observations have an average negative earnings surprise of -14 cents per share. The other two groups in the zero coding condition (small negative surprises and large positive surprises) have a mean FEERATIO of 0.53. In comparison, observations coded one (those that meet or the beat the forecast by +1 cent) have a mean FEERATIO of 0.55. These descriptive statistics show very clearly that FEERATIO is comparable across observations in the sample except for the subset of
11
The results in columns (2) to (5) of Table 3 are robust to estimation using a pooled sample of 1998-99 firmquarter observations. In column (2) which includes only large negative earnings surprises in the zero coding condition, FEERATIO is significant at p<0.01; while for the model in column (5) which is most consistent with Degeorge et al. (1999), FEERATIO is not significant at the 0.10 level.
13
observations with large negative earnings surprises. What this indicates is that the statistical significance of FEERATIO in column (1) of Table 3 is not driven by higher values of FEERATIO for those firms that meet or just beat the benchmark, which is what the fee dependence story implies. Rather, the statistical result is driven by significantly lower values of FEERATIO for just one subset of observations, those firms with large negative earnings surprises that are inappropriately included in the analysis. The statistical significance for FEERATIO in column (1) of Table 3 is caused by a subset of firms that average -14 cent negative earnings surprises, which does not support the argument that an auditor’s fee dependence increases the likelihood that its clients will management earnings to meet or beat analyst forecasts by +1 cent per share. In sum, column (1) of Table 3 uses a sample of quarterly observations to replicate the statistical result in FJN for the variable FEERATIO. However, based on the results in columns (2) to (5) of Table 3 we conclude that this result is a statistical artifact of the misspecification of the dependent variable and, by implication, the sample that is used in the benchmark test. Thus the result in column (1) of Table 3 is not interpretable as evidence of earnings management by firms that pay their auditors more for nonaudit services as measured by the variable FEERATIO. Alternative Measures of Auditor Fee Independence The variable FEERATIO has been criticized as a poor measure of the fee dependence created by nonaudit fees and the potential threat to auditor independence (Kinney and Libby 2003). A compelling argument can be made that the economic bond arises from the magnitude of fees rather than the relative mix of nonaudit fees and audit fees. Table 4 reports results using six alternative measures of auditor fee dependence that explicitly incorporate the magnitude of fees: FEERATIO1, FEERATIO2, NONAUDITFEE, NONAUDITFEERANK, TOTALFEE, and
14
TOTALFEERANK. For brevity we only report regression coefficients and p-values for the six measures of auditor fee dependence. All other variables are the same as those used in Tables 3 and their coefficients are not reported in the table. Column (1) of Table 4 uses the full sample based on the dependent variable coding rule of FJN and ALM, while column (2) excludes observations with large negative earnings surprises, which is more consistent with Degeorge et al. (1999). [Insert Table 4 Here] Column (1) replicates the dependent variable coding rule in FJN and ALM, and none of the six alternative fee variables are significant at the 0.10 level. Therefore we conclude that that the significant results on FEERATIO in column (1) of Table 3 which replicate FJN using quarterly data are not robust to alternative and more direct measures of auditor fee dependence and the potential economic bond created by large levels of fees. Column (2) excludes large negative earnings surprises from the sample, and only one of the six test variables is significant at p<0.10, the natural log of NONAUDITFEE which is significant at p=0.046. When the magnitude of nonaudit fees is expressed relative to the size of the auditor’s clientele using the variable NONAUDITFEERANK there is no significant association. 12 Overall there is no
consistent evidence in Table 4 that fees paid to auditors for either audit or nonaudit services affect the likelihood of meeting the benchmark earnings target. 13
12
Note the difference in the significance of the coefficient for NONAUDITFEE and NONAUDITFEERANK is not due to the exclusion of non-Big Five auditees in the regression of NONAUDITFEERANK. The coefficient on NONAUDITFEE remains marginally significant (p=0.07, two-tailed) if non-Big Five audited firms are excluded.
13
The results in Table 4 are robust to reestimation using a pooled sample of 1998-99 firm-quarter observations. None of the six fee variables in column (1) or column (2) has a significant positive association with the dependent variable (meeting or beating analysts forecast by +1 cent) at even the 0.20 significance level.
15
More Restrictive Test Based on Degeorge et al. (1999, Figure 8) We report one final benchmark test using a more restrictive sample based on an analysis in Degeorge et al. (1999, Figure 8, p. 24). 14 Degeorge et al. (1999) conclude that there is evidence of a hierarchy in earnings management behavior. The first objective is to report positive earnings, then to report a positive change in earnings over the prior period, and finally to meet or beat analysts’ earnings forecasts given that the first two objectives are met. Based on this hierarchy we restrict our test sample to those observations which have positive quarterly earnings and which also report increases in quarterly earnings (relative to the same quarter in the prior fiscal year), and then test if firms within this restrictive sample are more likely to meet or beat analysts’ earnings forecasts conditional on fees paid to auditors. Arguably, this subset of the sample is more likely to include firms that have actually managed earnings, ceteris paribus, and therefore should provide a stronger test setting. All proxies for auditor fee dependence are examined and the results are summarized in Table 5. For brevity we only report the regression coefficients and p-values for FEERATIO and the six alternative measures of auditor fee dependence. [Insert Table 5 Here] Two sets of estimations are reported in Table 5. The estimations in column (1) use the dependent variable coding rule from FJN and ALM, while the estimations in column (2) exclude observations with large negative earnings surprises and are more consistent with Degeorge et al. (1999). None of the fee variables in Table 5 are significant at the 0.10 level in this more restrictive test setting, which provides even more compelling evidence that there is no
14
We thank one of the referees for suggesting this analysis.
16
association between fees paid to auditors and the likelihood of earnings management behavior by clients in order to meet (or just beat) analysts’ earnings forecasts. 15
V. CONCLUSION Using a sample of quarterly earnings surprises, we can replicate the result in Frankel et al. (2002) of a positive association between FEERATIO (the ratio of nonaudit fees to total fees) and the likelihood of meeting analysts’ earnings forecasts which implies greater discretion to manage earnings. However, this result is not robust to other proxies for auditor fee dependence that more directly measure the magnitude of fees and the economic bond created by such fees that could potentially compromise auditor independence. More importantly, the definition of the dependent variable in Frankel et al. (2002) is problematic. Specifically, the inclusion of firms with large negative earnings surprises is not consistent with the distributional anomalies documented in Degeorge et al. (1999). Using a more appropriate definition of the dependent variable that is consistent with Degeorge et al. (1999), our study finds no systematic evidence of an association between the fees paid by a company to its auditor and the likelihood of managing earnings to meet quarterly benchmark earnings targets. The analyst benchmark test in Frankel et al. (2002) suggests that nonaudit services may impair the independence of auditors, and this rationale has been used by Congress and the SEC to limit such services. However, our study demonstrates the potential fragility of these results and their sensitivity to research design choices. Our replication of Frankel et al. (2002) using quarterly data is not robust to a more defensible definition of the dependent variable, nor to
15
The results in Table 5 are robust to reestimation using a pooled sample of 1998-99 firm-quarter observations. None of the seven fee variables in column (1) or column (2) has a significantly positive association with the dependent variable (meeting or beating analysts forecast by +1 cent) at the 0.10 significance level.
17
alternative measures of fee dependence that more directly measure the economic magnitude of fees paid to auditors. Based on the analyst benchmark tests in our study, as well as the findings in Ashbaugh et al. (2003) and Chung and Kallapur (2003), there appears to be no systematic evidence that auditors compromise their independence and allow clients with higher levels of fees (either nonaudit fees or total fees) greater discretion to manage earnings In the absence of such
evidence regulatory agencies should proceed with caution before placing further restrictions on the provision of nonaudit services to audit clients, particularly since current estimates are that the Big 4 accounting firms still earn around 50% of total revenues from nonaudit services which indicates a strong market demand for such services (Bryan-Low, 2002).
18
References American Institute of Certified Public Accountants. 1997. Serving the Public Interest: A New Conceptual Framework for Auditor Independence. American Institute of Certified Public Accountants, New York City. Ashbaugh, H., R. LaFond, and B. Mayhew. 2003. Do Nonaudit services compromise auditor independence?: further evidence. The Accounting Review (forthcoming). Brown, L. D. 2001. A temporal analysis of earnings surprises: Profits versus losses. Journal of Accounting Research 39 (2): 221-242. Brown, L.D., and M. Caylor. 2003. A Temporal analysis of earnings management thresholds. Working paper (Georgia State University). Bryan-Low, C. 2002. Accounting firms are still consulting. The Wall Street Journal, September 23, 2002, pp. C1 and C7 Chung, H., and S. Kallapur. 2003. Client importance, nonaudit services, and abnormal accruals. The Accounting Review (forthcoming). Copeland, J. E., 2000. Accounting ain’t broke, so don’t fix it. The Wall Street Journal, July 25. Ettredge, M. L., D. T. Simon, D. B. Smith, and M. S. Stone. 2000. The effect of the external accountant's review on the timing of adjustments to quarterly earnings. Journal of Accounting Research 38 (1), 195-208. Degeorge, F., J. Patel, and R. Zeckhauser. 1999. Earnings management to exceed thresholds. Journal of Business 72:1-33. Frankel, R., M. Johnson, and K. Nelson. 2002. The relation between auditors’ fees for nonaudit services and earnings management. The Accounting Review 77 (Supplement): 71-105. Kinney, W., and R. Libby. 2003. Discussion of the relation between auditors’ fees for nonaudit services and earnings management. The Accounting Review 77 (Supplement): 107-114. Kothari, S., A. Leone, and C. Wasley. 2002. Performance matched discretionary accruals measures. Working paper (University of Rochester). Levitt. A. 1998. The ‘Numbers Game.’ Remarks of SEC Chairman A. Levitt at the New York University Center for Law and Business, New York, N.Y., September 28, 1998. Manry, D., S. Tiras, and C. Wheatley. 2003. The influence of interim auditor reviews on the association of returns with earnings. The Accounting Review 78: 251-274.
19
Matsumoto, D. 2002. Management’s incentive to avoid negative earnings surprises. The Accounting Review 77: 483-514. Mendenhall, R. R., W. D. Nichols. 1988. Bad News and Differential Market Reactions to Announcements of Earlier-Quarters Versus Fourth-Quarter Earnings. Journal of Accounting Research 26, 63-86. Payne, J., and S. Robb. 2000. Earnings management: the effect of ex ante earnings expectations. Journal of Accounting, Auditing, and Finance 15 (4): 371 -392. Reynolds, J.K., and J. Francis. 2000. Does size matter? the influence of large clients on office-level auditor reporting decisions. Journal of Accounting and Economics 30: 375-400. Rogers, W.H. 1993. sg17: Regression standard errors in clustered samples. Stata Technical Bulletin 13, 19-23. Securities and Exchange Commission. 2000a. Proposed rule: Revision of the Commission’s auditor independence requirements. 17 CFR Parts 210 and 240 (Release Nos. 33-7870; 34-42994; 3527193; IC-24549; IA-1884; File No. S7-13-00). SEC: Washington, D.C. Securities and Exchange Commission. 2000b. Revision of the Commission’s auditor independence requirements. (Release Nos. 33-7919; 34-43602; File No. S7-13-00). SEC: Washington, D.C. Securities and Exchange Commission. 2000c. Hearing on auditor independence. September 21, 2000. http://www.sec.gov/rules/extra/audmin4.htm.
20
Table 1. Audit and nonaudit fees for fiscal year 2000 (N=1,588)a
Variable Mean Standard Deviation Q1 Median Q3 Minimum Maximum
Audit IS Other Nonaudit Total
0.705 0.352 1.380 1.733 2.438
1.512 2.527 3.856 5.442 6.708
0.146 0 0.100 0.106 0.283
0.285 0 0.309 0.330 0.662
0.611 0 1.051 1.126 1.781
0.026 0 0 0 0.029
25.000 46.800 68.200 79.700 103.600
Audit/Total IS/Total Other/Total Nonaudit/Total
a
0.481 0.037 0.482 0.519
0.221 0.122 0.216 0.221
0.314 0 0.316 0.357
0.461 0 0.500 0.539
0.643 0 0.648 0.686
0.015 0 0 0
1.000 0.929 0.931 0.985
The data reported in the table are hand collected from company proxy statements as required by the SEC’s recently issued rule on auditor independence (SEC 2000b). Audit is the aggregate fees billed for professional services rendered for the audit of the annual financial statements and the reviews of the quarterly financial statements (million); IS is the aggregate fees (million) billed for financial information systems design and implementation; Other is the aggregate fees (million) billed for all services rendered other than the services covered by Audit and IS; Nonaudit is the sum of IS and Other; Total is the total fees billed.
21
Table 2. Descriptive Statistics for fiscal year 2000 (N=5,208)a
Variable Mean median standard deviation
SURPRISE FEERATIO FEERATIO1 FEERATIO2 NONAUDITFEE NONAUDITFEERANK TOTALFEE TOTALFEERANK BM MV CF ROA LOSS LITIRISK BIG5
a
0.364 0.534 0.333 0.156 1.989
b
0 0.558 0 0 0.382 65 0.752 66 0.384 608 0.020 0.013 0 0 1
0.481 0.217 0.471 0.363 5.943 26.871 7.315 26.566 0.574 31,406 0.057 0.056 0.418 0.483 0.186
61.456 2.776
b
62.695 0.542 6,234 0.015 0.005 0.225 0.372 0.964
The sample includes all quarterly observations in fiscal year 2000. SURPRISE is coded 1 for firms whose earnings surprises (defined as actual earnings per share minus the analysts’ latest median forecast before the earnings announcement) are zero or +1 cent per share, and zero otherwise (i.e., surprises greater than +1 cent per share and all negative earnings surprises). FEERATIO is the ratio of nonaudit fees to total fees for year 2000. FEERATIO1 is a dummy that is equal to 1 if FEERATIO is greater than 0.50 and the dollar value of nonaudit fees is larger than the 75th percentile value of the respective Big5 and non-Big5 samples, and zero otherwise. FEERATIO2 is a dummy that is equal to 1 if FEERATIO is greater than 0.50 and the dollar value of nonaudit fees is larger than the 90th percentile value of the respective Big5 and non-Big5 samples, and zero otherwise. NONAUDITFEE is the dollar value (in millions) of nonaudit fees. NONAUDITFEERANK is the percentile ranking (1 to 100) of the dollar value of nonaudit fees for each of the Big-Five auditors. TOTALFEE is the dollar value of audit and nonaudit fees (in millions). TOTALFEERANK is the percentile ranking (1 to 100) of the dollar value of both audit and nonaudit fees for each of the Big-Five auditors. BM is the ratio of book value to market value of common equity at the beginning of a quarter. MV is the market value of common equity (millions of dollars) at the beginning of a quarter. CF is quarterly operating cash flow scaled by lag-one total assets. ROA is quarterly earnings per share scaled by lag-one total assets per share. LOSS is coded 1 if quarterly earnings per share is negative, and zero otherwise. LITIRISK is coded 1 if a firm is in high litigation industries (i.e., SIC 2833-2836, 3570-3577, 7370-7374, 3600-3674, 5200-5961), and zero otherwise. BIG5 is 1 if a firm is audited by one of the Big-Five accounting firms, and zero otherwise.
b
The sample size for NONAUDITFEERANK and TOTALFEERANK is 5,021 because these two variables are defined only for the Big Five auditors.
22
Table 3. Replication and extension of Frankel et al. (2002) using quarterly analyst forecasts for fiscal year 2000a Regression model: SURPRISEit = a + ß 1 FEERATIOi + ß 2 BMit + ß 3 ln(MVit ) + ß 4 CFit + ß 5 ROA it + ß 5 LOSSit +ß 6 LITIRISKit + ß 7 BIG5i + eit ;
(1) All earnings surprises BM FFEERATIO ln (MV) CF ROA LOSS LITIRISK BIG5 Constant Observations -0.074 (0.100) 0.256 (0.023)* 0.077 (0.000)** -0.427 (0.327) 0.490 (0.393) -0.422 (0.000)** 0.071 (0.142) -0.125 (0.271) -0.822 (0.000)** 5208 (3) (4) Dependent variable = SURPRISE Earnings surprises Earnings surprises Earnings surprises are <-1 cent and 0 or are >0 cent are from –1 cent to +1 cent +1 cent -0.177 -0.010 -0.123 (0.002)** (0.859) (0.082) 0.354 0.191 0.282 (0.026)* (0.139) (0.132) 0.141 0.061 0.045 (0.000)** (0.000)** (0.077) -0.659 -0.058 0.373 (0.375) (0.902) (0.648) 2.599 -1.340 0.854 (0.019)* (0.048)* (0.361) -0.806 -0.266 -0.378 (0.000)** (0.001)** (0.000)** 0.121 0.043 0.141 (0.070) (0.434) (0.081) 0.028 -0.221 -0.042 (0.838) (0.124) (0.824) -0.211 -0.440 0.879 (0.234) (0.012)* (0.000)** 2783 4052 2163 (2) (5) Earnings surprises are >-1 cent -0.026 (0.615) 0.209 (0.087) 0.060 (0.000)** -0.031 (0.947) -1.071 (0.102) -0.289 (0.000)** 0.059 (0.264) -0.202 (0.126) -0.516 (0.002)** 4320
a
See Tables 1 and 2 for variable definitions. The sample in column (1) includes all quarterly observations in 2000 and replicates Frankel et al. (2002). Columns (2) through (5) differ only in terms of the observations included in the zero condition. The sample in column (2) includes quarterly observations in 2000 whose earnings surprises (defined as actual earnings per share minus the analysts’ latest median forecast before the earnings announcement) are either small positive surprises (i.e., zero and +1 cent per share), or large negative surprises (i.e., in excess of –1 cent per share). The sample in column (3) includes quarterly observations in 2000 whose earnings surprises are either small positive surprises (i.e., zero and +1 cent per share), or large positive surprises (i.e., in excess of +1 cent per share). The sample in column (4) includes quarterly observations in 2000 whose earnings surprises are either small positive surprises (i.e., zero and +1 cent per share), or small negative surprises (i.e., – 1 cent per share). The sample in column (5) includes quarterly observations in 2000 whose earnings surprises are either small positive surprises (i.e., zero and +1 cent per share), or small negative surprises (i.e., –1 cent per share) and large positive surprises (i.e., in excess of +1 cent per share). The model in column (5) is most consistent with Degeorge et al. (1999). All regressions include calendar quarter dummies and fiscal quarter dummies, and the coefficients on these control variables are omitted for brevity. Two-tailed p-values are reported in parentheses, and t-statistics are adjusted for serial correlation and heteroskedasticity per Rogers (1993). *significant at 5% level; ** significant at 1% level (two -tailed).
23
Table 4. Regression coefficients on alternative auditor dependence proxies from the following regression model (fiscal year 2000)a SURPRISEit = a + ß 1 PROXYi + ß 2 BMit + ß 3 ln(MVit ) + ß 4 CFit + ß 5 ROA it + ß 5 LOSSit +ß 6 LITIRISKit + ß 7 BIG5i + eit ; where PROXY is one of the six auditor dependence proxies included in the table.
Alternative auditor dependence proxies 1. FFEERATIO1 2. FFEERATIO2 3. ln (NONAUDITFEE) 4. NONAUDITFEERANK 5. ln (TOTALFEE) 6. TOTALFEERANK Observationsb
(1) (2) Dependent variable = SURPRISE All earnings Earnings surprises surprises are >-1 cent
0.073 (0.203) 0.064 (0.390) 0.068 (0.119) 0.001 (0.281) 0.010 (0.693) 0.000 (0.994) 5,208
0.101 (0.100) 0.090 (0.255) 0.093 (0.046)* 0.001 (0.230) 0.034 (0.218) 0.001 (0.496) 4,320
a
The table reports the regression coefficients on alternative auditor dependence proxies for different samples. The sample in column (1) includes all quarterly observations in 2000. The sample in column (2) includes quarterly observations in 2000 whose earnings surprises are either small positive surprises (i.e., zero and +1 cent per share), or small negative surprises (i.e., –1 cent per share) and large positive surprises (i.e., in excess of +1 cent per share). The model in column (2) is consistent with Degeorge et al. (1999). All regressions include calendar quarter dummies and fiscal quarter dummies, and the coefficients on these control variables are omitted for brevity. See Tables 1 and 2 for variable definitions. Two-tailed p-values are reported in parentheses, and t statistics are adjusted for serial correlation and heteroskedasticity per Rogers (1993). *significant at 5% level; ** significant at 1% level (two-tailed).
b
The sample size for the regressions using NONAUDITFEERANK and TOTALFEERANK is 5,021 in column (1) and 4,178 in column (2).
24
Table 5. Regression coefficients on alternative auditor dependence proxies from the following regression model estimated using observations that report both positive current quarterly earnings and positive changes in earnings relative to the same quarter of the prior yeara SURPRISEit = a + ß 1 PROXYi + ß 2 BMit + ß 3 ln(MVit ) + ß 4 CFit + ß 5 ROA it + ß 5 LOSSit +ß 6 LITIRISKit + ß 7 BIG5i + eit ; where PROXY is one of the seven auditor dependence proxies included in the table.
Alternative auditor dependence proxies 1. FFEERATIO 2. FFEERATIO1 3. FFEERATIO2 4. ln (NONAUDITFEE) 5. NONAUDITFEERANK 6. ln (TOTALFEE) 7. TOTALFEERANK Observations b
a
(1) All earnings surprises 0.114 (0.438) -0.041 (0.572) 0.000 (0.997) 0.008 (0.893) -0.001 (0.377) -0.051 (0.150) -0.003 (0.046)* 2,867
(2) Earnings surprises are >-1 cent 0.107 (0.479) -0.018 (0.808) 0.017 (0.861) 0.017 (0.761) -0.001 (0.347) -0.045 (0.212) -0.003 (0.061) 2,663
The table reports the regression coefficients on alternative auditor dependence proxies for different samples. The sample in column (1) includes all quarterly observations in 2000. The sample in column (2) includes quarterly observations in 2000 whose earnings surprises are either small positive surprises (i.e., zero and +1 cent per share), or small negative surprises (i.e., – 1 cent per share) and large positive surprises (i.e., in excess of +1 cent per share). The model in column (2) is consistent with Degeorge et al. (1999). All regressions include calendar quarter dummies and fiscal quarter dummies, and the coefficients on these control variables are omitted for brevity. See Tables 1 and 2 for variable definitions. Two-tailed p-values are reported in parentheses, and t-statistics are adjusted for serial correlation and heteroskedasticity per Rogers (1993). *significant at 5% level; ** significant at 1% level (two-tailed).
b
The sample size for the regressions using NONAUDITFEERANK and TOTALFEERANK is 2,775 in column (1) and 2,579 in column (2).
25