Non-Audit Service Fees and Financial Reporting Quality: A Meta-Analysis

The purpose of this paper is to present a meta-analysis of the association between fees for non-audit services or non-audit fees (hereafter NAF) and financial reporting quality (hereafter FRQ). The well-known agency problem between shareholders and managers gives rise to the hiring of auditors who provide independent assurance to corporate stakeholders that financial statements prepared by corporate managers comply with generally accepted accounting principles (GAAP) (Watts and Zimmerman, 1983). Auditing also plays a significant role in enforcing and protecting investors' rights by detecting expropriation by insiders (Newman et al., 2005) and benefits management by providing a signalling mechanism regarding the reliability of management-provided financial information. However, the discharge of such responsibilities requires auditors to be independent, both in fact and in appearance. An auditor who is independent in fact has the ability to make independent audit decisions despite the perceived lack of independence or if the auditor is placed in a potentially compromising position. Independence in fact, however, does not necessarily imply that auditors are perceived as independent by the stakeholders (appearance of objectivity).

Although it is difficult to prove independence in fact, it is possible using large samples to determine the association between NAF and some accounting outcomes, for example, discretionary accruals, earnings conservatism to assess auditor independence (DeFond et al., 2002; Frankel et al., 2002; Ashbaugh et al., 2003; Chung and Kallapur, 2003). With respect to independence in appearance, archival research investigates the association between NAF and market valuation of accounting earnings to assess whether independence in appearance has been impaired due to the joint provision of audit and non-audit services (hereafter NAS) (Krishnan et al., 2005; Francis and Ke, 2006; Ghosh et al., 2009).

The idea that high NAF may have a detrimental effect on the FRQ is rooted in the theoretical model developed by DeAngelo (1981) who relates auditor independence with client-specific future quasi rents defined as ‘the excess of revenues over avoidable costs, including the opportunity cost of auditing the next-best alternative client’ (italics in original, p.116). She develops a two-dimensional definition of audit quality that sets the standard for addressing auditor independence issue. An auditor will be considered independent if he/she can detect a material misstatement and then report it to the audit service consumers. If no client-specific quasi-rents are expected from a given client relationship then the auditor should have no economic incentive to conceal financial misstatements and therefore should be considered perfectly independent from the client. However, if the client-specific future quasi rents are high then auditors may compromise independence to retain such quasi rents. It has been argued that NAS provides more quasi rents than do audit services because of the higher margin derived from the former and therefore is more likely to impair auditor independence (Securities and Exchange Commission [hereafter SEC], 2000), particularly independence in appearance (Lindberg and Beck, 2004).¹

The provision of NAS, however, provides considerable economies of scope. These economies of scope are broadly categorized into knowledge spillover benefits (benefits from transferring information and knowledge), and contractual economies (making better use of assets and/or safeguards already developed when contracting and ensuring quality in auditing) (Simunic, 1984; Beck et al., 1988; Arrunada, 1999). Both impairment of independence and economies of scope could co-exist although studies examined in this meta-analysis test for the effect of high NAF on impairment of independence irrespective of whether economies of scope is achieved. Narrative literature reviews on the consequences of NAS generally conclude that extant archival research on the effect of NAF on FRQ provides mixed evidence (Beattie and Fearnley, 2002; Schneider et al., 2006; Anandarajan et al., 2010).

Although narrative literature reviews may include a large number of studies on particular research themes, such reviews can be misleading and often inconclusive (Hunter and Schmidt, 1990). In some cases there may be several studies with varying results that are subject to variations in sample size, time period, and setting of the study. As a result, different researchers may reach different conclusions about a set of individual studies. A narrative literature review will report these as apparently inconsistent results and call for further research, which may also produce inconsistent results and further cloud the issue. By contrast, meta-analysis statistically aggregates results across individual studies and corrects for statistical artefacts like sampling and measurement error and, thereby, provides much greater precision with respect to the findings compared with narrative reviews (Hay et al., 2006). Meta-analysis is particularly effective in reconciling results that are inconsistent across studies like the effect of NAF on FRQ.

This meta-analysis is expected to contribute to the debate concerning whether regulation restricting the provision of NAS is justifiable. The recent most comprehensive corporate governance law in the form of the Sarbanes-Oxley Act 2002 (hereafter SOX-2002) prohibited nine specific NAS. The drafters of this regulation relied on an important piece of academic research conducted by Frankel et al. (2002), who document that firm-level absolute discretionary accruals and the probability of meeting and or beating analyst forecasts is positively associated with the level of NAF.² However, subsequent studies using different research methodologies fail to support Frankel et al.'s findings (Ashbaugh et al., 2003; Chung and Kallapur, 2003). These contradictory findings call for an overall assessment of the state of the current research on the association between NAF and FRQ. Such an assessment is expected to inform regulators who are in the process of considering whether to impose sanctions on the provision of NAS. Francis (2006, p. 748) notes, ‘it is clear that the appropriateness of auditor-provided NAS [non-audit services] continues to be controversial and viewed with skepticism by regulators. For this reason, it is important that there be ongoing research to facilitate well informed policy making by regulators with respect to the costs and benefits of restricting the scope of NAS to audit clients’ (italics added).

Prospective investors can use reliable meta-analysis results to assess the risk of making investment decisions based on accounting information audited by auditors who may have compromised their independence and, hence, the integrity of the financial statements because of the joint provision of audit and NAS. The importance of conducting meta-analysis is further justified because there has been less such research in the accounting discipline. Pomeroy and Thornton (2008) identify 14 previous meta-analyses in the accounting discipline on internal control judgment (Trotman and Wood, 1991), accountants' job satisfaction (Brierley, 1999), corporate characteristics and disclosure levels (Ahmed and Courtis, 1999), corporate social and financial performance (Orlitzky et al., 2003), and audit fees (Hay et al., 2006).

A recent meta-analysis by Lin and Hwang (2010) integrates results from 48 prior studies on the determinants of earnings management, and provides evidence that NAF is positively associated with corporate earnings management based on a sub-set of ten studies. This meta-analysis differs from Lin and Hwang in three important respects. First, this meta-analysis includes eight studies that provide evidence on the association between NAF and market perception of accounting earnings. This is particularly important because there is evidence to suggest that NAF is negatively associated with earnings response coefficients (hereafter ERC) although the evidence on auditor independence in fact is rather mixed (Beattie and Fearnley, 2002). Lin and Hwang (2010) do not include these studies because their dependent variable is some form of earnings management measures. Second, this study broadens the definition of FRQ by incorporating auditor reporting decisions as an important construct of the FRQ. The literature on the association between NAF and reporting decisions has produced very mixed evidence and is a good testing ground for meta-analysis. Third, the sample studies selected for this meta-analysis significantly outnumber Lin and Hwang's sample (45 versus 10 studies, respectively).

Following the meta-analysis procedures outlined in Lipsey and Wilson (2001), this study provides statistical evidence that client-provided NAF is associated with reduced FRQ, as is evident from an average effect size (hereafter ES) of 0.021 which is statistically significant at better than the 1% level. Further analysis decomposing the FRQ proxies into individual components reveals that earnings management increases and auditor's propensity to issue qualified audit opinion decreases with the level of NAF indicating impairment of independence in fact. Importantly, meta-analysis of the market-based studies reveals that ERC is negatively associated with the level of NAF, indicating a lack of independence in appearance because of the joint provision of audit and NAS.

1: NAS FEES AND AUDITOR INDEPENDENCE

An important theoretical rationale for the impairment of auditor independence caused by joint provision of audit and NAS was provided by DeAngelo (1981). She coined the term ‘quasi rent’ in the context of audit fee-setting and used this concept to theorize auditor independence. Incumbent auditors enjoy a comparative cost advantage over new auditors because the latter group must incur technological start-up costs with each audit. Clients also incur auditor switching costs. Both the technological and switching cost advantages allow the incumbent auditor to set the audit fees in a manner that generates future quasi rents for the incumbent auditors. Both the client and the incumbent auditor can impose real costs on the other. For example, the auditor could demand higher audit fees and the client could ask for clean audit opinion. The existence of future quasi rents, therefore, weakens the auditors' incentives to be independent (DeAngelo, 1981). Subsequent empirical research relied on this theoretical notion to test the effect of this quasi rent in the form of NAF on FRQ.

However, opponents of DeAngelo's (1981) theory argue that the provision of NAS actually increases audit efficiency. This argument is derived from the conjecture that providing both audit and NAS provides considerable economies of scope. These economies of scope are broadly categorized into knowledge spillover benefits (benefits from transferring information and knowledge) and contractual economies (making better use of assets and/or safeguards already developed when contracting and ensuring quality in auditing) (Simunic, 1984; Beck et al., 1988., Arrunada, 1999; Beattie and Fearnley, 2002). Simunic (1984) interprets the efficiencies from joint production of audit and NAS as follows:

Beck et al. (1988) distinguish NAS into recurring and non-recurring services and propose that recurring NAS give rise to knowledge spillovers and reduce the threat to independence, while the opposite effect is experienced with non-recurring NAS. Another argument for non-impairment of independence is advanced by Arrunada (1999), who suggests that the provision of NAS can also increase the auditor's investment in reputational capital, which the auditor is not likely to jeopardize to satisfy the demand of one particular client (Arrunada, 1999).

A number of narrative literature reviews have summarized the current state of archival research on the effect of NAF on FRQ. For example, Beattie and Fearnley (2002) and Schneider et al. (2006) review empirical studies of (a) the nature and magnitude of NAS fees; (b) the determinants of NAS purchasing decisions; (c) the impact of joint services provision from survey and experimental studies; (d) the association between joint provision, pricing and audit tenure; (e) the impact of NAS on auditor reporting decisions; and (f) the association between NAF and earnings quality. A brief summary of the review literature follows.

Beattie and Fearnley (2002) conclude from a comprehensive literature review that there is very little clear support for the view that joint provision of audit and NAS impairs independence in fact. They attribute this mixed evidence to the measurement of the proxy variable for auditor independence and the validity of the proxy itself. However, there is evidence to suggest that joint provision of audit and NAS adversely affects user perceptions. With respect to NAF and earnings quality, Schneider et al. (2006) conclude that extant archival studies provide some evidence consistent with NAF reducing earnings quality, but the findings are limited to specific circumstances and hence may not be generalizable. For audit opinion decisions, the authors conclude that U.S. studies provides little evidence that NAS reduces auditor independence. Evidence from Australia and the U.K. is, however, mixed and the differences in sample composition, reporting environment and the type of NAS provided may be responsible for these mixed findings. Anandarajan et al. (2010) conclude that existing research has provided mixed evidence on whether the provision of NAS impairs auditor independence. The authors raise concerns on the findings of the extant research since the research has unambiguously used NAS as the surrogate for impairment of independence. Survey evidence seems to suggest that a higher level of NAS is not necessarily perceived as causing impairment of independence (Dopuch et al., 2003). Ikin (2005) concludes that archival research provides no convincing evidence that high level of NAS impairs auditor independence in fact, as measured by auditor tenure, earnings management or earnings conservatism. However, the perception-based academic evidence seems to suggest that financial statement users perceive that NAS impairs auditor independence.

Figure 1 provides a schematic representation of the consequences of NAS categorized primarily into knowledge spillover and impairment of independence effects.³ Researchers have used three primary research methodologies, namely survey-based, experimental and large-sample archival research approaches to investigate the consequences of NAS.⁴ Archival research methodology has the advantage of‘incorporating all the richness of a capital market and its participants’ actual behaviour; however, a general weakness of archival studies is that the richness of the environment frequently makes it difficult, if not impossible, to pinpoint specific characteristics or situations' (Schneider et al., 2006, p. 205). It is evident from Figure 1 that the FRQ proxies used in the archival research are quite heterogeneous. Similarly, there is also disagreement among researchers regarding which NAF proxy better captures the economic dependence of incumbent auditors on the clients. The variables most frequently used to measure the importance of NAS to the auditing firm are the fee ratio (the ratio of NAF to total fees) and total fees (the sum of NAF and audit fees) paid to the external auditor; others include fee measures that adjust the amounts by client to construct a proxy for the client's importance to the auditor and percentile ranks, by auditor, of a firm's non-audit and audit fees.

Archival research uses an agency theory explanation for investigating the effect of NAF on FRQ and argues that auditors are rational wealth maximizers who would be intentionally biased towards compromising audit quality in order to generate wealth for themselves. Behavioural literature, on the other hand, suggests that psychological heuristics unconsciously lead auditors to make biased judgments. Although the cause of auditor bias differs in these two streams of literature, the ultimate effect remains the same: ‘auditors are more likely to acquiesce to client pressure, including pressure to allow earnings management, when the provision of nonaudit services generates economic rents’ (Frankel et al., 2002, p. 75). Frankel et al. (2002) report a positive association between NAF and earnings management consistent with this hypothesis. Their finding, however, has been challenged by at least three other subsequent papers (Ashbaugh et al., 2003; Chung and Kallapur, 2003; Reynolds et al., 2004). Francis (2004) offers a possible explanation for this stream of replicated research:

Another FRQ proxy that received significant research attention is the impact of NAF on audit reporting decisions. If the joint provision of audit and NAS impairs auditor independence then it is likely that auditors will issue far more unqualified audit reports than warranted. In an early study, Wines (1994) reports that NAF is negatively associated with audit qualification. Barkess and Simnett (1994) and Craswell (1999), however, fail to find any significant association between the two. But Sharma and Sidhu (2001) support Wines (1994) by focusing on a sample of distressed firms. Researchers have also used earnings conservatism, restatements, and market valuation of accounting earnings as useful FRQ proxies. The following section develops testable hypotheses on the association between NAF and FRQ proxies and attempts to provide a critical evaluation of the FRQ proxies.

2: DEVELOPMENT OF TESTABLE HYPOTHESES

Although FRQ occupies the central role in the information provision to outsiders, the precise definition of FRQ remains elusive. In considering FRQ, the overwhelming majority of research has focused on earnings quality. Earnings are widely used as a key performance indicator of business success, commonly employed in compensation and debt arrangements. A recent comprehensive survey of chief financial officers by Graham et al. (2005) show the GAAP earnings number, especially the earnings per share (EPS), is the key metric upon which the market focuses. This is mainly because, for evaluating a firm's performance, investors need a simple benchmark, like EPS, that reduces the costs of information processing (Graham et al., 2005, p. 21). Also, valuation theory has long posited a relationship between earnings and the value of common stock (Graham et al., 1962; Miller and Modigliani 1966). The most widely used earnings quality measure is some form of abnormal accruals measure as estimated by popular models like Jones (1991) and extended by Dechow et al. (1995). Following the agency-theory relationship developed earlier, auditors are more likely to acquiesce to client pressure, including pressure to allow earnings management, when the provision of NASs generates economic rents.⁵ Therefore, the following hypothesis is formulated:

However the reliability of the abnormal accruals models used in the literature is questionable. Unfortunately there is no best accruals model available and researchers usually validate the robustness of the findings by using alternative accruals models (Dechow et al., 2003). Another major problem with using abnormal accruals as a proxy for opportunistic earnings management is that managers could also use discretionary accruals as an efficient contracting device to provide credible signals to the marketplace about the future growth prospects of the organizations. However, earnings management-based accounting literature generally theorizes opportunistic motives for using abnormal accruals. Furthermore, managers could manipulate accounting information using a combination of accruals management, real activities manipulation (Roychowdhury, 2006), and classification shifting mechanisms (McVay, 2006).

Auditor provided qualified audit opinion has been used as another proxy for the FRQ. Unlike earnings management research which reflects actual reported numbers, qualified audit opinion is provided by the auditors by taking into consideration a large number of factors including quality of earnings before issuing such opinions. Issuance of such a qualified opinion is a matter of considerable importance to the auditors because such opinions could lead to company bankruptcy (the so-called self-fulfilling prophecy feature). The literature on the association between NAF and auditors' propensity to issue qualified audit opinions provides very mixed evidence. For example, two early Australian studies (Wines, 1994; Sharma and Shidhu, 2001), report a negative association between the two whereas Barkess and Simnett (1994) and Craswell (1999) find no significant association between NAF and qualified audit opinion decisions. None of the U.S. studies provide any evidence of an adverse effect of NAF on qualified audit opinion decisions; if anything there seems to be evidence that provision of tax services is associated with more accurate qualified audit opinion decisions (Robinson, 2008). The following testable hypothesis is developed to assess the effect of NAF on the propensity of auditors to issue qualified audit opinions:

FRQ has also been proxied by earnings conservatism and earnings restatement. Researchers use the notion of conservatism to imply conditional conservatism which requires a higher degree of verification for recognizing good news than bad news in financial statements (Basu, 1997). This timely loss recognition property mitigates agency problems between shareholders and managers (Kothari et al., 2010). The choice of conditional conservatism as an earnings quality measure is appropriate in an audit setting. Because published financial statements are a joint product of managers' representations and the audit process, earnings quality represented by the conditional conservatism property would also be expected to reflect audit quality (Ruddock et al., 2005). Because NAF are widely believed to contain higher profit margins than traditional audit services, auditors may be less inclined to issue conditionally conservative financial statements and hence compromise the integrity of financial statements with the level of NAF. Therefore the following hypothesis is formulated:

The conditional conservatism measure, however, has been challenged from different quarters. One common argument against this measure is that it proxies for some other properties of firms and/or financial reporting because of a negative relation between the conservatism measure and the market-to-book ratio and, hence, biases the asymmetric timeliness coefficient. Another serious concern is raised by Dietrich et al. (2007) who argue that the model is mis-specified, and produces biased coefficients since the model uses earnings as the dependent variable rather than the stock returns which is conventional in earnings response coefficient research. However, Ball et al. (2009) provide an econometric analysis of the Basu (1997) conservatism measure and conclude that these challenges are misconstrued.

With respect to earnings restatement, The U.S. General Accounting Office (GAO, 2002) documents a total of 919 restatements from January 1997 to June 2002. Academic research reveals that restating firms lose about 10% of the market value on the day of such announcements (Palmrose et al., 2004). Efendi et al. (2007) report a positive association between stock-based compensation arrangements and subsequent earnings restatements. If auditors compromise their independence because of the joint provision of audit and NAS, then these earnings manipulation and/or violation of GAAP transactions will go unreported, and lead to lower quality audits and reviews. Some of these violations are likely to be revealed through restatements of financial statements (Kinney et al., 2004). The following hypothesis is, therefore, developed:

The FRQ variables described above are all used in academic research to investigate the impact of NAF on the impairment of independence in fact. To examine independence in appearance, archival research has examined the market valuation of accounting earnings (conventional ERC) conditional on NAF. The underlying premise is that if investors perceive audit independence to be impaired because of NAF, ‘[they] will place less reliance on the auditors’ attestation of its client's financial statements, and perceive greater noise in the client's reported earnings' (Krishnan et al., 2005). The following hypothesis tests this conjecture:

The major problem associated with using capital market studies is the fact that the stock market is flooded with information on a continuous basis. It is difficult to control precisely for alternative information that could potentially confound the impact of NAF disclosures on stock returns.

3: METHODS

3.3: Meta-analysis Procedures

The purpose of this meta-analysis is to determine (a) the net effect of NAF on FRQ; and (b) whether there is homogeneity in the FRQ proxies used in the extant literature. The following standard meta-analysis procedures as suggested by Lipsey and Wilson (2001) are performed to fulfil these objectives.

• 1

  Step 1: Convert test statistics, such as p-values or t-statistics reported in different studies to a standard correlation measure called effect size (ES). The majority of the studies canvassed for meta-analysis in this paper reported p-values. Reported p-values are converted into t-statistics, and ES then is calculated as follows:

  ()
  where, t is the reported or converted t-statistics and df is the degree of freedom calculated as (n− 3) where n is the number of firm-years observations for each studies.

• Step 2: Compute mean ES using weighting factors based on assumptions regarding the homogeneity of variances across studies.

• Step 3: Test the homogeneity assumptions in step 2 above.

While conducting this meta-analysis a number of adjustments had to be made. First, the selected studies differ with respect to the proxy variable chosen to represent auditor independence. The most widely used proxy is the fee ratio (FEERATIO) defined as NAF divided by sum of audit and NAF for a particular client. This ratio has been used in earlier NAF-related research (Scheiner, 1984; Parkash and Venable, 1993; Barkess and Simnett, 1994; Firth, 1997) and is also consistent with the SEC (2000) position that this measure is useful for investors to gauge auditor independence. Another argument for using FEERATIO as a measure to assess auditor independence ‘could be that there are only certain specific types of nonaudit services that impair an auditor's independence . . . As the ratio of nonaudit to audit fee increases, it could become more likely that the nonaudit services provided include one of those that impair independence’ (Chung and Kallapur, 2003, p. 949). Table 2 Panels A and B provide meta-analysis evidence on the association between FRQ and the level of NAF using FEERATIO as the independence proxy based on 38 studies.

Table 2. EFFECT SIZE CALCULATION

Panel A: ES calculation and significance of the studies using FEERATIO as the independence proxy
FRQ	Author	Stat. value	N	ES	ESz	W (df)	W*ES	W*ES2
EM	Frankel et al. (2002)	0.01	2,472	0.05	0.05	2,469	123.45	6.17
	Ashbaugh et al. (2003)	0.02	3,069	0.04	0.04	3,066	122.64	4.91
	Chung and Kallapur (2003)	1.52	1,853	0.04	0.04	1,850	74.00	2.96
	Larcker and Richardson (2004)	1.69	5,103	0.02	0.02	5,100	102.00	2.04
	Reynolds et al. (2004)	0.233	2,507	−0.02	−0.02	2,504	−50.08	1.00
	Ferguson et al. (2004)	0.001	610	0.13	0.13	607	78.91	10.25
	Antle et al.(2006)	2.47	2,294	0.05	0.05	2,291	114.55	5.73
	Dee et al. (2006)	0.60	384	−0.027	−0.03	381	−11.43	0.28
	Habib and Islam (2007)	0.47	530	−0.03	−0.03	527	−15.81	0.47
	Mitra (2007)	1.20	71	0.14	0.14	68	9.52	1.33
	Srinidhi and Gul (2007)	0.01	4,282	0.04	0.04	4,279	171.16	6.85
	Lim and Tan (2008)	1.40	4,943	0.02	0.02	4,940	98.80	1.98
Audit opinion	Wines (1994)	−1.83	76	0.21	0.21	73	15.33	3.21
	Craswell (1997)	1.01	1,145	−0.03	−0.03	1,142	−34.26	1.03
	Lennox (1999)	2.05	2,244	−0.05	−0.05	2,241	−112.05	5.60
	Sharma and Shidhu (2001)	0.046	49	0.29	0.3	46	13.80	3.87
	DeFond et al. (2002)	0.49	1,158	0.02	0.02	1,155	23.10	0.46
	Firth (2002)	−1.96	1,112	0.06	0.06	1,109	66.54	3.99
	Raghunandan (2003)	3.92	300	0.22	0.22	297	65.34	14.38
	Hay et al. (2006)	0.52	643	0.03	0.03	640	19.20	0.576
	Fargher and Jiang (2008)	0.003	5,113	0.04	0.04	5,110	204.40	8.176
	Lim and Tan (2008)	1.61	1,692	−0.039	−0.04	1,689	−67.56	2.57
	Callaghan et al. (2009)	0.51	92	−0.07	−0.07	89	−6.23	0.44
	Liu et al. (2009)	3.60	194	0.25	0.26	191	49.66	11.94
Earnings conservatism	Ruddock et al. (2005)	1.23	2,497	0.024	0.02	2,494	49.88	1.44
	Zhang and Emanuel (2008)	−0.95	352	−0.05	−0.053	349	−18.78	0.87
Earnings restatement	Agrawal and Chadha (2005)	1.44	318	0.08	0.08	315	25.2	2.02
	Li and Lin (2005)	0.82	351	0.0439	0.044	348	15.31	0.67
	Bloomfield and Shackman (2008)	−1.86	500	−0.08	−0.082	497	−39.86	3.18
Capital market valuation	Brandon et al. (2004)	0.023	333	0.12	0.13	330	42.9	4.75
	Krishnan et al. (2005)	−3.22	2,816	0.06	0.06	2,813	168.78	10.13
	Gul et al. (2006)	−2.46	840	0.08	0.08	837	66.96	5.36
	Francis and Ke (2006)	−3.96	16,243	0.03	0.03	16,240	487.2	14.62
	Chin et al. (2007)	2.46	254	0.15	0.15	251	37.65	5.65
	Dhaliwal et al. (2008)	2.11	560	0.09	0.09	557	50.13	4.51
	Lim and Tan (2008)	−2.92	2,935	0.05	0.05	2,932	146.6	7.33
	Ghosh et al. (2009)	1.03	21,762	−0.007	−0.007	21,759	−152.31	1.07
	Niu (2008)	−0.43	814	0.02	0.02	811	16.22	0.32
			92,511			92,397	1,951	162.13
	Mean ES	0.021
	Standard error (SE)	0.0033
	z-statistic	6.37***
	Lower bound	0.015
	Upper bound	0.027
	Homogeneity test (Q-values)	118.33***

Panel B: Component of the FRQ proxies and ES analysis with FEERATIO as the primary independence proxy
	Earnings management (n= 12)	Audit opinion/ratification (n= 12)	Earnings conservatism (n= 2)	Earnings restatement (n= 3)	Capital market studies (n= 9)
Mean ES	0.029	0.017	0.011	0.0006	0.019
SE	0.0061	0.0085	0.019	0.0294	0.0046
z-statistic	4.88***	2.02**	0.583	0.019	4.01***
Lower limit	0.017	0.0005	−0.03	−0.057	0.0095
Upper limit	0.036	0.034	0.048	0.0581	0.028
Q-values	20.16***	52.15***	1.97	5.87**	37.68***
Within group Q	117.83***
[20.16+52.15+1.97+5.87+37.68]
Between group Q (Panel A)	118.33***

• Note: The sample consists of 38 archival studies and 92,511 observations investigating the effect of NAF (proxied by FEERATIO) on five different FRQ proxies. ES is calculated using the following formula:
• ()
• The formula is straightforward to apply when the sample studies reported t-statistics. The majority of the studies canvassed for meta-analysis, however, reported p-values. Reported p-values are converted into t-statistics using p-value tables.
• Standard error is calculated using the following formula:
• 
• 
• Lower and upper bound of the ES are calculated as ES- (1.96*SE) and ES+ (1.96*SE) respectively.
• To test whether the various ESs that are averaged into a mean value, are estimating the same population ES, a Q-statistic, which is distributed as a chi-square with k−1 degrees of freedom where k is the number of ES, is calculated using the following formula (Lipsey and Wilson, 2001, p. 115).
• ()
• Where effect size_i (ES) is the individual ES for i = 1 to k, and is the weighted mean ES over the k ES, and w_i is the individual weight for effect size_i.
• Note: Within group Q is distributed as a chi-square with 33 degrees of freedom (number of individual ES 38 minus 5 FRQ categories) and yields a critical value of 54.78 at 1% probability level. This rejects the hypothesis that the residual variability within each FRQ proxy is homogeneous. Between group Q is the reported Q in Panel A, i.e., 118.33. The difference between these two Q is not statistically significant at the critical level (critical value of Q with 4 degrees of freedom at 5% level is 9.49) and thereby rejects the presence of any between-group effect.
• *** and ** denote statistical significance at 1% and 5% level respectively (two-tailed test).

This measure, however, gives equal weight to small and large NAF whereas the incentives for auditors to compromise independence should be stronger when the size of NAF is larger. Ashbaugh et al. (2003), therefore, argue that FEERATIO does not capture the economic importance of the client when total client fees are immaterial to the audit firms and hence recommend using total fees (TOTFEE) as a more appropriate proxy to measure auditor's dependence. Table 3 Panels A and B provide meta-analysis evidence based on FEERATIO proxy. Other proxies include log of NAF, rank of NAF, and components of NAF, for example, audit-related fee ratio, tax ratio, and other fee ratios.⁷

Table 3. EFFECT SIZE MEASURE WITH ALTERNATIVE NAF POLICY

Panel A: ES measures using total fees (TOTFEE) as the NAF proxy
FRQ	Author	Stat. value	N	ES	ESz	W (df)	W*ESz	W*ES2
EM	Frankel et al. (2002)	0.16	2,472	0.03	0.03	2,469	74.07	2.222
	Ashbaugh et al. (2003)	0.59	3,069	0.01	0.01	3,066	30.66	0.307
	Larcker and Richardson (2004)	−6.21	5,103	−0.09	−0.09	5,100	−459.00	41.310
	Reynolds et al. (2004)	0.35	2,507	0.02	0.02	2,504	50.08	1.002
	Ferguson (2004)	0.04	610	0.08	0.08	607	50.08	3.89
	Dee et al. (2006)	0.02	384	0.12	0.12	381	45.72	5.486
	Habib and Islam (2007)	0.14	530	0.07	0.07	527	36.89	2.582
	Mitra (2007)	0.31	71	0.13	0.13	68	8.84	1.149
	Srinidhi and Gul (2007)	0.05	4,282	0.03	0.03	4,217	126.51	3.80
	Gul (2007)	0.05	4,720	−0.03	−0.03	4,717	−141.51	4.245
	Lim and Tan (2008)	0.65	4,943	0.0092	0.0092	4,940	45.45	0.42
Audit opinion	DeFond et al. (2002)	0.12	1,158	0.05	0.05	1,155	57.75	2.89
	Lim and Tan (2008)	8.31	1,692	−0.20	−0.20	1,689	−337.80	67.56
	Callaghan et al. (2009)	0.83	92	0.03	0.03	89	2.67	0.08
Capital market valuation	Brandon et al. (2004)	0.001	333	0.18	0.18	330	60.06	10.69
	Krishnan et al. (2005)	−1.68	2,816	0.03	0.03	2,813	84.39	2.53
	Chin et al. (2007)	0.96	254	0.06	0.06	251	15.06	0.90
	Dhaliwal et al. (2008)	2.45	560	0.10	0.10	557	55.70	5.57
	Ghosh et al. (2009)	−2.25	21,762	0.015	0.015	21,759	326.39	14.30
	Lim and Tan (2008)	−3.16	2,935	0.06	0.06	2,932	175.92	10.56
	Mean ES	0.0051
	SE	0.004
	z-statistic	1.26
	Lower bound	−0.0029
	Upper bound	0.013
	Homogeneity test	170.60***

Panel B: Component of FRQ proxies and ES analysis (TOTFEE as the independence proxy)
FRQ	Earnings management (11 studies)	Audit opinion/auditor ratification (3 studies)	Capital market studies (6 studies)
Mean ES	−0.0046	−0.09	0.025
SE	0.0059	0.018	0.006
z-statistic	−0.78	−5.12***	4.24***
Lower limit	−0.016	−0.13	0.013
Upper limit	0.007	−0.06	0.037
Q-values	65.85***	44.30***	17.17***
Within group Q	127.32***
Between group Q (from Panel A)	170.60***

• Note: The sample consists of 20 archival studies and 60,293 observations investigating the effect of NAF (proxied by TOTFEE) on five different FRQ proxies. ES is calculated using the following formula:
• ()
• The formula is straightforward to apply when the sample studies reported t-statistics. The majority of the studies canvassed for meta-analysis, however, reported p-values. Reported p-values are converted into t-statistics using p-value tables.
• Standard error is calculated using the following formula:
• 
• 
• Lower and upper bound of the ES are calculated as ES- (1.96*SE) and ES+ (1.96*SE) respectively.
• To test whether the various ESs that are averaged into a mean value, are estimating the same population ES, a Q-statistic, which is distributed as a chi-square with k− 1 degrees of freedom where k is the number of ES, is calculated using the following formula (Lipsey and Wilson, 2001, p. 115).
• ()
• Where effect size_i (ES) is the individual ES for i = 1 to k, and is the weighted mean ES over the k ES, and w_i is the individual weight for effect size_i.
• Note: Within group Q is distributed as a chi-square with 17 degrees of freedom (number of individual ES 20 minus 3 FRQ categories) and is significantly higher than the critical Q-value of 33.4 at 1% probability level. This rejects the hypothesis that the residual variability within each FRQ proxy is homogeneous. Between group Q is the reported Q in Panel A, i.e., 170.6. The difference between these two Q is significantly higher than the critical Q-value of 9.21 (2 degrees of freedom at 1% level) and reveals the presence of between-group variability.
• *** denotes statistical significance at 1% level (two-tailed test).

Another difficulty associated with conducting meaningful meta-analysis relates to the heterogeneity of the FRQ proxies used. For example, when considering earnings management, some researchers used discretionary accruals in the primary analysis and the propensity of managers to meet or beat earnings targets (benchmark beating phenomena) in the supplementary analysis as proxies for FRQ (Frankel et al., 2002; Ashbaugh et al., 2003; Dee et al., 2006). This concern, however, is primarily limited to only four studies in the earnings management category that use DACCR and earnings benchmarks as alternative earnings management proxies. I provide a sensitivity analysis to assess the impact of these studies with multiple dependent variables on the overall results in section 5.1.

Finally, the ‘file drawer’ problem is tackled by calculating the fail-safe number, which reflects the number of studies failing to report significant results that would be required to reverse a conclusion about a significant relationship between the dependent and independent variables. The number is calculated as follows (Rosenthal, 1991, p. 261):

• 1

  I converted all t-statistics into their respective p-values and then all p-values are converted into z-statistics. Individual z-statistics are then combined using the following formula:

  ()
  Where N is the number of studies included in the meta-analysis and Z is the converted z-statistics.

• 2

  Fail-safe number is calculated using the following formula proposed by Rosenthal (1991):

  ()

where k is the number of studies in the meta-analysis and z is the combined standard z-value for the meta-analysis. The file drawer issue as represented by the file safe number only becomes a problem when N_fs exceeds the critical value calculated as follows:

()

Notwithstanding the popularity of meta-analysis in aggregating studies to provide statistical evidence, a caveat is in order. All the studies selected for meta-analysis use some form of multivariate regression analysis to investigate the research questions. Lipsey and Wilson (2001, pp. 15–16) note that such multivariate analysis ‘can not generally be represented in an effect size statistic but a study may report the correlation matrix upon which the multiple regression is based . . . Meta-analysts have not yet developed effect size statistics that adequately represent this form [multivariate regression] of research finding.’ Notwithstanding this limitation, researchers have used meta-analysis in a number of academic areas to resolve inconsistent findings and this meta-review is one such example.

4: RESULTS

Table 2 Panel A reports the ES measures of the selected studies. The ES measure is an essential feature of meta-analysis which ‘produces a statistical standardization of the study findings such that the resulting numerical values are interpretable in a consistent fashion across all the variables and measures involved’ (Lipsey and Wilson, 2001, p. 4). ES calculations for studies that report t-statistics were straightforward using the procedure outlined in 3.3 above. The majority of the selected studies, however, report p-values, and those values had to be converted into t-statistics. To do that I used two-tailed t-values from the t-distribution table in Lipsey and Wilson (2001). Table 2 also reports effect size_zr which is the transformed correlation and standard error (SE) of the distribution using the following formula (Lipsey and Wilson, 2001, p. 64):

()

I coded every study ES as positive if it supported the expectation that NAF are negatively associated with FRQ. For example, studies documenting a positive association between earnings management and NAF imply low FRQ. On the other hand, studies documenting a negative association between NAF and unexpected earnings also imply low FRQ. If I had taken the original signs reported in the study then it would fail to capture the fact that both these groups of studies provide evidence that high NAF reduce FRQ. Therefore, to be consistent with the underlying interpretation, I coded every effect size as positive if it supported the hypothesis that high NAF reduce FRQ, even though the actual coefficient signs reported in the original studies could show otherwise. Reported results in Table 2 Panel A reveal that the overall mean ES is 0.021 with an associated z-statistic of 6.37 which is statistically significant at the 1% level, implying that the level of NAF has a negative effect on FRQ (the signs of these ES are reversed from negative to positive to be consistent with the definition of low FRQ).

Looking at the confidence interval of the ES distribution, it is concluded that there is a 0.95 probability that the true association between NAF and FRQ is between 0.015 and 0.027. Cohen (1977, 1978, cited in Lipsey and Wilson, 2001) reports a general rule of thumb statistic for interpreting ES values with ES≤0.20 representing the small range (analogous values for the correlation ES is 0.10). Pomeroy and Thornton (2008) report an average ES of −0.08 in their meta-analysis of the effect of audit committee independence on FRQ.

Table 2 also reveals that nine of the 12 studies on the association between earnings management and the level of NAF report positive ES supporting the conjecture that earnings management increases with the level of NAF. Eight of the 12 studies on the association between NAF and qualified audit opinion find that auditor propensity to provide clean opinion increases with the level of NAF. Archival research on the effect of NAF on the impairment of independence in appearance provides more robust evidence. Six of the nine studies on the association between NAF level and ERC and cost of debt reveal a detrimental effect of the NAF on these proxy measures. For example, Krishnan et al. (2005) document a negative ERC for firms receiving high NAF, indicating that investors perceive such higher fees as a threat to independence. Similar evidence is provided by Gul et al. (2006) in the Australian market. Two other studies by Brandon et al. (2004) and Dhaliwal et al. (2008) in the bond market setting report that there is a positive association between NAF and cost of raising bond capital.

A very limited number of studies investigate the effect of NAF on two other FRQ proxies, namely, earnings conservatism using Basu (1997)-type conservatism measure (two studies), and earnings restatement (four studies). Neither Ruddock et al. (2005) in Australia nor Zhang and Emanuel (2008) in New Zealand find any evidence that earnings conservatism is negatively associated with the level of NAF. The association between NAF and earnings restatement, too, has produced very mixed evidence. Agrawal and Chadha (2005) and Li and Lin (2005) report a positive, but statistically insignificant, association between NAF and the probability of earnings restatement. The positive coefficient is consistent with the argument that higher NAF increases the economic ties between the auditor and the clients, which in turn makes the auditors less willing to challenge questionable accounting practices. Some of these accounting practices may require restatements in the future. Kinney et al. (2004) and Bloomfield and Shackman (2008), on the other hand, show that earnings restatement is negatively associated with the level of NAF.

To ensure that these results are not driven by publication bias, I calculate the fail-safe number for these 38 studies and compare this with the critical fail-safe number. The calculated fail-safe number N_fs= 4,608 for the 38 studies included in Table 2 far exceeds the critical value of 200 calculated using equation (5). Publication bias, therefore, can be ruled out as a concern for the reported results. For individual categories too, this conclusion holds.

4.1: Test of Heterogeneity Among FRQ Proxies

An important consideration in conducting meta-analysis is to determine whether the various ES that are averaged into a mean value are estimating the same population ES. According to Lipsey and Wilson (2001, p. 115), ‘in a homogenous distribution an individual effect size differs from the population mean only by sampling error. A statistical test that rejects the null hypothesis of homogeneity indicates that the variability of the ES is larger than would be expected from sampling error and, therefore, each effect size does not estimate a common population mean.’ The homogeneity of variance test is based on the Q-statistic, which is distributed as a chi-square with k− 1 degrees of freedom where k is the number of ES. The formula to calculate Q is given below:

()

Where ES is the individual effect size for i = 1 to k, and is the weighted mean ES over the k ES, and w_i is the individual weight for effect size_i. The reported Q-statistic in Table 2 is 118.33, which is highly significant (the critical value is 73.4 at the p= 0.001 level, with 37 df and strongly rejects the homogeneity hypothesis, suggesting that that the FRQ proxies used in the extant literature do not measure the same underlying construct).

To further explore the source of such heterogeneity, I conduct an analysis of between-study variability as suggested by Pomeroy and Thornton (2008, p. 316):

• •

  Group the financial reporting proxies into common categories (e.g., earnings management studies);

• •

  Calculate an overall ES for each FRQ group following the ES calculation procedure outlined above;

• •

  Calculate a Q-statistic for each group and test the homogeneity of variances assumption for each group;

• •

  Sum the Q-values across all individual groups to obtain a within group Q and test for homogeneity;

• •

  Finally, compare the within-group Q with the between group Q and test whether the difference is statistically significant.

Panel B of Table 2 presents the between-study variability result for three groups, and allows for testing the hypotheses developed in section 2. The mean ES of the studies investigating the association between earnings management and NAF is 0.03, with an associated z-statistic of 4.90 which is significant at the 1% level and therefore supports H1. Looking at the confidence interval of the ES distribution, it is concluded that there is a 0.95 probability that the true association between earnings management and NAF is between 0.017 and 0.04. The auditors' going-concern opinion-based FRQ proxy provides an overall ES of 0.017, with a z-statistic of 2.02 which is statistically significant at better than the 5% level, and thus supports H2. The true association between NAF and this FRQ proxy is 0.005 to 0.034. The overall ES of the group of studies investigating the association of NAF and market pricing of earnings is 0.019 with a highly significant z-statistic of 4.00. The true association is between 0.0095 and 0.028. H5 is therefore strongly supported. Two studies investigate the bond market effect of higher NAF and reach the conclusion that bond investors also consider high NAF to be a threat to independence. However, they report coefficients which are opposite in sign. Dhaliwal et al. (2008) measure the costs of debt using yield on the first bond issue where the higher the yield the higher the cost of debt. On the other hand, Brandon et al. (2004) use Moody's bond rating as a proxy for cost of debt, where a higher score represents a stronger bond rating. What this suggests is not only heterogeneity among FRQ proxies but also intra-group-heterogeneity.

Panel B also reports the Q-statistics for each component of FRQ proxies. Test of homogeneity of variances for the five categories of FRQ reveals that earnings management, qualified audit opinion and capital market-based studies suffer from heterogeneity of variances (reported Q-stat is 20.16, 52.15, and 37.38, respectively, all significantly larger than the critical chi-square value at better than the 5% level). This implies that empirical research on the association between NAF and FRQ proxies gives rise to significant intra-group heterogeneity with respect to the actual findings.

4.3: Sensitivity Analysis

Multiple dependent variables  Some of the earnings management and NAF studies used both DACCR and meeting some earnings benchmarks (analyst forecasts, report small increase in earnings) as earnings management proxies. Reported results in Table 2 include only DACCR-based results. The standard solutions for overcoming multiple dependent variable problem may be summarized as follows (DeCoster, 2004):

• 1

  ES can be calculated for each construct and be entered into the same model. This is the easiest approach but violates the independence assumption.

• 2

  ES can be calculated for each construct and perform a separate analysis on each. For example in this meta-analysis we can calculate ES for DACCR and benchmark beating separately. This, however, is only feasible if the each response measure is used in a number of different studies.

• 3

  Mathematically combine the two ES into one. This is the most preferred method.

Rosenthal and Rubin (1986) present a method for computing more accurate combinations of ES. To combine several ES, the following formula can be used:

()

Where, ES_i is the ES for the i_th measure, ρ is the typical intercorrelation between the response measures, and m is the number of alternative constructs being combined. One problem with using Rosenthal and Rubin's (1986) equations is that many studies do not report intercorrelation between the alternative measures. This is particularly acute in the present case as none of the earnings management studies that uses benchmark beating as an alternative earnings management construct provides the required inter-correlation to conduct this analysis.

The next preferred course of action is (2) above which requires a separate estimation of the ESs associated with benchmark beating constructs. I included four studies that use meeting analyst forecasts as an earnings benchmark for this analysis. Research has found that investors pay a premium for firms that meet and/or beat analyst forecasts (Kasznik and McNichols, 2002; Mikhail et al., 2004; Brown and Caylor, 2005). Following this procedure I find the mean ES for this very small sub-group of studies to be 0.022 with a z-statistic of 2.56 which is statistically significant at better than the 5% level. This evidence suggests that managerial propensity to manage earnings is positively associated with the level of NAF when alternative earnings management proxy is used.⁹

Studies conducted using data from the U.S.A. versus other countries  Of the 38 studies selected for meta-analysis in the primary analysis (Table 2), 23 studies use data from the U.S. market while the remainder are from Australia, New Zealand, the United Kingdom, Canada and Bangladesh. Whether the results derived from the U.S. market is generalizable to other countries requires empirical testing because of country-specific differences in reporting incentives attributed to a number of factors including differences in legal system (Habib, 2007). It should be recognized, however, that all the countries studied share common law legal origins and therefore are expected to share a lot of similarities with respect to reporting environment. But a major difference in the reporting environment between the U.S.A. and the rest of the countries is that U.S.A. is characterized by a relatively more litigious environment compared to the rest of the sample countries which may constrain managerial propensity to compromise independence and compel managers to provide more credible financial information. Khurana and Raman (2004) report that Big 4 audit is associated with a lower cost of equity (a proxy for audit quality) in the U.S.A. but not in Canada, Australia or the U.K.

I compute mean overall ES for these two groups of countries using FEERATIO as the independence proxy variable and report the results in Table 4 Panel A. Results reveal that the mean ES is positive and statistically significant for both country groups implying that overall FRQ is negatively associated with the level of NAF. Individual FRQ category analysis shows that earnings management impairs independence in fact both inside and outside the U.S.A. Capital market-based studies, too, provide evidence that the level of NAF is perceived as a threat to independence, again both inside and outside the U.S.A. The association between NAF and the propensity to issue qualified opinions is found to be insignificant for both country categories. Taken together, the evidence indicates that capital market participants consider the provision of NAS as a threat against auditor independence and this may provide some justification for the regulatory ban on NAS.

Table 4. SENSITIVITY RESULTS

Panel A: The association between NAF and FRQ using study results from two broad country categories [FEERATIO] as the primary independent variable
	Number of studies		All categories	Earnings management	Audit opinion	Earnings conservatism	Earnings restatement	Capital market studies
U.S.A.	23	Mean ES	0.02	0.026	0.019	None	0.0006	0.017
		z-stat	5.33***	4.08***	1.10	–	0.02	3.54***
		Q-values	77.16***	10.90***	28.57***	–	5.87***	30.02***
Outside U.S.A.	15	Mean ES	0.027	0.05	0.011	0.011	–	0.04
		z-stat	3.69***	3.03***	1.32	0.58	–	2.68***
		Q-values	42.93***	7.25***	24.32***	1.97		9.06***

Panel B: The association between NAF and FRQ using study results from three journal categories
	Number of studies		All categories	Earnings management	Audit opinion	Earnings conservatism	Earnings restatement	Capital market studies
TOP 5	11	Mean ES	0.03	0.034	−0.016	–	–	–
		z-statistic	5.28***	5.17**	0.87	–	–	–
		Q-values	19.20***	8.52***	2.34	–	–	–
Others	20	Mean ES	0.01	0.03	−0.0074	–	0.0006	0.013
		z-statistic	2.46**	1.69*	−0.88	–	0.02	2.65***
		Q-values	58.30***	4.94	17.47***	–	5.87*	13.17***
AJPAT	7	Mean ES	0.023	–	0.06	–	–	0.07
		z-statistic	3.17***	–	4.15***	–	–	3.78***
		Q-values	94.80***	–	17.68***	–	–	0.62

• Notes: The top five accounting journals are The Accounting Review (TAR), Journal of Accounting Research (JAR), Journal of Accounting and Economics (JAE), Accounting, Organizations and Society (AOS) and Contemporary Accounting Research (CAR). None of the papers selected for the present meta-analysis appeared in the JAE and AOS.
• ES is calculated using the following formula:
• ()
• The formula is straightforward to apply when the sample studies reported t-statistics. The majority of the studies canvassed for meta-analysis, however, reported p-values. Reported p-values are converted into t-statistics using p-value tables.
• Standard error is calculated using the following formula:
• 
• 
• Lower and upper bound of the ES are calculated as ES− (1.96*SE) and ES+ (1.96*SE) respectively.
• To test whether the various ESs that are averaged into a mean value are estimating the same population ES, a Q-statistic, which is distributed as a chi-square with k− 1 degrees of freedom where k is the number of ES, is calculated using the following formula (Lipsey and Wilson, 2001, p. 115).
• ()
• Where effect size_i (ES) is the individual ES for i = 1 to k, and is the weighted mean ES over the k ES, and w_i is the individual weight for effect size_i.
• ***,** and *** denote statistical significance at 1%, 5% and 10% level, respectively (two-tailed test).

Journal quality and the association between NAF and FRQ  Whether empirical evidence published in top quality journals differs from other journals is also a relevant issue. Hay et al. (2006, p. 157) argue that top quality journals ‘publish higher-quality studies with more robust findings; but there is also a greater possibility of bias, as it is possible that editors are more likely to reject findings that are not interesting because they are not significant’. Table 4 Panel B, therefore, presents a sensitivity analysis comparing studies published in the top five accounting versus other journals (accounting as well as other non-accounting academic journals) to determine whether the results of the research questions vary between these two groups of journals. Eleven of the 38 studies included in this meta-analysis appear in the top three accounting journals (none of the studies appears in Journal of Accounting and Economics and Accounting, Organizations and Society which are ranked among the top five accounting journals). The other 20 studies appear in other accounting and non-accounting journals. I have also grouped seven studies that appear in the specialized auditing journal Auditing: A Journal of Practice and Theory (AJPAT). Evidence reveals that for all these there sub-groups, it appears that increased NAF reduces FRQ, as is evident from a positive and statistically significant overall ES. For individual FRQ proxies, it appears that earnings management increases with the level of NAF in studies published in the top five journals (ES 0.026, z-statistic 4.53, significant at the 1% level). For other journal category, too, this findings hold although the ES is marginally significant. The association between NAF and audit opinion is only significant for studies published in the AJPAT (ES 0.06, z-statistic 4.15, significant at the 1% level). Also for this group of studies, there is strong evidence that ERC decreases with the increase in NAF (ES 0.07, z-statistic 3.78, statistically significant at the 1% level).

5: IMPLICATIONS

6: CONCLUDING REMARKS

It has been well documented that auditing serves as an effective corporate governance mechanism. However, audit quality has come under serious scrutiny because of the joint provision of audit and NAS, and regulators consider such services to be a threat to auditor independence. A sizable volume of archival research has developed over the years to address whether such a presumption is supported by actual data. The purpose of this paper is to conduct a systematic meta-analysis to provide a statistical assessment of the association between NAF and FRQ. Meta-analysis is superior to narrative literature review because it aggregates results across individual studies statistically and corrects for statistical artefacts like sampling and measurement error, thereby providing much greater precision with respect to the findings compared with narrative reviews. Meta-analysis results provide some evidence that investors perceive NAF as a threat to independence, as is evident from the negative association between the level of NAF and the ERC. However, the lack of homogeneity in the underlying studies remains to be a concern.