Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
USING PEER FIRMS TO EXAMINE WHETHER AUDITOR
INDUSTRY SPECIALIZATION IMPROVES AUDIT QUALITY
AND TO ENHANCE EXPECTATION MODELS FOR
ANALYTICAL AUDIT PROCEDURES
by
Miguel Angel Minutti Meza
A thesis submitted in conformity with the requirements
for the degree of Doctor of Philosophy
Joseph L. Rotman School of Management
University of Toronto
June, 2011
© Copyright by Miguel Angel Minutti Meza (2011)
ii
USING PEER FIRMS TO EXAMINE WHETHER AUDITOR INDUSTRY
SPECIALIZATION IMPROVES AUDIT QUALITY AND TO ENHANCE
EXPECTATION MODELS FOR ANALYTICAL AUDIT PROCEDURES
Miguel Angel Minutti Meza
A thesis submitted in conformity with the requirements
for the degree of Doctor of Philosophy
Joseph L. Rotman School of Management
University of Toronto
June, 2011
ABSTRACT
This dissertation investigates how economically-comparable peer firms can be used to
obtain inferences about a company’s accounting quality in two different research settings. The
first Chapter examines whether auditor industry specialization, measured using auditor market
share by industry, improves audit quality. After matching clients of specialist and non-specialist
auditors according to industry, size and performance, there are no significant differences in audit
quality between these two groups of auditors. In addition, this Chapter uses two analyses that do
not rely primarily on matched samples. First, examining a sample of Arthur Andersen clients that
switched auditors in 2002, there is no evidence of industry-specialization effects following the
auditor change. Second, using a simulation approach, this study shows that client characteristics,
and particularly client size, influence the observed association between auditor industry
specialization and audit quality. Overall, these findings do not imply that industry knowledge is
not important for auditors, but that the methodology used in extant studies examining this issue
may not fully parse out the effects of auditor industry expertise from client characteristics. The
second Chapter examines whether account-level expectation models for analytical audit
iii
procedures can be enhanced by using information from economically-comparable peer firms.
This Chapter assesses the effectiveness of three main types of expectation models, with and
without including information from peer firms: heuristic, time-series, and industry cross-
sectional models. Information from peer firms improves the accuracy of all models and improves
the detection power of time-series and industry cross-sectional models. Comparing between
models, one-period heuristic models are generally unreliable, and industry cross-sectional
models can be more effective than time-series models. These findings may help auditors of
public companies and financial analysts in selecting expectation models and finding peer firms to
assess the reasonability of a company’s financial information at the account-level.
iv
ACKNOWLEDGMENTS
I thank my supervisors, Gordon Richardson and Ping Zhang, for their continuous guidance
during my doctoral studies. I am also grateful to the other members of my dissertation
committee, Jeffrey Callen and Gus De Franco, and to Michel Magnan (the external reviewer) for
providing excellent comments and suggestions.
I thank Francesco Bova, Ole-Kristian Hope, Franco Wong, and Baohua Xin for giving me
helpful insights to improve this dissertation.
I acknowledge the financial support of the Canadian Public Accountability Board (CPAB)
through the 2010 Keith Boocock Doctoral Scholarship.
I am indebted to my colleagues that made available their help and advice in a number of ways
during the past five years, in particular to Stephanie Larocque, Alastair Lawrence, Hamed
Mahmudi and Dushyant Vyas. Outside school, some exceptional friends were always there to
give me encouragement. Among several others, I thank Flor Yunuen Garcia, Daniel Molina,
Carlos Sergio and Andrew Snowball.
Finally, I owe my deepest gratitude to my wife Nicole, my parents Miguel Angel and Carmen,
and my brother Luis Arturo for being my constant inspiration and support. I dedicate this thesis
to them.
***
OM SRI GURUBHYO NAMAH HARI OM
v
TABLE OF CONTENTS
ABSTRACT ii
ACKNOWLEDGEMENTS iv
TABLE OF CONTENTS v
LIST OF TABLES viii
LIST OF FIGURES x
LIST OF APPENDICES xi
CHAPTER 1 –DOES AUDITOR INDUSTRY SPECIALIZATION IMPROVE AUDIT
QUALITY? EVIDENCE FROM COMPARABLE CLIENTS
1
1.1 INTRODUCTION 1
1.2 AUDIT QUALITY AND ECONOMIC COMPARABILITY 6
1.2.1 Peer matching and economic comparability 6
1.2.2 Peer-matched test of audit quality 6
1.2.3 Matched-sample estimators of the specialization effects 7
1.2.4 Advantages and disadvantages of using matched samples 8
1.2.5 Selection of matching variables and matching approach 9
1.3 RELATED EMPIRICAL STUDIES AND MEASURES OF AUDITOR INDUSTRY
SPECIALIZATION
11
1.4 AUDIT-QUALITY PROXIES, REGRESSION MODELS AND SAMPLE
SELECTION
14
1.4.1 Discretionary accruals 14
1.4.2 Discretionary revenue 15
1.4.3 Going-concern opinions 16
1.4.4 Multivariate regression models 16
1.4.5 Analysis of matched samples –pooled and pair-wise differences models 18
vi
1.4.6 Sample selection 19
1.5 RESULTS 20
1.5.1 Discretionary accruals –full sample analyses 20
1.5.2 Discretionary accruals –matched sample analyses 22
1.5.3 Discretionary revenue –full sample analyses 24
1.5.4 Discretionary revenue –matched sample analyses 25
1.5.5 Propensity to issue a going-concern opinion –full sample analyses 26
1.5.6 Propensity to issue a going-concern opinion –matched sample analyses 27
1.6 ANALYSES OF AUDITOR SWITCHES 28
1.7 SIMULATION ANALYSES 33
1.8 ADDITIONAL MATCHING ANALYSES AND SENSITIVITY TESTS 34
1.8.1 Propensity-score matching 34
1.8.2 Size and industry matching 36
1.8.3 Bootstrap, random sub-samples, and stratified samples 36
1.8.4 Alternative measures of auditor industry specialization 37
1.9 CONCLUSION –CHAPTER 1 38
CHAPTER 2 –DOES USING INFORMATION FROM PEER FIRMS IMPROVE
ACCOUNT-LEVEL EXPECTATION MODELS?
41
2.1 INTRODUCTION 41
2.2 ACCOUNT-LEVEL EXPECTATION MODELS 46
2.2.1 Firm-level heuristic models 46
2.2.2 Firm-level time-series models 50
2.2.3 Industry cross-sectional models 51
2.3 PEER INFORMATION IN ACCOUNT-LEVEL EXPECTATION MODELS 53
2.3.1 Advantages of using peer-firm information in AAP 53
2.3.2 Selection of matching variables and matching approach 54
vii
2.3.3 Firm-level heuristic models including peer-firm information 55
2.3.4 Firm-level time-series models including peer-firm information 55
2.3.5 Industry cross-sectional models including peer-firm information 56
2.4 ASSESSING THE SPECIFICATION AND POWER OF EACH MODEL 56
2.4.1 firm-level error measures 56
2.4.2 Tests of statistical specification and power using a simulation procedure 58
2.4.3 Sample selection 59
2.5 RESULTS 60
2.5.1 Descriptive statistics for firm-level heuristic models samples 60
2.5.2 Specification of firm-level heuristic models 61
2.5.3 Detection power of firm-level heuristic models 63
2.5.4 Specification and detection power of firm-level time-series models 65
2.5.5 Specification and detection power of industry cross-sectional models 67
2.5.6 Idiosyncratic error in AAP expectation models 69
2.6 ADDITIONAL ANALYSES 70
2.6.1 Increasing the number of peer firms 70
2.6.2 Using industry averages instead of matched peer firms 71
2.6.3 Changing the structure of peer-firm information in the expectation models 71
2.6.4 Simulating manipulation in opposite direction and varying the amount of
manipulation
72
2.7 CONCLUSION –CHAPTER 2 72
REFERENCES 75
viii
LIST OF TABLES
Table 1: Discretionary Accruals Analyses
PANEL A: Descriptive Statistics Ranking Auditors by Industry Market
Share at National Level
PANEL B: Full Sample Partition by Industry Specialization
at National Level (NLEAD1)
PANEL C: Full Sample Partition by Industry Specialization
at City Level (CLEAD1)
88
89
90
Table 2: Discretionary Accruals Analyses: Pooled Multivariate Tests
Full and Matched Samples
92
Table 3: Discretionary Accruals Analyses: Multivariate Pair-Wise Differences Tests
Matched Samples
93
Table 4: Discretionary Revenue Analyses
PANEL A: Full Sample Partition by Industry Specialization
at National Level (NLEAD1)
PANEL B: Full Sample Partition by Industry Specialization
at City Level (CLEAD1)
94
95
Table 5: Discretionary Revenue Analyses: Pooled Multivariate Tests
Full and Matched Samples
97
Table 6: Discretionary Revenue Analyses: Multivariate Pair-Wise Differences Tests
Matched Samples
98
Table 7: Going-Concern Analyses
PANEL A: Full Sample Partition by Industry Specialization
at National Level (NLEAD1)
PANEL B: Full Sample Partition by Industry Specialization
at City Level (CLEAD1)
99
100
Table 8: Going-Concern Analyses: Pooled Multivariate Tests
Full and Matched Samples
102
ix
Table 9: Going-Concern Analyses: Conditional Logistic Regression Tests
Matched Samples
103
Table 10: Clients that Switched from Arthur Andersen 2001–2002
PANEL A: Pre-Post Switch Analyses for Arthur-Andersen Clients
PANEL B: Pre-Post Switch Analyses for Arthur-Andersen Clients
and Matched Control Group
104
105
Table 11: Heuristic Expectation Models
Descriptive Statistics
PANEL A: Full Sample
PANEL B: Sub-Sample with Peer Firms
106
Table 12: Heuristic Expectation Models
Correlation Table –Sub-Sample with Peer Firms
107
Table 13: Heuristic Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
108
Table 14: Time Series Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
109
Table 15: Industry Cross-Sectional Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
110
x
LIST OF FIGURES
Figure 1: LEAD1 Simulation Results from 1,000 replications
Assigning Clients to Five Auditors at Random
Full Sample
87
Figure 2 LEAD1 Simulation Results from 1,000 replications
Assigning Clients to Five Auditors at Random
Sub-Sample of Industries Without a Specialist Auditor
87
xi
LIST OF APPENDICES
Appendix A: Variable Definitions –Chapter 1 80
Appendix B: Summary of Discretionary Accruals and Discretionary Revenue Estimates 81
Appendix C: Issues Identified by PCAOB during the Inspections of Big 4 Firms from
2003 to 2008 Involving the Application of Analytical Procedures
82
Appendix D: Variable Definitions –Chapter 2 84
Appendix E: Summary of Expectation Models 85
Appendix F: Summary of Simulation Procedure 86
1
CHAPTER 1 –DOES AUDITOR INDUSTRY SPECIALIZATION IMPROVE AUDIT
QUALITY? EVIDENCE FROM COMPARABLE CLIENTS
1.1 INTRODUCTION
Accounting firms recognize the importance of industry expertise in providing high-quality
audits and organize their assurance practices along industry lines. In large firms, individual
auditors specialize by auditing clients in the same industry. For example, PwC highlights that ―our
audit approach, at the leading edge of best practice, is tailored to suit the size and nature of your
organisation and draws upon our extensive industry knowledge (PwC 2010).‖ A report on the U.S.
audit market issued by the U.S. General Accounting Office (GAO) in 2008 also acknowledges the
importance of industry expertise, noting that ―a firm with industry expertise may exploit its
specialization by developing and marketing audit-related services which are specific to clients in
the industry and provide a higher level of assurance (GAO 2008; p. 111).‖ Asserting the benefits of
auditor industry specialization is relevant for public companies choosing among auditors, to
regulators concerned with high concentration on the U.S. audit market, and to audit firms aiming
to perform high-quality audits while maintaining their competitive position in each industry.1
Auditing researchers have extensively studied the consequences of auditor expertise.
Experimental auditing research provides evidence that industry expertise enhances auditors’
judgements. The findings of prior studies suggest that knowledge of the industry may increase
audit quality, improving the accuracy of error detection (Owhoso et al. 2002; Solomon et al. 1999),
1 Since four audit firms hold the majority of the U.S. audit market for public companies, specialization may lead to
dominance of a single audit firm within an industry. Dominance by a single audit firm in an industry may have
undesirable consequences such as high audit fees and low audit quality. Extant research shows that auditors may be
able to obtain a specialization fee premium by improving efficiency and creating barriers to entry. Francis et al.,
(2005) find an association between fee premiums and joint national and city specialist auditors in the U.S. audit
market; DeFond et al. (2000) find a specialization premium in addition to audit quality effects in the Hong Kong audit
market; however, Carson and Fargher (2007), focusing on the Australian audit market, find that the association
between the specialist fee premium and auditor specialization is concentrated in audit fees paid by the largest clients in
each industry.
2
enhancing the quality of the auditor’s risk assessment (Low 2004; Taylor 2000), and influencing
the choice of audit tests and the allocation of audit hours (Low 2004). Empirical auditing research
has also examined the effects of auditor industry expertise; however, empirical researchers cannot
directly observe expertise at the firm, office, or auditor level, and this area of the literature has
used each audit firm’s within-industry market share, or auditor industry specialization, as an
indirect proxy for auditor industry expertise. A specialist is a firm that has ―differentiated itself
from its competitors in terms of market share within a particular industry‖ (Neal and Riley 2004;
p. 170). Previous studies that use within-industry market share proxies for industry expertise have
shown that the clients of specialist auditors have better financial reporting quality, exhibiting on
average from 0.3 to 2.0 percent lower absolute discretionary accruals, compared to clients of non-
specialist auditors (Balsam et al. 2003; Krishnan 2003; Reichelt and Wang 2010).
Measuring the effects of auditor industry expertise on audit quality is problematic because
the proxies for industry specialization and audit quality are associated with underlying client
characteristics. For example, large clients have lower absolute discretionary accruals and large
clients are often audited by industry specialists. For determining causal inference in observational
studies, empirical researchers should aim to compare treated and control groups that have similar
client characteristics, ideally approximating experimental conditions. A potential way to achieve
this objective is by matching treatment and control observations on all relevant observable
dimensions except for the treatment and outcome variables. This study proposes a methodology to
find economically-comparable clients and applies it to mitigate the effect differences in client
characteristics between specialists and non-specialist auditors.
Controlling for confounding factors is particularly important in studying the effects on
industry specialization for two main reasons. First, an audit firm may have extensive industry
3
knowledge even when its within-industry market share is small relative to other audit firms.
Industry knowledge could be gained through other means; for instance, by the number of years an
audit team has audited clients in the industry, by providing training to individual auditors, by
auditing private clients in the same industry, by providing consulting services, and by hiring
experts from within the industry or from other audit firms.2 Thus, it is not obvious that auditors
with larger market share will have higher quality. Second, the evidence in Boone et al. (2010) and
Lawrence et al. (2011) shows that the previously documented association between auditor size and
audit quality could be attributed to differences in client characteristics, particularly to differences
in client size. The separation of specialist and non-specialist auditors by within-industry market
share also creates two groups of auditors with different client characteristics. For example,
specialist auditors have larger and more profitable clients compared to non-specialist auditors.
Prior studies of auditor industry specialization control for the impact of client
characteristics by including client size, performance, growth, and other linear control variables in
multivariate regression analyses. There are two problems with the linear control approach:
important variables such as client size and performance are nonlinear to both the auditor choice
decision and the proxies for audit quality (Kothari et al. 2005; Hribar et al. 2009; Lawrence et al.
2011), and differences in client characteristics are partially a result of endogenous self-selection.
Furthermore, previous research by Rubin (1979), Heckman et al. (1998), Rubin and Thomas
(2000), and Rubin (2001), shows that linear regression may increase bias in the estimation of
treatment effects when there are even moderately nonlinear relations between the dependent and
independent variables, and this problem is exacerbated when there are significant differences in
2 For example, a recent article in Bloomberg’s BusinessWeek notes that ―Deloitte recruiters say they're doing better
head-to-head against such old-shoe firms as McKinsey and BCG Consulting, both in recruiting and getting new
business‖ and that this firm ―typically gets more than 85 percent of the experienced hires it makes an offer to‖ (Byrnes
2010).
4
means and variances in the independent variables between treated and control groups. To
overcome the endogeneity problem, some studies use econometric designs that explicitly model
the mechanism that results on differences in client characteristics between auditors, such as the
Heckman (1979) self-selection model or two-stage models. A limitation of these research designs
is that they require identifying appropriate exogenous instrumental variables or exclusion
restrictions in the first stage, which is a difficult condition to meet in models predicting auditor
choice (Francis et al. 2010). Moreover, two-stage models may perform poorly when there is
insufficient overlap between treatment and control observations (Glazerman et al. 2003; Dehejia
and Wahba 2002). The matching models used in this study constitute an alternative to determine
the auditor treatment effects. 3
This study uses three audit-quality proxies: discretionary accruals, a revenue manipulation
proxy from Stubben (2010), and the auditor’s propensity to issue a going-concern opinion.
Consistent with prior studies, this study first shows a relation between three audit-quality proxies
and auditor industry specialization at the U.S. national and city level. After matching clients of
specialist and non-specialist auditors, there are no statistically significant differences in audit
quality-proxies between the two groups of auditors. The main matching approach used in this
study is based on three fundamental economic dimensions: industry, size, and performance. These
findings are robust to using alternative matching approaches, to using various proxies for auditor
industry specialization and audit quality, and to controlling for the effect of imperfectly matched
characteristics.
This study also documents confirmatory evidence from two additional analyses. First, there
3 Heckman (2005) discusses extensively the advantages and disadvantages of matching versus explicit modelling of
the selection process. Both approaches are acceptable for estimating treatment effects; however, the matching
approach does not require identification of exclusion restrictions. Conversely, matching relies on the assumption that
selection is strictly based on observables or that treatment assignment is ―strongly ignorable,‖ and also requires some
degree of overlap or ―common support‖ between treatment and control observations.
5
are insignificant pre-post differences in discretionary accruals or revenue manipulation for Arthur
Andersen’s clients that exogenously switched to auditors with a different degree of specialization
in 2002. Second, using a simulation procedure, clients are assigned to five simulated auditors at
random and specialist and non-specialist auditors are designated based on within-industry market
share. The auditor that is assigned the largest clients of the industry is often designated as
specialist, and designated specialist auditors appear to have higher audit quality compared to non-
designated specialist auditors, highlighting the confounding effect of client size on tests of auditor
industry specialization.
Overall, the combined evidence provided in this Chapter suggests that the extant empirical
methodology may not fully parse out the confounding effects of client characteristics in tests of
auditor industry specialization and audit quality. The findings documented in this study do not
imply that industry knowledge is not important for auditors. Furthermore, the results of this study
are subject to the intrinsic limitations of matching for estimating causal effects, resulting from a
trade-off between internal and external validity, and to the proxies for audit quality and auditor
industry specialization used in this study. Finally, beyond the audit literature, this study contributes
to the broad accounting literature on matching and economic comparability. The methodology
used here could be adapted to other studies in accounting research comparing treated and control
groups, particularly where it is difficult to specify a correct model or to find exogenous predictors
of treatment choice.4
4 For example, a study using discretionary accruals as a dependent variable and a treatment variable correlated with
firm size and performance (e.g., management compensation, corporate governance, or financial analyst following)
may benefit from using the methodology applied in this study.
6
1.2 AUDIT QUALITY AND ECONOMIC COMPARABILITY
1.2.1 Peer-Matching and economic comparability
Using peer firms as a benchmark is common among practitioners and researchers. Peer
firms are used by financial analysts to support their price-earnings ratios, earnings forecasts, and
overall stock recommendations (De Franco et al. 2011), by investment managers in structuring
their portfolios (Chan et al. 2007), by compensation committees in setting executive compensation
(Albuquerque 2009), by business valuators in determining valuation multiples (Bhojraj et al.
2002), and by auditors in conducting analytical procedures (Hoitash et al. 2006). In using peer
firms as benchmarks, practitioners rely on comparability or uniformity of financial information
and on the overall quality of the mapping of economic events into financial reporting. Moreover,
several prior studies in accounting research have used peer matching ―as a research design device
for isolating a variable of particular interest‖ (Bhojraj et al. 2002; p. 410), to simplify data
collection (Geiger and Rama 2003), to provide more reliable inferences in market-based research
(Barber and Lyon 1997), and to mitigate the effect of nonlinearities (Kothari et al. 2005).5 A
primary objective of this study is to use fundamental economic characteristics to match peer firms
in order to obtain inferences about the relative accounting quality between two groups of auditors.
1.2.2 Peer-matched test of audit quality
To investigate the difference in audit quality between two auditors, researchers must
ascertain that the observed differences between the auditors’ clients are the result of the auditors’
effect. A peer-based approach could be useful in identifying the auditor treatment effects under two
general scenarios.
5 Other disciplines have investigated the benefits and drawbacks of matching to identify causal effects; for example,
applied statistics (Stuart 2009; Rubin 2006; Rosenbaum 2002), epidemiology (Brookhart et al. 2006), sociology
(Morgan and Harding 2006), applied econometrics (Imbens 2004), and political science (Ho et al. 2007). Zhao (2004,
p.100) notes that ―Selection bias due only to observables is a strong assumption. But with a proper data set and if the
selection-on-observables assumption is justifiable, matching methods are useful tools to estimate treatment effects.‖
7
In the first scenario, assume that (1) clients do not engage routinely in earnings
management, (2) low-quality auditors allow random noise in accounting accruals as a result of
inconsistent enforcement of accounting principles, and (3) two clients are economically
comparable and have the same drivers of accounting accruals, but one client has a low-quality
auditor and the other client has a high-quality auditor. Under perfect economic comparability, the
only difference between these two clients’ accruals is the random noise introduced by the low-
quality auditor.
In the second scenario, assume that (1) clients engage routinely in earnings management,
(2) low-quality auditors are not able to fully uncover earnings management, and (3) two clients are
economically comparable and have the same drivers of accounting accruals, but one client has a
low-quality auditor and the other client has a high-quality auditor. Under these conditions, the
effect of earnings management should be the only difference between these two clients’ accruals.
Along these lines, researchers may identify differences between the accruals of clients of
specialist and non-specialist auditors if specialist auditors are better at enforcing the right
accounting policies and at constraining earnings management. In a general setting where the true
accrual function is unknown, the overall difference in accrual quality between two clients can be
approximated by employing a combination of a discretionary accruals model and matching on
economic comparability. Similarly, a test of the differences in propensity to issue a going-concern
opinion between specialist and non-specialist auditors could be well specified if the matching
process mitigates differences in client characteristics that could influence the probability of
bankruptcy.
1.2.3 Matched-sample estimators of the specialization effects
A univariate t-test of the differences in means between perfectly matched clients constitutes
8
a direct estimator of treatment effects (Zhao 2004). However, if the matching process is not
perfect, it is still important to control for unmatched client characteristics using multivariate
analyses. The multivariate analyses in all matched samples of specialist and non-specialist
auditors’ clients are performed using two approaches. Under the first approach, the same model
estimated on the full sample is estimated in the pooled matched sample of clients, while under the
second approach, the pair-wise differences in the dependent variables between peer-matched
clients of specialists are regressed on the pair-wise differences of the independent variables in the
original model (Rubin 1973; Imbens 2004; Cram et al. 2009). The intercept of this pair-wise
differences model is interpreted as the average difference resulting from the specialist’s treatment
effects. Following Cram et al. (2009), the analysis of the matched pairs of clients of specialist and
non-specialist auditors is done using a conditional fixed effect logistic regression.
1.2.4 Advantages and disadvantages of using matched samples
An advantage of the peer-matched approach is that it imposes weak stationarity or linearity
conditions on the relation between the matched firm characteristics and the proxies for audit
quality. Although the peer-based approach reflects the relative quality between peer firms,
idiosyncratic differences should be mitigated in large samples, allowing researchers to assess the
average treatment effects of specialist auditors. This argument is similar to that in Kothari et al.
(2005); however, this approach aims to isolate a wider set of client characteristics from the proxies
for audit quality. Another advantage of the peer-based approach is that it does not require
identification of exclusion restrictions. Finally, this approach is suitable for a differences-in-
differences test of the effect of auditor specialization for clients that switch auditors as a result of
an exogenous shock.
Using matched samples comes at a cost, resulting from a trade-off between internal and
9
external validity. Four underlying threats to matching approaches are (1) firms deemed to be
economically similar may not be truly comparable, (2) the results from matched samples may not
be immediately extended to the entire population, (3) matching reduces sample sizes, and (4) it is
not possible to match on pre-treatment attributes or to control for alternative treatments.6 Section
1.8.3 describes the results of additional analyses performed to mitigate these threats.
1.2.5 Selection of matching variables and matching approach
There are two primary research-design choices applicable to matched samples. The first
choice is the set of variables or dimensions used for matching; the second is the mechanism to
aggregate across dimensions and to find comparable observations. The choice of matching
variables is important because in a strict sense, matching assumes that bias is only due to
observables. The source of bias is the difference between observables in the treatment and control
groups. The bias due to non-matched characteristics decreases as the number of matching variables
increases. On the other hand, the complexity and structure of the methods needed to aggregate
across dimensions increases as the number of matching variables increases.
When the number of matching variables is small, the researcher can directly match on the
variables of interest or within a specified distance from each variable of interest without requiring
a weighting approach to aggregate across dimensions. This type of matching is known as
attributes-based or covariate matching. The main approach used in this study is a form of covariate
matching. This study proposes that the three most important fundamental variables that affect the
audit-quality proxies and also influence the differences between auditor groups are the client’s
industry, size, and performance. The literature on discretionary accruals has repeatedly highlighted
6 The fourth threat could result in a bias if the matching variables are affected by the auditor choice. It is not possible
to fully rule out concerns that inferences based on matched samples are affected by ex post matching or alternative
treatments.
10
the importance of these three dimensions and recommends estimating discretionary accruals by
industry, scaling by total assets and controlling for firm performance.
To match on these dimensions, for a given fiscal year-end, industry (defined by two-digit
SIC code), and size distance (firms that are within a size distance of 50 percent), firm i is matched
to firm j with the most comparable performance, measuring performance as stock returns’
covariance over the preceding 48 months, where higher covariance indicates higher
comparability.7 As per the De Franco et al. (2011) methodology, returns covariance is measured
using the adjusted R2 of the following regression of firm i’s monthly returns on firm j’s monthly
returns8:
RETURNSi,t = Φi,t + Φi,tRETURNSj,t + εi,t (1)
In addition, matched firms are required to have their fiscal year-end on the same month to
reduce differences from timing in financial reporting. Allowing for 50 percent distance in total
assets results in more than one potential control for every treatment observation, and the final
selection among all possible controls is based on returns’ covariance. This procedure is likely to
closely match peer firms deemed economically comparable by the market. Compared to other
matching approaches, it does not rely on a specific functional form to predict comparability,
beyond a returns covariance structure, and can be used not only in case-control research settings,
but also in situations where a company needs to be matched with its economic peers; for example,
to form benchmark groups for valuation or to perform analytical audit procedures.
In order to mitigate any bias resulting from imperfect matching, the pair-wise differences
7 As noted by Chan et al. (2007, p. 57), ―if equity market participants consider a set of companies closely related, then
shocks in the group of stocks should experience coincident movements in their stock returns.‖ 8 The Kendall’s (1938) Tau or rank correlation coefficient is also estimated for the matched peer firms. This non-
parametric statistic measures co-movement or serial dependence and can be directly interpreted as the probability of
observing concordant or discordant pairs of observations. Both correlation measures produce similar matched pairs.
11
analyses controls for differences in size, performance and other variables between matched
observations. Furthermore, as robustness test, this study also uses propensity-score matching,
including several additional variables in the matching. Using propensity score, control
observations are matched to treatment observations based on a specified distance between their
overall probabilities of undergoing treatment. These probabilities are estimated using a number of
covariates that predict choice, effectively aggregating multiple dimensions into a single matching
variable. This alternative matching requires specifying a functional form for the choice model and
an acceptable distance between observations in terms of probability. Matching on the three
proposed covariates or using propensity-score matching produces qualitatively similar results.
These two approaches are complementary in examining the specialization effects, and confirm that
the previously documented specialization effects may be attributable to differences in client
characteristics. The propensity-score matching results are described in Section 1.8.1.9
1.3 RELATED EMPIRICAL STUDIES AND MEASURES OF AUDITOR INDUSTRY
SPECIALIZATION
Prior studies primarily measure auditor industry specialization using the auditor’s within-industry
market share. For each auditor and year, industry market share is calculated as:
I
i
J
j kij
J
j kij
ki
S
SEMARKETSHAR
ik
1 1
1 (2)
where MARKETSHAREki is the market share of auditor i in industry k, Skij represents the total
assets of client firm j in industry k audited by auditor i, J represents the number of clients that are
9 Zhao (2004) concludes that there is no clear alternative between covariates and propensity-score matching methods.
When the correlation between covariates and treatment choice are high, propensity-score matching is a good choice;
however, when the sample size is small, covariate matching performs better. Hahn (1998) shows that covariate
matching is asymptotically efficient because it attains the efficiency bound, and Angrist and Hahn (2004) show that
covariate matching may be more efficient in finite samples than propensity-score matching.
12
served by audit firm i in industry k, and I is the number of audit firms in industry k. 10
The two
main proxies for auditor industry specialization used in this study are:
NLEAD1 = ―1‖ for auditors that have the largest market share in a given industry
and year at the U.S. national level and have more than 10 percent
greater market share than the closest competitor, and ―0‖ otherwise;
and,
CLEAD1 = ―1‖ for auditors that have the largest market share in a given industry
and year at the U.S. city level, where city is defined as a Metropolitan
Statistical Area following the 2003 U.S. Census Bureau MSA
definitions, and have more than 10 percent greater market share than
the closest competitor, and ―0‖ otherwise. 11
The main analyses presented in all tables use these two proxies, and similar results using an
alternative cut-off for market share, and combined national and city level specialization proxies,
are described in Section 1.8.4.
Balsam et al. (2003) find a negative relation between auditor specialization and the client’s
absolute discretionary accruals. Discretionary accruals are calculated using the industry cross-
sectional Jones (1991) model and auditor industry specialization is proxied by six measures of
auditor market share by industry. Similarly, Krishnan (2003) documents a negative relation
between auditor specialization and the client’s absolute discretionary accruals. Discretionary
accruals are calculated using the industry cross-sectional Jones (1991) model and auditor industry
specialization is proxied by two measures of auditor market share by industry.
10
Prior studies have also used total sales or auditor fees to compute within-industry market shares. This study uses
total assets to calculate specialization measures at national and city level because total assets are available for most
firms in the sample period. 11
Francis et al. (2005) and Reichelt and Wang (2010) also use MSA definitions to identify city level specialists. Cities
with less than three observations are deleted from the sample. MSA definitions are available at the U.S. Census
Bureau’s website: http://www.census.gov/population/www/metroareas/metrodef.html.
13
Reichelt and Wang (2010) also document a negative relation between auditor specialization
and the client’s absolute discretionary accruals. Discretionary accruals are calculated using the
industry cross-sectional Jones (1991) model, including return on assets (ROA) as per Kothari et al.
(2005), and auditor industry specialization is measured using two proxies at the national level, city
level, and both levels combined: the first proxy equals one for auditors that have the largest market
share in a given industry and have more than 10 percent greater market share than the closest
competitor, and zero otherwise; and the second proxy that equals one for auditors that have over
30 percent market share in a given industry and year, and zero otherwise. In addition, Reichelt and
Wang (2010) shows a positive association between the city level and combined national and city
level measures and the auditor’s propensity to issue a going-concern opinion.
Lim and Tan (2008) investigate the moderating effect of auditor specialization on the
relation between non-audit fees and absolute discretionary accruals and find no statistically
significant association between absolute discretionary accruals and auditor specialization in a
model without non-audit-fee measures. At the same time, they find an interaction effect between
auditor specialization and non-audit fees, suggesting that clients audited by specialists are
associated with higher absolute levels of discretionary current accruals as non-audit fees increase.
Lim and Tan (2008) calculate discretionary accruals using the industry cross-sectional Jones
(1991) model, including ROA as per Kothari et al. (2005), and their measure of specialization
equals one if the auditor has the largest market share in the client’s industry, and zero otherwise.
Furthermore, Lim and Tan (2008) document a positive association between auditor industry
specialization and the auditor’s propensity to issue a going-concern opinion; however, the
association is negative once the specialization proxy is interacted with audit fees.
The studies summarized above show a positive association between auditor industry
14
specialization and audit quality. Consistent with prior studies, this study uses the client’s absolute
discretionary accruals and the auditor’s propensity to issue a going-concern opinion as audit-
quality proxies. Additionally, this study uses a proxy for discretionary revenue, proposed by
Stubben (2010), which considers a number of cross-sectional characteristics in the estimation
process. This measure is arguably better specified at detecting revenue manipulation than the
previously used discretionary accruals measures.12
1.4 AUDIT-QUALITY PROXIES, REGRESSION MODELS AND SAMPLE SELECTION
1.4.1 Discretionary accruals
As a first audit-quality proxy, this study uses absolute discretionary accruals, estimated
using an annual cross-sectional model for each industry. Absolute discretionary accruals are
calculated using two different approaches: ADA is based on a model including ROA (Kothari et al.
2005) as an additional predictor (Equation (3) below), and ADA_FULL is based on a more
comprehensive model (Equation (4) below) including ROA (Kothari et al. 2005), cash flows in
periods t and t-1 scaled by total assets (McNichols 2002), and a non-linear interaction term based
on the sign of cash flows in period t (Ball and Shivakumar 2006). The main analyses present the
results based on the absolute value of discretionary accruals from the Kothari et al. (2005) model
(ADA). All results are qualitatively similar using ADA_FULL, or estimating the Kothari et al.
(2005) model using prior year’s ROA instead of current year’s ROA.13
ACi,t = α + β1ΔRi,t + β2PPEi,t + β3ROAi,t + εi,t (3)
ACi,t = α + β1ΔRi,t + β2PPEi,t + β3ROAi,t + β4CFOi,t-1 + β5CFOi,t (4)
12
Table 3 in Stubben (2010, p.707) shows that this discretionary measure detects a combination of 1 percent simulated
manipulation in both revenue and expenses in 23.6 percent of the samples with manipulation, compared to 11.6
percent using the Jones model or 11.2 percent using the performance-matched modified Jones model. 13
All continuous variables are winsorized at the 1 and 99 percent levels before estimating the discretionary accruals
and discretionary revenue models.
15
+ β6CFOi,t+1 + β7Di,t + β8D×CFOi,t + εjt
where for client i and fiscal year-end t:
ADA = absolute value of error term εi,t in Equation (3) ;
ADA_FULL = absolute value of error term εi,t in Equation (4);
AC = (cash flow from operations - income before extraordinary
items)/average total assets;
ΔR = (revenuet - revenuet-1)/average total assets;
PPE = gross property, plant and equipment/average total assets;
ROA = (net income before extraordinary items)/average total assets;
CFO = (cash flow from operations)/average total assets; and,
D = indicator variable equal to ―1‖ if CFOt is negative, and ―0‖ otherwise.
1.4.2 Discretionary revenue
As a second audit-quality proxy, this study uses ADREV, the absolute value of discretionary
revenue, as proposed by Stubben (2010). The revenue manipulation or discretionary revenue
model (Equation (5) below) is related to the discretionary accruals model, relying on the
association between changes in accounts receivable and changes in revenue to predict earnings
management. Moreover, the estimation of this measure allows for variation in the model
coefficients across client characteristics and also considers nonlinear terms, compared to
discretionary accruals models that assume the same coefficient for all clients in the same industry.
ΔARi,t = α + β1ΔRi,t+ β2ΔRi,t×SIZEi,t + β3ΔRi,t×AGEi,t + β4ΔRi,t×AGE_SQi,t
+ β5ΔRi,t×GRR_Pi,t + β6ΔRi,t×GRR_Ni,t + β7ΔRi,t×GRMi,t
+ β8ΔRi,t×GRM_SQi,t + εi,t (5)
16
where for client i and fiscal year-end t:
ADREV = absolute value of error term εi,t in Equation (5);
ΔAR = change in accounts receivable reported in the cash flow statement;
ΔR = (revenuet - revenuet-1)/average total assets;
SIZE = natural logarithm of total assets;
AGE = natural logarithm of the number of years since the firm has data in
COMPUSTAT;
GRR_P = industry-median-adjusted revenue growth (=0 if negative);
GRR_N = industry-median-adjusted revenue growth (=0 if positive);
GRM = industry-median-adjusted gross margin; and
_SQ = square of variable.
1.4.3 Going-concern opinions
As a third audit-quality proxy, this study uses the auditor’s propensity to issue a going-
concern opinion. The variable for going-concern opinion (GCONCERN) is directly taken from
Audit Analytics and is coded as ―1‖ if the auditor gave a going-concern opinion to a client in the
fiscal year, and ―0‖ otherwise.
1.4.4 Multivariate regression models
This study replicates the findings of prior studies that examine the relation between auditor
specialization and each audit-quality proxy using the following model14
:
14
All continuous variables are winsorized at the 1 and 99 percent levels before estimating the main models in the full
samples of each audit-quality proxy.
17
QUALITY_PROXYi.t= ω0 + ω1LEADi,t + ω2BIG4i,t + ω3LOGMKTi,t + ω4LEVi,t
+ ω5ROAi,t+ ω6ROALi,t+ ω7LOSSi,t+ ω8CFOi,t + ω9BTMi,t
+ ω10ABS(ACCRL) i,t + ω11GROWTHi,t+ω12ALTMANi,t
+ ω13STDEARNi,t + ω14TENUREi,t + ω15YEAR F.E. + vi,t (6)
where for client i and fiscal year-end t:
QUALITY_PROXY = audit quality proxies as defined above;
LEAD = indicator variable for each measure of auditor industry specialization
as defined above (NLEAD1 or CLEAD1);
BIG4 = ―1‖ if the client has a Big 4 auditor, and ―0‖ otherwise;
LOG_MKT = natural logarithm of market value;
LEV = (total liabilities) / average total assets;
ROA = (net income) / average total assets;
ROAL = (net incomet-1) / average total assets t-1;
LOSS = indicator variable equal one if net income is negative, and ―0‖
otherwise;
CFO = (cash flow from operations)/average total assets;
BTM = (book value of equity) / market value of equity;
ABS(ACCRL) = (absolute value of total accrualst-1)/average total assetst-1;
GROWTH = sales growth calculated as (sales – salest-1)/salest-1;
ALTMAN = Altman’s (1983) scores;
STDEARN = standard deviation of income before extraordinary items in the past
four years;
18
TENURE = ―1‖ if the client has kept the same auditor for three or more fiscal
years, and ―0‖ otherwise; and,
YEAR F.E. = year fixed effects.
Prior studies document that auditor industry specialization increases audit quality, reducing
the absolute value of discretionary accruals and discretionary revenue, and increasing the auditor’s
propensity to issue a going-concern opinion. Consistent with Balsam et al. (2003) and Reichelt and
Wang (2010), lower discretionary accruals are expected for larger firms (LOG_MKT), firms with
higher operating cash flow (CFO), firms with higher leverage (LEV), firms audited by a Big 4
auditor (BIG4), and firms with longer tenure (TENURE). Higher absolute discretionary accruals
are expected for growth firms (GROWTH and BTM), firms with losses (LOSS), firms with extreme
performance (ROA and ROAL), firms with high-income volatility (STDEARN), firms with high
probability of bankruptcy (ALTMAN), and for firms with higher prior total accruals
(ABS(ACCRL)). The signs of all controls variables should be the same in the discretionary revenue
model as both proxies are equally influenced by incentives and opportunities for earnings
management.
In the going-concern model, the probability of going concern should be lower for larger
and more stable clients (LOG_MKT, BIG4, TENURE), and decrease as liquidity (CFO, ALTMAN)
and profitability increases (ROAL, ROA). On the other hand, the probability of going concern will
increase as risk (STDREARN, ABS(ACCRL), LOSS) and leverage increases (LEV).15
1.4.5 Analysis of matched samples –pooled and pair-wise differences models
For each specialization measure, NLEAD1 and CLEAD1, clients of specialist and non-
specialist auditors are pair-matched by fiscal year-end month and industry, within a 50 percent size
15
Consistent with prior studies, the discretionary accruals and discretionary revenue models do not include industry
fixed effects because these audit-quality proxies are estimated by industry. The logistic going-concern model includes
both year and industry fixed effects.
19
distance, selecting the peer with the highest stock return covariance from all the possible matches.
Two alternative models are estimated using the matched samples. First, Equation (6) is estimated
in the pooled matched sample of clients of specialist and non-specialist auditors. Second, the
following pair-wise differences model is estimated for the discretionary accruals and discretionary
revenue audit-quality proxies:
QUALITY_PROXYijt= γ0 + γ1BIG4ijt + γ2LOGMKTijt + γ3LEVijt + γ4ROAijt
+ γ5ROALijt+ γ6LOSSijt+ γ7CFOijt + γ8BTMijt
+ γ9ABS(ACCRLij)t + γ10GROWTHijt+ γ11ALTMANijt
+ γ12STDEARNijt + γ13TENUREijt + γ14YEAR F.E. + ut (7)
where for fiscal year-end t, ij denotes the pair-wise difference between the value of each variable
(as previously defined) for the client of the specialist auditor minus the value of the same variable
for the matched client of a non-specialist auditor, and the intercept γ0 represents the average pair-
wise difference between matched observations, controlling for the effect of differences resulting
from imperfectly matched variables. For the going-concern audit-quality proxy, a conditional fixed
effects logistic regression is estimated using the observations with intra-pair variation in going-
concern opinions.16
1.4.6 Sample selection
For the discretionary accruals analyses, this study uses U.S. public company data for the
years 1988 to 2008 from COMPUSTAT and data for the years 2000 to 2008 from Audit
Analytics.17
Firms in the financial services industries (SIC codes 6000–6999), firms with negative
assets, market price, or sales, and firms without the necessary data to calculate the control
16
An alternative approach is to include an indicator variable for each matched pair. The results obtained by using this
alternative approach are similar to those documented in the main tables. 17
The main sample is resricted to this time period because reported operating cash flows, needed to calculate
discretionary accruals, are only available starting from 1988 as per SFAS No. 95 (FASB 1987).
20
variables in the main regression model are deleted from the sample. This results in a full sample
consisting of 75,188 firm-year observations with the national level measure. The sample size is
reduced to 23,307 firm-year observations with the city level measure. This measure is calculated
for the years 2000 to 2008 with auditor city data in Audit Analytics and a corresponding city in the
U.S. Census Bureau MSA classification.
For the discretionary revenue analyses, firms without the additional variables required to
calculate the discretionary revenue proxy as per Stubben (2010), in particular the changes in
accounts receivables from the statement of cash flows, are deleted from the full samples used in
the discretionary accruals analyses. This results in a full sample consisting of 69,512 firm-year
observations with the national level measure and 21,914 firm-year observations with the city level
measure.
For the propensity to issue going-concern analyses, this study uses U.S. public company
data for the years 2000 to 2008 from COMPUSTAT and auditor opinion data from Audit
Analytics. Firms in the financial services industries, firms with negative assets, market price, or
sales, and firms without the necessary data to calculate the control variables in the main regression
model are deleted from the sample. This results in a full sample consisting of 35,406 firm-year
observations with the national level measure and 23,349 firm-year observations with the city level
measure.18
1.5 RESULTS
1.5.1 Discretionary accruals –full sample analyses
Table 1 presents the descriptive statistics of the full sample used in the discretionary
18
Some prior studies, such as Balsam et al. (2003) and Krishnan (2003) eliminate clients of the non-Big 4 firms from
their sample in order to get a cleaner test of specialization separate from a possible Big 4 effect. In order to get the
largest possible sample size, clients of all firms are included in the main analyses, controlling for the Big 4 effect using
an indicator variable for clients of the Big 4 auditors. This is consistent with Reichelt and Wang (2010).
21
accruals analyses. Table 1, Panel A, shows how client size, performance, and total accruals vary
across the top eight auditors ranked by within-industry market share at U.S. national level. Ranks
are calculated by industry and year. For example, for two-digit SIC 49, in year 2007, the auditor
with the highest market share for that two-digit SIC will be in the first rank, the auditor with the
second highest will be in the second rank, and so on. Clients of auditors with high market share are
larger, have better performance, lower absolute total accruals and lower absolute discretionary
accruals and discretionary revenue. This pattern is persistent regardless of the cut-off value used to
divide specialist and non-specialist auditors.
Table 1, Panel B, shows the descriptive statistics of the national level full sample in
Column (I), and for a partition using NLEAD1 as measure of auditor industry specialization in
Columns (II) to (IV). Clients of national level specialist auditors represent 10.84 percent of the
total sample, similar to the 11.6 percent reported in Reichelt and Wang (2010, p.658). The
univariate tests of differences in means in Column (IV) show that clients of national level
specialists have on average approximately 2 percent lower absolute discretionary accruals
(difference in ADA = -0.017); are more than two times larger (difference in Total assets = 2,190);
have more leverage (difference in LEV = 0.029); have lower total accruals (difference in
ABS(ACCRL) = -0.049); have better performance in terms of profitability (difference in ROA =
0.046); have lower incidence of losses (difference in LOSS = -0.088); have higher cash flow from
operations (difference in CFO = 0.036); and have higher sales growth (difference in GROWTH =
0.008), compared to clients of non-specialist auditors.
Table 1, Panel C, shows the descriptive statistics of the city level full sample in Column (I),
and for a partition using CLEAD1 as measure of auditor industry specialization in Columns (II) to
(IV). Clients of city level specialist auditors represent 33.8 percent of the total sample, similar to
22
the 35 percent reported in Reichelt and Wang (2010, p.658). Panel C, Column (IV), shows that the
city level partition has similar differences in characteristics compared to the national level partition
in Panel B, Column (IV); however, the size difference (Panel C, Column IV, difference in Total
assets = 3,251) is more pronounced in the city level partition than in the national level partition
(Panel B, Column IV, difference in Total assets = 2,190).
Table 2 presents the results of the full-sample regression analyses using NLEAD1 and
CLEAD1 as measures of auditor specialization. In line with previous studies, Column (I) shows
that the coefficient on NLEAD1 is -0.0037, and Column (IV) shows that the coefficient on
CLEAD1 is -0.0026, and both coefficients are statistically significant (at one and five percent
level, respectively). These coefficients indicate that clients of specialist auditors have between 2.6
and 3.7 lower discretionary accruals compared to clients of non-specialists auditors.
1.5.2 Discretionary accruals –matched sample analyses
Table 1, Panel B, presents the descriptive statistics for the national level matched sample of
clients of specialist and non-specialist auditors. Using NLEAD1 as measure of auditor industry
specialization, 5,479 clients of the specialist auditors are matched to the same number of clients of
the non-specialist auditors within the specified criteria. Column (VII) of Table 1, Panel B, shows
that in the matched sample, clients of the specialist auditors have statistically insignificant
differences in absolute discretionary accruals (difference in ADA = -0.001); are on average
approximately 1.2 times larger (difference in Total assets = 537); have more leverage (difference
in LEV = 0.009); have statistically weak differences (at ten percent level) in performance in terms
of profitability (difference in ROA = 0.008) and cash flow from operations (difference in CFO =
0.007); and have statistically insignificant differences in the incidence of losses (difference in
LOSS = -0.009), total accruals (difference in ABS(ACCRL) = -0.004), and sales growth (difference
23
in GROWTH = -0.002), compared to the clients of specialist auditors. Other variables still exhibit
statistically significant differences, but the magnitude of the differences is considerably smaller
than in the full sample (Panel B, Column IV). These results show that the matching procedure
balances performance and growth, and mitigates size differences, but it does not fully mitigate
differences in all variables across the two auditor groups.
Table 1, Panel C, presents the descriptive statistics for the city level matched sample of
clients of industry specialist and non-specialist auditors. Using CLEAD1 as measure of auditor
industry specialization, 4,979 clients of the specialist auditors are matched to the same number of
clients of the non-specialist auditors within the specified criteria. At the city level, the matching
procedure is not as effective in mitigating size and performance differences, primarily due to a
different number of potential control observations for each treatment observation in the full
sample. Column (VII) of Table 1, Panel C, shows that in the matched sample, there is a weak
statistical difference of -0.002 (at ten percent level) in mean absolute discretionary accruals
between specialist and non-specialist auditor clients; however, the magnitude of the difference is
only 7 percent of the difference in means for the full sample (-0.002 in Panel C, Column VII;
compared to -0.028 in Panel C, Column IV).
Table 2 presents the results of the pooled matched sample regression analyses using
NLEAD1 and CLEAD1 as measures of auditor specialization. Column (III) shows that the
coefficient on NLEAD1 is 0.0006, and Column (IV) shows that the coefficient on CLEAD1 is
0.0009, and both coefficients are statistically insignificant. These results are confirmed by the
results in Table 3, where the statistically insignificant coefficients on the intercepts (-0.0013 in
Column I, and -0.0014 in Column II) of the pair-wise differences regression models, estimated
using the NLEAD1 and CLEAD1 matched samples, indicate that there are no differences in
24
absolute discretionary accruals between specialist and non-specialist auditors, even after
controlling for the effect of unmatched characteristics between observations. The combined
evidence from the univariate difference in means, pooled multivariate regressions, and pair-wise
differences regressions documented in Tables 1 to 3, suggests that after controlling for differences
in client characteristics between the two auditor groups by matching, the extant research design is
unable to detect differences in absolute discretionary accruals as a result of auditor industry
specialization.
1.5.3 Discretionary revenue –full sample analyses
Table 4 presents the descriptive statistics of the full sample used in the discretionary
revenue analyses. Table 4, Panel A, shows the descriptive statistics of the national level full sample
and for a partition using NLEAD1 as measure of auditor industry specialization. Clients of national
level specialist auditors have similar proportions as those in the discretionary accruals sample.
Panel A, Column (IV), shows that clients of national level specialist have on average
approximately 0.8 percent lower absolute discretionary revenue (difference in ADREV = -0.008).
The separation of client characteristics is very similar to those in the discretionary accruals
samples shown in Table 1.
Table 4, Panel B, shows the descriptive statistics of the city level full sample and for a
partition using CLEAD1 as measure of auditor industry specialization. Panel B, Column (IV),
shows that clients of city level specialist auditors exhibit similar characteristics as the clients of
national level specialist auditors, compared to the clients of non-specialist auditors; however, the
size difference (Panel B, Column IV, difference in Total assets = 3,265) is more pronounced in the
city level partition than in the national level partition (Panel A, Column IV, difference in Total
assets = 2,162).
25
Table 5 presents the results of the full-sample regression analyses using NLEAD1 and
CLEAD1 as measures of auditor specialization. Column (I) shows that the coefficient on NLEAD1
is -0.0013, and Column (IV) shows that the coefficient on CLEAD1 is -0.0020, and both
coefficients are statistically significant (at five and one percent level, respectively). These full-
sample results are consistent with the discretionary accruals full-sample results.
1.5.4 Discretionary revenue –matched sample analyses
Table 4, Panel A, presents the descriptive statistics for the national level matched sample of
clients of industry specialist and non-specialist auditors. Using NLEAD1 as measure of auditor
industry specialization, 5,053 clients of the specialist auditors are matched to the same number of
clients of the non-specialist auditors within the specified criteria. Column (VII) of Table 4, Panel
A, shows that in the matched sample, clients of the specialist auditors have a statistically
insignificant difference in mean absolute discretionary revenue (difference in ADREV = 0.001),
compared to the clients of non-specialist auditors. Furthermore, Column (VII) of Table 4, Panel A,
shows that the matching procedure balances growth (difference in GROWTH = -0.003) and the
incidence of losses (difference in LOSS = -0.012); and that the matching procedure mitigates
differences in size (difference in Total assets = 540), performance (difference in ROA = 0.007) and
cash flow from operations (difference in CFO = 0.006).
Table 4, Panel B, presents the descriptive statistics for the city level matched sample of
clients of industry specialist and non-specialist auditors. Using CLEAD1 as measure of auditor
industry specialization 4,695 clients of the specialist auditors are matched to the same number of
clients of the non-specialist auditors within the specified criteria. Column VII shows that at the
city level, the matching procedure is not as effective at mitigating differences in size (difference in
Total assets = 706) and performance (difference in ROA = 0.020); nevertheless, in this matched
26
sample, there is a statistically insignificant difference in mean absolute discretionary revenue
between specialist and non-specialist auditor clients (difference in ADREV = -0.001).
Table 5 presents the results of the pooled matched sample analyses using NLEAD1 and
CLEAD1 as measures of auditor specialization. Column (III) shows that the coefficient on
NLEAD1 is 0.0011, and Column (IV) shows that the coefficient on CLEAD1 is 0.0004, and both
coefficients are statistically insignificant. These results are confirmed by the results in Table 6,
where the statistically insignificant coefficients on the intercepts (0.0005 in Column I, and -0.0002
in Column II) of the pair-wise differences regression models, estimated using the NLEAD1 and
CLEAD1 matched samples, indicate that there are no differences in absolute discretionary revenue
between specialist and non-specialist auditors, even after controlling for unmatched characteristics
between observations. These matched sample results are in line with the discretionary accruals
matched sample results.
1.5.5 Propensity to issue a going-concern opinion –full sample analyses
Table 7 presents the descriptive statistics of the full sample used in the going-concern
analyses. This sample is smaller than the previous two samples because auditor opinion data in
Audit Analytics is only available since 2000. Table 7, Panel A, shows the descriptive statistics of
the national level full sample and for a partition using NLEAD1 as measure of auditor industry
specialization. Clients of national level specialist auditors have a slightly higher proportion than
those in the previous two full samples and exhibit approximately 4.3 percent incidence of going
concern, compared to 10.2 percent for clients of non-specialist auditors. This is consistent with
auditor specialists having larger and more profitable clients, which are generally less likely to go
bankrupt. Table 7, Panel B, shows the descriptive statistics of the city level full sample and for a
partition using CLEAD1 as measure of auditor industry specialization.
27
Table 8 presents the results of the full-sample analyses using NLEAD1 and CLEAD1 as
measures of auditor specialization. Column (II) shows that only the coefficient on the city level
measure of specialization of 0.2722 is significant (at one percent level) and in the right direction,
indicating that clients of city level specialist auditors are more likely to issue going-concern
opinions.
1.5.6 Propensity to issue a going-concern opinion –matched sample analyses
Table 7, Panel A, presents the descriptive statistics for the national level matched sample of
clients of industry specialist and non-specialist auditors. Using NLEAD1 as measure of auditor
industry specialization, 2,539 clients of the specialist auditors are matched to the same number of
clients of the non-specialist auditors within the specified criteria. Column (VII) of Table 7, Panel
A, shows that in the matched sample, clients of the specialist auditors have statistically
insignificant differences in the incidence of going-concern opinion (difference in GCONCERN =
0.003). Panel B shows the city level matched sample. Using CLEAD1 as measure of auditor
industry specialization, 4,951 clients of the specialist auditors are matched to the same number of
clients of the non-specialist auditors within the specified criteria. Column (VII) of Table 7, Panel
B, shows that in the matched sample there is a statistically insignificant difference in the incidence
of going-concern opinion between clients of specialist and non-specialist auditor clients
(difference in GCONCERN = -0.004).
Table 8 presents the results of the pooled matched sample regression analyses using
NLEAD1 and CLEAD1 as measures of auditor specialization. The coefficient on both variables, in
Columns (III) and (IV), is statistically insignificant. In Table 9, the statistically insignificant
coefficients on the NLEAD1 and CLEAD1 variables in the conditional logistic regressions indicate
that there are no differences in propensity to issue a going-concern opinion between specialist and
28
non-specialist auditors, controlling for unmatched characteristics between observations. The
results of this procedure should be interpreted with caution because they are based on a small
sample of matched pairs with intra-pair variation in going-concern opinions; however, these results
confirm the results from the overall difference in means between specialist and non-specialist
auditors and from the pooled logistic regression.19
After matching economically-comparable clients between specialist and non-specialist
auditors, in all the matched samples, the treatment effects of specialist auditors are insignificantly
different from those of non-specialist auditors with respect to absolute discretionary accruals,
absolute discretionary revenue, and auditor’s propensity to issue a going-concern opinion.
1.6 ANALYSES OF AUDITOR SWITCHES
Taking advantage of the setting created by the demise of Arthur Andersen (AA), after the
firm’s indictment for obstruction of justice in 2002, this Section examines the pre-post changes in
absolute discretionary accruals and absolute discretionary revenue for a sample of former AA
clients that switched to an auditor with a different degree of industry specialization.
Previous studies examining the consequences of this unique exogenous shock find that
there was a negative market reaction for AA clients during the key dates in the AA trial (Chaney
and Philipich 2002; Callen and Morel 2003), although there were some confounding market events
around the same dates (Nelson et al. 2008). In addition, successor auditors required more
conservative accounting for former AA clients (Nagy 2005; Cahan and Zhang 2006); however, the
differences pre-post switch were related to whether former AA employees continued auditing the
client after they were hired by the successor auditor in 2002 (Blouin et al. 2007). Finally, Knechel
et al. (2007a) examines 318 non-AA clients that switched auditors between 2000 and 2003, and
19
These models are estimated using the Stata command clogit. Tenure, year, and industry-specific intercepts are not
included because there is insufficient intra-pair variation in these variables.
29
documents that those clients who switched from non-specialist to specialist auditors experienced
statistically significant positive abnormal returns of 2.5 percent surrounding the date of the auditor
change, providing additional evidence of a perceived specialist auditor effect.
The following regression model is estimated for AA clients that switched auditors in 2002
in order to test whether there was a pre-post effect of a switch to an auditor with a different degree
of industry specialization:
ΔQUALITY_MEASUREi = δ0 + δ1ΔLEADi + δ2ΔBIG4i + δ3ΔLOGMKTi + δ4ΔLEVi
+ δ5ΔROAi+ δ6ΔROALi+ δ7ΔLOSSi+ δ8ΔCFOi + δ9ΔBTMi
+ δ10ΔABS(ACCRL) i + δ11ΔGROWTHi+ δ12ΔALTMANi
+ δ13ΔSTDEARNi + vi (8)
where Δ denotes the difference between the level of each variable (as previously defined) in 2002
and the level of that variable in 2001. This model uses each client as its own control. The intercept
δ0 represents the average change in the dependent variable controlling for changes in other client
characteristics, and the coefficient δ1 on ΔLEAD represents the incremental change as a result of
switching between specialist and non-specialist auditors (ΔNLEAD and ΔCLEAD at national and
city level respectively). If specialist auditors are better at detecting and undoing earnings
management, it is expected that a switch to a specialist auditor will decrease absolute discretionary
accruals and absolute discretionary revenue.
In order to isolate the specialization effects from other potential year-specific effects,
influencing changes in absolute discretionary accruals and absolute discretionary revenue from
2001 to 2002, it is possible to compare the AA clients to a control group of non-AA clients. For
each specialization measure, AA clients are matched to non-AA clients in 2001 by fiscal year-end
month and industry, within a 50 percent size distance, keeping the pairs with the highest stock
30
return covariance from all the possible matches as defined in Section 1.2.5. The following pre-post
regression model is estimated using the matched sample of AA and non-AA clients:
ΔQUALITY_MEASUREi = ρ0 + ρ1AAi + ρ2ΔLEAD1i + ρ3AA_ ΔLEADi + ρ4ΔBIG4i
+ ρ5ΔLOG_MKT + ρ6ΔROAi+ ρ7ΔROALi+ ρ8ΔLOSSi
+ ρ9ΔCFOi + ρ10ΔBTMi + ρ11ΔABS(ACCRL) i
+ ρ12ΔGROWTHi+ ρ13ΔALTMANi + ρ14ΔSTDEARNi + zi (9)
where Δ denotes the level of each variable (as previously defined) in 2002 minus the level of that
variable in 2001 for all clients in the sample, and for client i:
AA = ―1‖ for AA clients and ―0‖ otherwise; and,
AA_ΔLEAD = interaction term between AA and changes between specialist
auditors, where ΔLEAD= ―-1‖ for clients that switched in industries
where AA was a specialist and the successor is not a specialist
auditor, ΔLEAD= ―1‖ for clients that switched in industries where
AA was not a specialist and the successor is a specialist auditor, and
ΔLEAD= ―0‖ for all other cases.
The pre-post analyses uses a sample of 393 AA clients that switched to a different auditor
during 2002, and for which discretionary accruals, discretionary revenue, and auditor
specialization proxies can be calculated in both years.20
From these 393 AA clients, using the
national level measure of specialization NLEAD1, 47 clients switched to a specialist auditor in
industries where AA was not a specialist (ΔNLEAD1 = 1), 23 clients switched to a non-specialist
auditor in industries where AA was a specialist (ΔNLEAD1 = -1), and 323 clients stayed at the
same level of specialization (ΔNLEAD1 = 0); similarly, using the city level measure of
20
Going-concern opinions are not used in these analyses due to the low incidence of this variable within the clients the
AA sample. The industry leadership variable is calculated using market share of each auditor by industry in each year
2001 and 2002.
31
specialization CLEAD1, 72 clients switched to a specialist auditor in industries where AA was not
a specialist (ΔCLEAD1 = 1), 42 clients switched to a non-specialist auditor in industries where AA
was a specialist (ΔCLEAD1 = -1), and 279 clients stayed at the same level of specialization
(ΔCLEAD1i = 0). There are more upgrades to a specialist auditor and downgrades to a non-
specialist auditor in the sample at the city level than at the national level because there are is a
larger proportion of city level specialists.
Next, the sample of AA clients is matched to an economically-comparable group of non-
AA clients in 2001 to form a differences-in-differences sample with 287 AA clients and 287 non-
AA clients. From the 287 AA clients, using the national level measure of specialization NLEAD1,
33 clients switched to a specialist auditor in industries where AA was not a specialist (ΔNLEAD1 =
1), 11 clients switched to a non-specialist auditor in industries where AA was a specialist
(ΔNLEAD1 = -1), and 243 clients stayed at the same level (ΔNLEAD1 = 0). From the 287 non-AA
clients, using the national level measure of specialization NLEAD1, 23 clients upgraded to a
specialist auditor (ΔNLEAD1 = 1), 11 clients downgraded to a non-specialist auditor (ΔNLEAD1 =
-1), and 253 clients stayed at the same level of specialization (ΔNLEAD1 = 0). The switches using
the city level measure of specialization CLEAD1 are proportionally similar to those using the
national level measure NLEAD1.
Table 10, Panel A, shows the results of the multivariate pre-post analyses for the sample of
AA clients that switch to a different auditor in 2002. There is no evidence of a pre-post change in
discretionary accruals from switching between specialist and non-specialist auditors. The
coefficients on the ΔNLEAD1 variable (-0.0081) and the ΔCLEAD1 variable (0.0115) in Columns
(I) and (II) are statistically insignificant. Similarly, there is no evidence of a pre-post change in
discretionary revenue from switching between specialist and non-specialist auditors. The
32
coefficients on the ΔNLEAD1 variable (-0.0057) and the ΔCLEAD1 variable (-0.0011) in Columns
(III) and (IV) are also statistically insignificant.
Table 10, Panel B, shows the result of the differences-in-differences analyses, comparing
changes between 2002 and 2001 for AA and non-AA clients. In this Table, the main variable of
interest is the interaction between the AA indicator variable and the specialization variables,
AA_ΔNLEAD1 at the national level and AA_ΔCLEAD1 at the city level. This interaction variable
captures the effect of switching to an auditor with a different level of specialization for the former
AA clients. There is no evidence of a change in discretionary accruals or absolute discretionary
revenue from switching between specialist and non-specialist auditors for AA clients at the
national level. The coefficients on the AA_ΔNLEAD1 and AA_ΔCLEAD1 variable in Columns (I),
(III), and (IV) are statistically insignificant. Moreover, the coefficient on the city level
specialization measure for former AA clients AA_ΔCLEAD1 in the second column (0.0324) is
statistically significant (at ten percent level) in the opposite direction.
The results in Table 8 are robust to excluding clients in industries where AA was a
specialist, and switching to a non-specialist auditor (ΔLEAD1= -1), or clients in industries where
AA was a specialist, and switching to a specialist auditor (ΔLEAD1= 0 and LEAD1= 1). The
results in Table 8 are also robust to standard errors calculated using 1,000 bootstrap replications,
mitigating the concerns that the low statistical significance could be a result of small sample size.
An advantage of this research design is that it uses an exogenous shock to test whether
specialist auditors have a direct and immediate impact on the client’s financial reporting quality.
Nevertheless, there are limitations inherent to these analyses. First, as noted by Blouin et al.
(2007), in several instances, former AA employees were hired by the successor auditors and
continued to audit the same clients. Second, there were changes in the environment that may have
33
motivated all auditors, specialist and non-specialist, to be more conservative in 2002. Third, the
effect of auditor specialization may not be immediately reflected in the two proxies for financial
reporting quality used in these analyses. These results are incremental to the matched sample
analyses, providing additional evidence on the shortcomings of the extant methodology to test the
association between auditor industry expertise and audit quality.
1.7 SIMULATION ANALYSES
Using a simulation procedure, this Section examines whether the observed association
between audit quality and auditor industry specialization could be observed when clients are
assigned to five auditors generated at random. This simulation approach aims to examine the
effectiveness of the extant methodology to isolate the effects of client characteristics from the
effects of auditor industry specialization. The simulation is conducted in four steps. First, all
clients in the full discretionary accruals sample are assigned to five auditors generated at random
each year. Second, industry specialists are designated based on their industry market share at
national level. The auditor with the highest market share is designated as specialist (NLEAD1 = 1)
and the other four auditors are designated as non-specialists (NLEAD1 = 0). Third, the main
regression model (Equation 6) is estimated using these simulated conditions. Finally, the first three
steps are repeated 1,000 times. 21
Figure 1 shows the coefficients for 1,000 iterations of the simulation procedure, employing
ADA as dependent variable and the simulated NLEAD1 as measure of auditor industry
specialization. The coefficient estimate for NLEAD1 is negative in 97.6 percent of the iterations
and has a mean statistically different from zero (at one percent level) of -0.0020. In these
21
This approach is similar to the one used by Carson and Fargher (2007) to assess the importance of client size to
determine audit fee premiums. For example, if in a given industry there are 20 clients, four clients would be assigned
to each auditor at random; following, the specialist auditor is the one with the largest industry market share.
34
simulations, the auditors that are randomly assigned a sufficient number of the largest clients in
each industry are often designated as specialists, and these designated specialists appear to be of
higher quality compared to non-designated specialists. Overall, the results from these simulations
are suggestive that client characteristics, and particularly client size, influence the observed
association between audit quality and auditor industry specialization. There is a limitation inherent
to this analysis. The randomly assigned clients from the original full sample were actually audited
by specialist or non-specialist auditors. In some cases, the designated specialist in the simulation is
the same as the actual specialist. On average, the designated specialist auditor is also the actual
specialist for 16 percent of the clients. This limitation is mitigated by the number of replications
and by the fact that the same simulated specialization effects are observed in samples drawn from
industries without an actual specialist auditor. Figure 2 shows the coefficients for 1,000 iterations
of the simulation procedure, employing ADA as dependent variable and NLEAD1 as measure of
auditor industry specialization for industries without a specialist. The coefficient estimate for
NLEAD1 is negative in 69.3 percent of the iterations and has a mean statistically different from
zero (at one percent level) of -0.0009.
Employing ADREV as dependent variable and NLEAD1 as measure of auditor industry
specialization yields qualitatively similar results to those described in this Section.
1.8 ADDITIONAL MATCHING ANALYSES AND SENSITIVITY TESTS
1.8.1 Propensity-score matching
Another approach that can be used to find comparable firms is the propensity-score
matching methodology proposed by Rosenbaum and Rubin (1983). Propensity-score matching is a
widely used methodology to find a group of comparable cases and control observations to mitigate
the effect of self-selection in observational causal studies. In general, this approach can be used to
35
pair match observations that belong to two different regimes, in the context of this study, to find
comparable clients audited by specialist and non-specialist auditors. A potential drawback of this
approach is that it depends on the specification of the choice model, known as the ―strongly
ignorable treatment assignment assumption.‖ The main advantage of propensity-score matching is
that it is usually effective at selecting observations that are closely matched in the predefined
covariates.
For each audit-quality proxy and specialization measure, clients of specialist and non-
specialist auditors are matched using propensity scores. The propensity of choosing specialist
auditors at national or city level is predicted using a logistic regression where the dependent
variable is the specialist indicator variable and the independent variables are all the control
variables in Equation (6), including industry and year-indicator variables. Observations are
matched by propensity score, within common support, without replacement, using a caliper
distance of 0.03. Following, the main multivariate model is estimated in the matched sample of
clients of industry specialist and non-specialist auditors.22
Using this methodology, clients of the
industry specialist auditors have statistically insignificant differences in absolute discretionary
accruals, absolute discretionary revenue, and incidence of going-concern opinions, compared to
clients of non-specialist auditors.
22
These propensity score settings are consistent with Lawrence at al. (2011) and generally result in balanced
covariates between auditor groups. Similar results are obtained by reducing the caliper distance, although this reduces
the sample size further. Using the logarithm of total assets in the propensity score calculation as a size variable, instead
of the logarithm of market value, produces qualitatively similar results in the matched samples as those documented in
the main tables. In general, the logarithm of total assets in the model results in more balanced client characteristics
between auditor groups. There is only a difference in the city level going-concern analysis, where using the logarithm
of market value, in addition to the other covariates of the propensity score model, results in a statistically significant
coefficient on the variable CLEAD1 (at five percent level) in the matched sample regression model; while using the
logarithm of total assets, in addition to the other covariates of the propensity score model, results in a statistically
insignificant coefficient.
36
1.8.2 Size and industry matching
Following Lawrence et al. (2011), clients of specialist and non-specialist auditors are also
matched by year, industry, and total assets. Individual observations are matched using propensity
score, estimated using the logarithm of total assets and industry and year indicator variables as
predictors in the logistic regression, within common support, without replacement, and using a
caliper distance of 0.03. In the industry and size matched samples, clients of the specialist auditors
have statistically insignificant differences in absolute discretionary accruals, absolute discretionary
revenue, and incidence of going-concern opinions, compared to clients of non-specialist auditors,
estimated using the main model on each matched sample.
In general, the industry and size matching is successful at balancing size, performance, and
leverage between auditor groups, but is not successful at mitigating differences in cash flow from
operations and absolute accruals between auditor groups. In the large samples used in this study,
matching on industry, within 50 percent size, and returns covariance is similar as matching closely
only on industry and size. These alternatives might not be equivalent in small samples where
idiosyncratic differences need to be more closely matched between the case and control groups,
and for those samples researchers should aim to use the comparability measures that produce the
best possible balance between matched observations. In general, companies of very similar size
within an industry have correlated stock returns and exhibit similar performance, and matching
clients on these criteria using alternative specifications shows that the extant research design
cannot distinguish between the clients of specialist and non-specialist auditors.
1.8.3 Bootstrap, random sub-samples, and stratified samples
To mitigate concerns that the lack of significance in the matched samples analyses is a
result of smaller sample sizes, this Section documents the results of three additional sensitivity
37
analyses. First, bootstrap standard errors are estimated for all the matched sample models using
1,000 replications. These analyses produce qualitatively similar results as those shown in the main
tables. Second, for the national and city level full samples of the three audit-quality proxies, a
random sub-sample of the same size as the matched sample is drawn from the full sample, and the
main models are estimated in that sub-sable with bootstrap standard errors estimated using 1,000
replications. Sample size does not affect the results for the discretionary accruals and discretionary
revenue models; however, for the going-concern samples, the results are more sensitive to sample
size due to the low incidence of going-concern opinions, equal to 9.5 percent in the full sample.
Third, the matched sample results hold separately for industries where auditor specialization could
matter incrementally to detect earnings management or to determine the probability of going
concern. Managers could have more opportunities for manipulation in industries with high total
accruals and high volatility of earnings, and may also face higher incentives to meet expectations
in competitive or high-growth industries. Likewise, determining the probability of going concern
is difficult for low-growth industries, where competition is intense and there is high-earnings
volatility. For each industry and year in the matched samples, median total accruals is calculated
using the variable ABS(ACCRL), median sales growth is calculated using the variable GROWTH,
median industry concentration is calculated using the Herfindahl index based on total assets, and
median earnings volatility is calculated using the variable STDEARN. Next, industries are ranked
by year using the industry median for each of these variables, and the main models are estimated
separately for observations in the top and bottom quartiles. Separating industries by these variables
produces similar results to those documented in the main tables using the full matched samples.
1.8.4 Alternative measures of auditor industry specialization
All the full and matched sample analyses are also repeated using an alternative market
38
share cut-off for the national and city level specialist measures, LEAD30p equal to ―1‖ for auditors
that have over 30 percent market share in a given industry and year at the national or city level,
and ―0‖ otherwise. This measure results in a greater number of clients deemed to be audited by a
specialist, and larger matched samples than in the results shown in the main tables. Using this
alternative measure, the effects of auditor industry specialization are also statistically insignificant
in the matched samples for all three audit-quality proxies and in the pre-post switch analysis for
Arthur Andersen clients. Moreover, using this alternative measure, the coefficient on the city level
specialist variable (CLEAD30p = 0.0042) for the discretionary accruals matched sample is
significant (at one percent level) in opposite direction, mitigating the concerns that the lack of
statistical significance in the match samples is a result of smaller sample size.
Similarly, the simulation analyses are also repeated using this alternative market share cut-
off. The results of the simulation are qualitatively similar to those described in Section 1.7. Client
characteristics, and particularly client size, influence the observed association between audit
quality and auditor industry specialization.
I addition, all the full and matched sample analyses are repeated using a combined measure
of and national and city level auditor industry specialization, equal to ―1‖ for auditors that are
specialists at both levels in a given industry and year, and ―0‖ otherwise. Clients of combined
national and city level specialists are matched with clients of other auditors. The effect of
combined national and city level auditor industry specialization is statistically insignificant in the
matched samples for all three audit-quality proxies.
1.9 CONCLUSION
To determine causal inference in observational studies, empirical researchers should aim to
compare treated and control groups that have similar client characteristics, ideally approximating
39
experimental conditions. A potential way to achieve this objective is by matching treatment and
control observations on all relevant observable dimensions except for the treatment variable. This
Chapter proposes a methodology to find economically-comparable clients and applies it to
mitigate the effect of differences in client characteristics in comparing the relative quality between
industry specialist and non-specialist auditors.
This study employs discretionary accruals, discretionary revenue, and propensity to issue a
going-concern opinion as proxies for audit quality, and within-industry market share proxies for
auditor industry specialization at the U.S. national and city level. After matching clients of
specialist and non-specialist auditors, the previously documented association between auditor
industry specialization and the audit-quality proxies may be attributed to client characteristics.
This study proposes a methodology to match clients on economic comparability, by year, industry,
size, and returns covariance. Alternatively, clients are matched using different specifications of a
propensity score model.
Furthermore, there is no evidence of a pre-post change in discretionary accruals or
discretionary revenue resulting from an exogenous switch between specialist and non-specialist
auditors for a sample of Arthur Andersen clients that changed auditors in 2002. Finally, using a
simulation approach, this study documents that client characteristics, and particularly client size,
influence the observed association between audit quality and auditor industry specialization.
The results of this study do not imply that industry knowledge does not contribute to audit
quality, but that the extant methodology does not fully capture the effects of auditor industry
expertise. The main arguments of this study can be summarized in three steps. First, there is a
strong prior that auditor industry expertise should improve audit quality. Second, the extant
archival evidence seems to confirm that prior; however, this study shows that the previously
40
proposed research design might not be robust to controlling for differences in client characteristics
between specialist and non-specialist auditors. Finally, if we assume that the strong prior holds, the
findings of this study suggest that the previously proposed research design is not capturing the
effects of auditor industry expertise. The methodology used in this Chapter could be useful to
other studies examining the consequences of auditor industry specialization, and may motivate
further research on alternative proxies and research designs to investigate the effects of auditor
industry specialization.
41
CHAPTER 2 –DOES USING INFORMATION FROM PEER FIRMS IMPROVE
ACCOUNT-LEVEL EXPECTATION MODELS?
2.1 INTRODUCTION
Determining accurate account-level expectations in order to assess the reasonability of a
company’s financial information is both an open empirical issue for researchers and a matter of
practical importance for auditors, regulators, and financial analysts. In particular, auditors employ
account-level expectations in analytical audit procedures (AAP) to identify abnormal line-item
balances that differ from expected values by a significant amount or that appear inconsistent with
other relevant information. This study examines the effectiveness of account-level expectation
models for accounts receivables and inventory, relying on historical and contemporaneous
information from a company and its economically-comparable peer firms to detect potential
account-level errors or manipulation.
This study fits within the AAP literature on expectation models (Vandervelde et al. 2008;
Hoitash et al. 2006), but also draws from related studies aiming to detect earnings management in
total accruals (Kothari et al. 2005; Dechow et al. 1995), to detect earnings management in specific
accounts (Caylor 2010; Stubben 2010; Marquardt and Wiedman 2004), to detect the overall
probability of fraud (Dechow et al. 2010; Beneish 1999), and to assess financial information at the
account-level (Melumad and Nissim 2009). The results of this study confirm the findings of the
earnings management literature that industry cross-sectional prediction models can be equally or
better specified than firm-level time-series models. The cross-sectional approach is consistent with
the International Audit Standard 520 recommending using ―similar industry information, such as a
comparison of the entity’s ratio of sales to accounts receivable with industry averages or with other
entities of comparable size in the same industry‖ (IFAC 2009, p. 436). This approach eliminates
42
the need for a long time-series in estimating the expectation models and ignores the information
from prior periods beyond last year; however, in fast growing or changing industries, information
from prior periods may have lower relevance than cross-sectional information. Furthermore, an
advantage of cross-sectional models is that they can be estimated when historical data is not
available, such as in the years following the adoption of IFRS. This may help in overcoming a
limitation pointed out by Ng (2008, p. 50) that ―the lack of historical information arising from a
wholesale change in accounting basis may result in reduced effectiveness of analytical
procedures.‖
The general theme linking the earnings management and the AAP literature is the need to
find abnormal patterns in accounting information; however, a practical limitation in conducting
AAP is that all data that enters the expectation models has to be available to the auditor during the
course of the audit, before year-end financial statements are released. For example, in earnings
management studies, a cross-sectional discretionary accruals model, such as the Jones (1991)
model, is usually estimated by industry using year-end data for a sample of firms available to the
researcher. In contrast, the auditor faces a single-client situation where year-end financial
statements for the client and its potential peers are usually not available during the planning phase
of the audit. The methodology examined in this study is mostly applicable during the planning
phase of the audit process, assuming that the auditor has access to two or three quarters of publicly
available financial information for a client and its potential peer firms.
The U.S. Audit Standard on analytical procedures (AU 329, formerly SAS No. 56, AICPA
1988) and the International Audit Standards (ISA 520, IFAC 2009) prescribe performing AAP in
the planning stages and at the end of the audit process. AAP include comparison of the client’s
financial statement accounts and ratios over time, comparison of the client’s financial statement
accounts and ratios against those of similar companies, and comparison of the client’s financial
43
and non-financial information. In essence, the idea behind AAP is that auditors may detect high
risk areas, direct their attention towards potential misstatements, and obtain confirmatory evidence
for other audit tests by examining the reasonability of the client's financial information.23
Auditors
conduct AAP in three general steps, first they set an expectation for a particular financial statement
account or ratio, second they compare the difference between actual and expected numbers, and
finally they perform additional audit procedures to explain any unexpected differences. For
example, if the client's accounts receivable, or the ratio of accounts receivable to sales, are
significantly higher this year compared to last year, or higher than the industry average, this could
potentially signal aggressive revenue recognition, and the auditor should perform additional
auditing procedures to mitigate the risk of revenue overstatement.
Audit regulators have expressed concerns about the quality of analytical procedures in the
audits of public companies in the US. For example, the Public Company Accountability Board
(PCAOB) has noted deficiencies in the application of analytical procedures in 13 out of 24 annual
inspections of Big 4 accounting firms performed between 2003 and 2008. The deficiencies
identified by PCAOB comprise: failure to establish precise expectations for account balances,
inappropriate follow-up of differences between actual and expected balances, and inappropriate
reliance on analytical procedures as substantive audit tests (for a summary of these deficiencies see
Appendix C). The evidence from the PCAOB inspections corroborates the conclusions of the
review conducted by the Panel on Audit Effectiveness (Public Oversight Board 2000, p.42)
highlighting that ―analytical procedures were not as effective as primary procedures when
expectations were not properly developed, materiality thresholds for investigating differences were
not clearly established, the analytical procedures were superficial or explanations were not
23
AU 329 (AICPA 1988, p.339) states that ―A basic premise underlying the application of analytical procedures is that
plausible relationships among data may reasonably be expected to exist and continue in the absence of known
conditions to the contrary.‖
44
corroborated, and the documentation was inadequate.‖ Furthermore, a recent study by Trompeter
and Wright (2010) shows that accounting firms are focusing more on AAP following the passage
of the Sarbanes-Oxley Act of 2002, increasing the time dedicated to AAP in audit engagements,
estimated by a group of Big 4 partners to be approximately 25 percent of the total audit hours.
This study examines the specification and power of three different approaches to develop
expectation models for AAP. The first approach resembles the process commonly used by auditors
to conduct AAP, relying on a one-period firm-specific heuristic model that uses prior working
capital account balances and current year growth in related income statement accounts to predict
current year’s working capital balances.24
The second approach follows the extant literature on
AAP, using multivariate expectation models estimated using a time-series of historical data at the
firm-level. The third approach contributes to the AAP literature showing the advantages of
implementing industry cross-sectional expectation models. Moreover, this study assesses whether
each approach can be improved by introducing information from economically-comparable peer
firms into the analysis. In general, a reliable expectation model should be reasonably well
specified or unbiased, showing zero mean error for random and stratified samples (e.g., across sub-
samples of high-growth and low-growth companies), and should also have power for detecting
errors or manipulation. As expectation models become more precise, it is easier to attribute
significant differences between actual and expected amounts to manipulation and not to random
noise.
The main contribution of this study is to provide a framework to use information from
economically-comparable peer firms in account-level expectation models. This study proposes a
peer-matching methodology in a setting where multiple dimensions of economic comparability
24
This study also examines a heuristic model based on a moving average of the twelve quarters prior to the current
year. Auditors may choose to use averages as expectorations instead of relying only on adjusted or incremental year-
over-year expectations.
45
need to be aggregated, and where other techniques such as propensity-score matching are not
applicable.25
Moreover, this study provides evidence that the proposed economic-peer approach
improves the specification of the three types of models, and improves the detection power of time-
series and industry cross-sectional models. Comparing between models, one-period incremental
heuristic models are generally unreliable, and industry cross-sectional models can be more
effective than previously proposed time-series models.
The simulation results in this study, seeding directional account-level errors equal to one
percent of total assets, show that the time-series models including information from peer firms
detect manipulation in accounts receivables in approximately 22 percent of the simulated samples,
and in inventory in approximately 60 percent of the simulated samples. In contrast, industry cross-
sectional models including information from peer firms detect manipulation in accounts
receivables in 80 percent of the simulated samples, and in inventory in over 91 percent of the
simulated samples. Without information from peer firms, the time-series models perform better for
inventory, detecting manipulation in 55 percent of the samples, and the cross-sectional models
perform better for accounts receivables, detecting manipulation in 32 percent of the samples.
Finally, the results from including multiple peers in the analysis suggest that auditors may benefit
from focusing on finding a small number of closely-matched peer firms in performing AAP.
The models proposed here can be potentially applied to a large number of public U.S.
firms; however, it may be possible for auditors to construct increasingly more precise expectations,
25
Propensity-score matching is a widely used methodology to find a group of comparable case and control subjects in
order to estimate treatment effects. In general, propensity-score matching can be used to match observations that
belong to two different regimes. It is applicable in observational causal studies where treatment choice is not random.
Control observations are matched to treatment observations based on a specified distance between their overall
probabilities of undergoing treatment. These probabilities are usually estimated using a logistic regression model
where a number of covariates predict treatment choice, effectively aggregating multiple dimensions into a single
matching variable. In the context of AAP there are not treated and non-treated firms, and thus propensity score models
are not applicable to find matched peer firms.
46
by using multiple relations between financial statement accounts and more detailed information for
a particular client. Nevertheless, using aggregated data is in line with common practice among
auditors, as documented by Hirst and Koonce (1996, p. 463), and with U.S. Auditing Standard AU
329 noting that ―analytical procedures used in planning the audit generally use data aggregated at
the account level‖ (AU 329, p. 348 AICPA, 1988). Furthermore, the main objective of this study is
to gain insights about the incremental value of peer-firm information. The approaches examined
here imply a trade-off between conducting a detailed examination using a small sample of firms
and multiple interrelated accounts (Vandervelde et al. 2008; Leitch and Chen 2003), and a general
analysis using quarterly data for a large sample of companies.
Overall, the findings of this study may be helpful to auditors of public companies, earnings
management researchers, audit regulators, and financial analysts for selecting expectation models
and identifying comparable peer firms to assess a company’s financial data at the account-level.
2.2 ACCOUNT-LEVEL EXPECTATION MODELS
2.2.1 Firm-level heuristic models
Heuristic models resemble the process followed by auditors to develop account-level
expectations for AAP. Studying the effectiveness of relatively simple and commonly used heuristic
models is important because these models are base-case benchmarks against which we can
evaluate more complicated statistical models. Hirst and Koonce (1996, p. 463) note that ―the most
common procedure at planning is fluctuation analysis, whereby current year unaudited account
balances are compared to prior year audited account balances‖, and also that ―analytical
procedures at planning are generally performed account by account at the financial statement
level.‖ Knechel et al. (2007b, p. 342) note that ―the comparison of financial statement balances
over time is the oldest and most common analytical procedures used by auditors.‖ Gauntt and
47
Glezen (1997, p. 56) mention that ―the most common approach to develop expectations is to use
prior period amounts‖ and that ―in some cases, however, an incremental approach to expectations
is used, whereby the prior year actual amounts are adjusted for known changes‖. Finally, Kinney
(1987, p.61) points out that ―a reasonable expectation for the ratio in month t is assumed to be the
average audited value of the same ratio for the previous audit year adjusted for changes in the
industry.‖
This study uses quarterly data and focuses on two working capital accounts, accounts
receivables and inventory, also using year-quarter changes in sales and cost of goods sold as
drivers of expected changes in accounts receivables and inventory.26
These two accounts are
suitable to test the incremental benefit of using peer information in AAP because their treatment is
common among similar firms within an industry, they appear in most balance sheets, and they
involve large volumes of transactions. Account balances are scaled by total assets to make the
expectation errors comparable across companies and models in terms of percentage of assets.
Previous research has found that accounts receivables and inventory are important for
financial statement users and could be targets for management’s manipulation. Lev and
Thiagarajan (1993) find that the market reacts to changes in these two accounts. Dechow et al.
(2010) find that change in accounts receivables and change in inventory are positively associated
with the probability of material misstatements. Melumad and Nissim (2009), note that a significant
increase of accounts receivables relative to sales may indicate that sales are overstated, and that the
valuation of inventory involves substantial estimates which can be exploited by management,
allowing them to overstate inventory write-downs to reduce cost of goods sold in subsequent
periods. Callen et al. (2008) show that the likelihood of revenue manipulation is positively
26
Using quarterly aggregated data is in line with U.S. Auditing Standard AU 329 (AU 329, p. 348 AICPA, 1988)
noting that the procedures ―may consist of reviewing changes in account balances from the prior to the current year‖
or ―might involve an extensive analysis of quarterly financial statements‖.
48
associated with high levels of accounts receivables adjusted for credit policies. Marquardt and
Wiedman (2004) show that firms undergoing an equity offering have high unexpected accounts
receivable but no significantly different unexpected inventory, compared to a sample of control
firms. Nelson et al. (2003) examine a sample of 515 earnings management attempts from a survey
of 253 experienced auditors, and document that revenue and reserve recognition, including
inventory reserves, constitute the most common earnings-management approaches. Beneish (1999)
finds that the change in the ratio of accounts receivables to sales is higher for a sample of firms
that manipulated earnings compared to a sample of control firms.
This study examines two types of heuristic models for accounts receivables and inventory,
first a one-year incremental approach relying on prior-year balances adjusted for expected changes
(Marquardt and Wiedman 2004), and also a historical approach relying on average prior
balances.27
Under the incremental approach, the expected accounts receivables in a given quarter
(E[ARi,t]) is obtained by multiplying the accounts receivables in the same quarter of the previous
year (ARi,t-4) by the change in sales (CHSALESi,t):
E[ARi,t] = ARi,t-4*CHSALESi,t (AR1)
In the accounts receivable incremental approach, change in sales (CHSALESi,t) is calculated
as the ratio of the current quarter sales to the sales in the same quarter of the previous year, and
account receivables is scaled by total assets at the end of the quarter. Under the historical
approach, the expected accounts receivables in a given year-quarter (E[ARi,t]) is obtained by
averaging 12 quarters of accounts receivables, as a percentage of assets, ending in the fourth
quarter of the prior audited fiscal year:
27
These models are similar to the Healy (1985) and DeAngelo (1986) models in the earnings management literature,
where non-discretionary accruals are expected to be the average total accruals over a number of prior periods, or last
year’s total accruals, respectively. If non-discretionary accruals are constant over time and discretionary accruals have
a mean of zero in the estimation period, then both models measure non-discretionary accruals without error.
49
𝐸[𝐴𝑅𝑖,𝑡] = ∑𝐴𝑅𝑖,𝑡−4−𝑘
12
12
𝑘=1
(AR2)
Similarly, under the incremental approach, the expected inventory in a given quarter
(E[INi,t]) is obtained by multiplying the inventory in the same quarter of the previous year (INi,t-4)
by the change in cost of goods sold (CHCOGSi,t):
E[INVi,t] = INVi,t-4*CHCOGSi,t (INV1)
In the inventory incremental approach, change in cost of goods sold (CHCOGSi,t) is
calculated as the ratio of the current quarter cost of goods sold to the cost of goods sold in the
same quarter of the previous year, and inventory is scaled by total assets at the end of the quarter.28
Under the historical approach, the expected inventory in a given quarter (E[INVi,t]) is obtained by
averaging 12 quarters of inventory, as a percentage of assets, ending in the fourth quarter of the
prior audited fiscal year:
𝐸[𝐼𝑁𝑉𝑖,𝑡] = ∑𝐼𝑁𝑉𝑖,𝑡−4−𝑘
12
12
𝑘=1
(INV2)
For the four models described above, the quarterly unexpected balance, or forecast error, is
calculated as the difference between actual and expected balances as a percentage of total assets.
Section 2.4, explains in more detail the measures used to assess the reliability of each model and
how quarterly estimates are averaged to calculate a year firm-level measure using the quarterly
estimates.
28
Using a heuristic inventory model with changes in sales instead of the changes in cost of goods sold yields similar
results to those documented in the main tables. In addition, each variable is scaled by average assets in all models,
instead of using total assets at the end of each quarter, and all results are similar to those documented in the main
tables.
50
2.2.2 Firm-level time-series models
Time-series models, estimated at the firm-level, are the most commonly examined
expectation models in the prior AAP literature (Wheeler and Pany 1990; Chen and Leitch 1999;
Leitch and Chen 2003; Hoitash et al. 2006; Vandervelde et al. 2008).29
Prior studies have assessed
the effectiveness of a variety of models including: martingale, Census X-11, ARIMA, structural
equations, and multivariate regression. Prior studies generally find that models predicting an
account balance or ratio using historical data for that account or ratio and concurrent data from
related accounts are more effective at predicting account-level errors than models relying only on
historical data. In these studies, first the expectation model is estimated at the firm-level using a
series of monthly or quarterly observations from previously audited years; and following, the
model parameters are fitted to the monthly or quarterly observations in the current year under
audit.
This study examines two multivariate models that use past information for each account
and concurrent information from a related account (e.g., receivables in the same quarter of the
prior-year and change in sales).30
The past ratio of accounts receivables or inventory to assets is
used as the first predictor of the current year-quarter ratio, and changes in sales or cost of goods
sold is used as an additional variable that proxies for shocks that impact the current year-quarter’s
ratios:
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + εi,t (AR3)
29
This study is closely related to Hoitash et al. (2006), examining how information from peer firms improves time-
series multivariate and single-account models; however, this study proposes a different procedure to select peer firms,
the peer-firm information is not averaged across peers, the effectiveness of each model is assessed using simulations,
and the main focus is on predicting current asset accounts (receivables and inventory) instead of their related income
statement accounts (sales and cost of goods sold). 30
Accounts receivable and sales, as well as inventory and cost of goods sold, are related by the nature of these
accounts and prior studies have shown that they are strongly correlated (Vandervelde et al. 2008; Hoitash et al. 2006).
51
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + εi,t (INV3)
where, for firm i and year-quarter t, variable definitions are as follows:
ARi,t = accounts receivables in year-quarter t, scaled by total assets;
ARi,t-4 = accounts receivables in year-quarter t-4, scaled by total assets;
ΔSALESt = sales in year-quarter t minus sales in year-quarter t-4, scaled by total assets;
INVi,t = inventory in year-quarter t, scaled by total assets;
INVi,t-4 = inventory in year-quarter t-4, scaled by total assets; and,
ΔCOGSi,t = cost of goods sold in year-quarter t minus cost of goods sold in year-quarter t-4,
scaled by total assets.
The models above are similar to models (3) and (9) in Hoitash et al. (2006, p.62); however,
in this paper the focus is on predicting abnormal balances for working capital accounts using
income statement accounts as right hand side variables. In Hoitash et al. (2006) the focus is on
predicting income statement accounts. A second important difference is that in this study all
variables are scaled by total assets.
For the two models described above, the quarterly unexpected balance, or forecast error, is
calculated as the difference between quarterly account balances scaled by total assets, and the
expectation resulting from fitting the estimated firm-level parameters to the quarterly observations
from the current-year under audit.
2.2.3 Industry cross-sectional models
Industry cross-sectional models are commonly examined expectation models in the
earnings management literature, aiming to detect earnings management in total accruals (Kothari
52
et al. 2006; Dechow et al. 1995), or account-specific earnings management (Caylor 2010; Stubben
2010; Marquardt and Wiedman 2004). Earnings management studies generally use an expectation
model for total accruals, or for individual components of working capital accruals such as changes
in accounts receivables, to separate expected or ―normal‖ accruals from unexpected or
―discretionary‖ accruals. The discretionary accruals models are estimated annually for a cross-
section of firms within the same industry, commonly defined as all firms within the same two-digit
SIC code. A limitation of using industry cross-sectional models in AAP is that information from all
firms in the cross-section has to be available to the auditor at the time of the audit. This limitation
can be mitigated by using quarterly data for the first two or three quarters of the fiscal year, usually
known to the auditor in the planning and execution phase of the audit.
The cross-sectional approach is consistent with the International Audit Standard 520
recommending using ―similar industry information, such as a comparison of the entity’s ratio of
sales to accounts receivable with industry averages or with other entities of comparable size in the
same industry‖ (IFAC 2009, p. 436). This approach eliminates the need for a long time-series in
estimating the expectation models and ignores the information from prior periods beyond last year;
however, in fast growing or changing industries, information from prior periods may have lower
relevance than cross-sectional information. Moreover, for years immediately following the
adoption of IFRS, the auditor only has the current year IFRS balances and comparatives for one
year prior, thus industry-cross sectional models could be useful in such context.
This study also examines the performance of the accounts receivables and inventory
models described above (AR3 and INV3), estimated for each account and current year-quarter
under audit in cross-section, by industry (denoted as AR4 and INV4). The quarterly unexpected
balance, or forecast error, is calculated as the differences between quarterly account balances
scaled by total assets, and the expectation resulting from fitting the estimated cross-sectional
53
parameters to the firm-level quarterly observations from the current-year under audit.
2.3 PEER INFORMATION IN ACCOUNT-LEVEL EXPECTATION MODELS
2.3.1 Advantages of using peer-firm information in AAP
As explained in Chapter 1, using peer firms as a benchmark is common among
practitioners and researchers. Contemporaneous peer-firm information may improve the
effectiveness of AAP models by providing independent information about the current period’s
conditions, for example, capturing unexpected changes in demand, or increasing competition,
within the industry. In general, the evidence from prior studies in the AAP literature suggests that
historical account-level information needs to be supplemented with relevant contemporaneous data
in order to increase the effectiveness of the expectation models. One way to incorporate
contemporaneous information in the expectation models is by including related contemporaneous
accounts either directly as additional predictors in the model, or indirectly through a structural
model. These approaches require firm-level analysis of interrelations between multiple accounts,
influenced by business complexity, overall uncertainty, and the joint structure of errors involving
related accounts in multiple years (Vandervelde et al. 2008). Other potential useful sources of
contemporaneous data are economic indicators, industry averages (Lev 1980), or comparable peer
firms (Hoitash et al. 2006).
While industry averages or other economic indicators could be useful for the analysis,
closely-matched economically-comparable peer firms are more likely to capture individual
characteristics applicable to each firm. Using matched firms is in line with the requirements of the
International Audit Standard 520, prescribing that the reliability of data should be assessed by the
auditor ―for example, broad industry data may need to be supplemented to be comparable to that
of an entity that produces and sells specialized products‖ (IFAC 2009, p. 438). Furthermore,
54
Trompeter and Wright (2010, p.680) mention that in the current audit environment, ―the access to
benchmarking and competitive data makes the analytical tools so much more powerful‖ and that
auditors ―can get more comparisons with competitors or just other companies in the marketplace‖.
2.3.2 Selection of matching variables and matching approach
Peer firms are matched using the approach previously described in Chapter 1. For a given
fiscal year-end, industry (defined by two-digit SIC code), and size distance (firms that are within a
size distance of 50 percent), firm i is matched to firm j with the most comparable performance,
estimated using stock returns’ covariance over the preceding 48 months, where higher covariance
indicates higher comparability. As per the De Franco et al. (2011) methodology, returns covariance
is measured using the adjusted R2 of the following regression of firm i’s monthly returns on firm
j’s monthly returns31
:
RETURNSi,t = Φi,t+ Φi,t RETURNSj,t+ εi,t (1)
In addition, matched firms have their fiscal year-end on the same month to reduce
differences from timing in financial reporting. Allowing for 50 percent distance in total assets
results in more than one potential control for every treatment observation, and the final selection
among all possible controls is based on returns’ covariance. This procedure is likely to closely
match peer firms deemed economically comparable by the market, and is suitable for research
settings where there are no specified case-control groups. In addition it does not rely on a specific
functional form to predict comparability, beyond a returns covariance structure. Finally, in order to
mitigate any bias resulting from imperfect matching, peer information is introduced as an
additional predictor in the multivariate models, instead of using a peer-differencing approach.
31
The Kendall (1938) Tau or rank correlation coefficient is also reported for the matched peer firms. This non-
parametric statistic measures co-movement, or serial dependence, and can be directly interpreted as the probability of
observing concordant or discordant pairs of observations. Using both Tau and R2 criteria result in similar selection of
peer firms.
55
2.3.3 Firm-level heuristic models including peer-firm information
All heuristic models are also estimated at the firm-level including peer-firm information for
each account, comparing the unexpected balance of each firm i to the unexpected balance of its
peer firm j. Under this approach, peer firms are selected in the last quarter of the fiscal year before
the audit, and the unexpected quarterly receivables and inventory are first calculated for firm i and
its peer firm j independently as explained before (Models AR1, AR2, INV1, and INV2).
Following, the unexpected receivables or inventory of firm j are subtracted from the unexpected
receivables or inventory of firm i. The peer-adjusted difference (denoted AR1P, AR2
P, INV1
P, and
INV2P) is an indicator of abnormal receivables or inventory for a given company beyond normal
fluctuations from a comparable peer firm.
2.3.4 Firm-level time-series models including peer-firm information
The time-series models are also estimated including the lag accounts receivable or
inventory, and the concurrent changes in these two accounts, from a peer firm as additional
predictors in each model.32
Peer firms are selected in the last quarter of the fiscal year before the
audit, and peers stay constant throughout the estimation period and the year under audit (i.e., if
firm j is deemed to be the closest peer for firm i, the model is estimated for firm i using 16 quarters
of data from firm i and firm j and then the model estimates are fitted to the data from the year
under audit). The time-series models including peer-firm information are as follows:
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + β3jARi,t-4 + β4jΔSALESi,t + εi,t (AR3P)
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + γ3jINVi,t-4 + γ4ΔCOGSi,t + εi,t (INV3P)
32
Including peer-firm information as an additional predictor may be preferable to a peer-differences estimate (i.e.,
subtracting from firm i’s prediction error the corresponding peer firm j’s error) because this approach does not assume
that the matching results in firms with identical drivers for changes in accounts receivables or inventory.
56
where, for firm i and year-quarter t:
jARi,t = accounts receivables from closest peer firm in year-quarter t, scaled by total assets;
jARi,t-4 = accounts receivables from closest peer firm in year-quarter t-4, scaled by total
assets;
jINVi,t = inventory from closest peer firm in year-quarter t, scaled by total assets; and,
jINVi,t-4 = inventory from closest peer firm in year-quarter t-4, scaled by total assets.
2.3.5 Industry cross-sectional models including peer-firm information
The industry cross-sectional models (AR4 and INV4) are also estimated including the lag
accounts receivables or inventory, and the concurrent changes in these two accounts, from a peer
firm as additional predictors in each model. The cross-sectional models including peers are
denoted AR4P and INV4
P, and peer firms are selected in the last quarter of the prior fiscal year
before the audit.
2.4 ASSESSING THE SPECIFICATION AND POWER OF EACH MODEL
2.4.1 Firm-level error measures
Let Yi,t denote a firm-quarter observation at time t and Ei,t the expectation from a given
model at the same firm-quarter t. Then, define the difference between actual and expected amount
as model prediction error ei,t = Yi,t – Ei,t. The expectation model can be used to compute out-of-
sample expectations En+1 ,..., En+m based on historical data from past observations t = 1,..., n.
Alternatively, the expectation model can be used to compute expectations Ei,t ,..., Ej,t based on
cross-sectional data for all firms in a given industry and quarter t.
A main objective of this study is to compare the accuracy or specification of each model in
predicting actual accounts receivables and inventory balances, as a percentage of total assets, when
57
there is no error or manipulation. Under this scenario, the model with the lowest prediction error
for a given firm or sample of firms is a better expectation model. The forecasting literature
(Hyndman and Koehler 2006; Makridakis et al. 1998) proposes to use the mean (or median) of
either signed or absolute non-scaled errors if the scale is invariant across observations and
datasets.33
Given that all observations entering the models are scaled by total assets, the errors are
expressed as a percentage of total assets and are directly comparable across firms, thus this study
uses the following accuracy measures:
Mean Absolute Error (MAE) = mean(|ei,t|)
Median Absolute Error (MdAE) = median(|ei,t|)
Mean Error (ME) = mean(ei,t)
Median Error (MdE) = median(ei,t)
To compute the above accuracy measures at the firm-year level, the forecast errors are
averaged across quarters for the year under audit. The statistics in Tables 13 to 15 are calculated
averaging the first two quarters of the year under audit, information that could be available for the
auditor during the planning phase. In addition, Tables 13 to 15 present results averaging all four
quarters of the year under audit for comparative purposes. Similar inferences to those shown in the
Tables 13 to 15 are obtained by averaging the first three quarters.
Turning to the power to detect manipulation, this study simulates two common errors,
suggested by extant studies in the AAP literature (Leitch and Chen 2003, p. 153; Chen and Leitch
1999, p. 44): fictitious sales resulting in overstatement of accounts receivables and sales, and cut-
33
Prior studies in the AAP literature that use non-scaled account balances use scaled errors to aggregate across firms,
dividing the forecast error by the account balance, and compute the Mean Absolute Prediction Error (MAPE) instead
of the Mean Absolute Error MAE.
58
off error resulting in understatement of inventory and overstatement of cost of goods sold.34
In the
main analyses, manipulation is equal to one percent of total assets in these four related accounts,
increasing receivables and sales, and decreasing and increasing inventory and cost of goods sold
respectively by the same amount. In the presence of simulated errors, the model that is the closest
to the seeded manipulation for a given firm or sample of firms is the one most effective for AAP,
and thus this study uses the following measures of detection power:
Mean Error (ME) = mean(ei,t)
Median Error (MdE) = median(ei,t)
In addition, the standard deviation of mean errors is also a useful statistic. Even when two
models have similar mean error, the model that produces estimates with a lower standard deviation
is more likely to detect error or manipulation by better isolating noise from true manipulation.
2.4.2 Tests of statistical specification and power using a simulation procedure
Each expectation model has different sample sizes and the expectations for each firm and
industry are correlated in cross-section and over time. In order to test the statistical specification
across expectation models for a large sample of firms, a simulation procedure is performed for
each model, selecting 250 samples of 100 firms from the full sample. The MPE is calculated for
each of the 250 randomly selected samples and across all samples. Next, the statistical significance
of the MPE is assessed using a two-tailed t-test. Finally, the percentage of samples in which the
null of no manipulation is rejected is computed for all models. The t-test is defined as the equal-
weighted sample MPE divided by its standard error and assumes cross-sectional independence in
the estimated MPE of the sample firms, given that samples are randomly selected across time
periods and industries. Tables 13 to 15 report the mean and standard error of the 250 samples, as
34
This study also examines the specification and power of all models using manipulation in opposite direction. The
results of these analyses are described in Section 2.6.4.
59
well as the percent of the 250 times that the two-tailed t-test rejects the null hypothesis of no
manipulation. Using an expected rejection rate of five percent when there is no seeded error, and
based on a 95 percent confidence interval, an actual rejection rate above five percent indicates that
the expectation model is misspecified.
All models proposed in this study are estimated at the quarter-level and then averaged by
firm to compute a firm-level measure of manipulation using two, three or four quarters of errors by
firm. In order to test the statistical power of each approach to detect errors, directional
manipulation in accounts receivables or inventory is induced in each quarter of the year under
audit. Following, the error is averaged across quarters by firm to calculate a firm-year measure.
Finally, the same simulation procedure explained above is performed for the samples with
manipulation.
This study provides evidence of the detection power of each model by comparing the
estimates of each model against a known quantity of manipulation. In this setting, the best measure
is the one with the highest probability of rejecting the null of no manipulation, and ideally the
rejection rate should be 100 percent (This simulation steps are similar to those in Stubben 2010
and Kothari et al. 2005. See also Appendix F for a summary of the simulation procedure).
2.4.3 Sample selection
The initial sample consists of all U.S. public firms from COMPUSTAT for the years 1988
to 2008 with available quarterly data. Firms without four quarters of data or SIC codes, and zero
total assets, accounts receivable, inventory, sales and cost of goods sold in a given fiscal year are
deleted. All firms in the sample have both accounts receivable and inventory data for two
consecutive years; this eliminates financial and other firms without inventory. The full sample of
firms used to estimate the heuristic models has 53,215 firm-year observations, and this number is
60
reduced to 33,152 observations with at least one peer firm.35
The full sample of firms used to
estimate the time-series models has 47,235 firm-year observations, and this number is reduced to
38,510 observations with at least one peer firm. The full sample of firms used to estimate the
cross-sectional models has 72,074 firm-year observations, and this number is reduced to 38,959
observations with at least one peer firm.36
2.5 RESULTS
2.5.1 Descriptive statistics for firm-level heuristic models samples
Table 11 presents the descriptive statistics for both the full sample (Panel A) and the
sample with peer firms (Panel B) used to assess the heuristic models. In both samples, the mean
accounts receivables (ARt) is 19 percent of total assets, and the median is 16 percent of total assets,
these mean and median percentages are similar to those in the same quarter of the prior year (ARt-
4). The mean inventory (INVt) is 16 percent of total assets, and the median is 13 percent of total
assets, these mean and median percentages are similar to those in the same quarter of the prior year
(INVt-4). The mean sales (SALESt) are 32 percent of total assets, and the median is 28 percent of
total assets. The mean cost of goods sold (COGSt) is 22 percent of total assets, and the median is
18 percent of total assets. Annual growth, measured as the ratio of current year-quarter balances to
the balances in the same quarter in the prior year, has a mean of 11 percent and a median close to
of 7 percent for sales, and a mean of 12 percent and a median of 7 percent for cost of goods sold.
The descriptive statistics for the firms in the sample highlight three characteristics of the
data. First, accounts receivables and inventory are on average both material accounts as a
percentage of assets. Second, accounts receivables and inventory are reasonably stable over time
35
In addition, in the sample used to estimate the heuristic models, firms with changes in sales or cost of goods sold
(CHCOGSi,t or CHSALESi,t) above 5.0 or below 0.2 are deleted to mitigate the large influence that outliers in these
ratios have on the firm-level expectations. 36
In addition, to estimate the model in cross-section, industries with less than 10 observations are excluded from the
sample.
61
as a percentage of assets. Third, on average sales and cost of goods sold increase over time,
making it important to account for firm growth in expectation models for these accounts.
Examining the descriptive statistics for the matched peer firms in Table 11 (Panel B), peer
firms are very close in total assets (jATt), have a mean R2 from the pair-wise returns regression of
16 percent (PEER-R2) and a mean Kendall’s tau of 17 percent (PEER-TAU), and overall exhibit
very similar characteristics in all variables of interest.
Table 12 presents the pair-wise correlation for the sample with peer firms. The first column
shows that accounts receivables (ARt) is strongly correlated with same-quarter last-year accounts
receivables (ARt-4 correlation =0.952), sales (SALESt correlation =0.126), growth in sales
(CHSALESt =0.016), peer accounts receivables in the same-quarter last-year (jARt-4 correlation
=0.591), and peer sales (jSALESt correlation = -0.043). Similarly, the third column shows that
inventory (INVt) is strongly correlated with same-quarter last-year inventory (INVt-4 correlation
=0.958), cost of goods sold (COGSt correlation = 0.374), growth in cost of goods sold (CHCOGSt
= -0.036), peer inventory in the same-quarter last-year (jARt-4 correlation =0.517), and peer cost of
goods sold (jCOGSt = 0.337).
The descriptive statistics (untabulated) of the time-series and industry-cross sectional full
and matched-peers samples are similar compared to the heuristic full sample, with small variations
mostly due to differences in the data requirements.
2.5.2 Specification of firm-level heuristic models
Table 13 presents the main analyses of statistical specification and power of the heuristic
models. AR1 is the receivables model using the incremental approach and AR2 is the receivables
model using the historical approach. Similarly, INV1 is the inventory model using the incremental
approach and INV2 is the inventory model using the historical approach.
62
The top four rows (AR1-4qtr to INV2-4qtr) of Table 13 show the results for each heuristic
expectation model without peers, using the average error of four quarters for each firm. The
following four rows (AR1P-4qtr to INV2
P-4qtr) show the results for each heuristic expectation
model with one peer for each audit client, using the average error of four quarters for each firm.
Similarly the bottom eight rows show the results, with and without peers, using the average error
of two quarters for each audit client. The statistics shown in Tables 13 to 15 are calculated using
250 random samples of 100 firms.
Columns (I) and (II) of Table 13 show the mean and median absolute error (MAE and
MdAE respectively) for the samples without manipulation. In the samples without manipulation,
the model that predicts the smallest absolute mean or median errors is the best model. Examining
the models without peers, the historical expectations based on 12 quarters averages for both
accounts receivables and inventory (AR2 and INV2) are better specified, exhibiting smaller mean
and median absolute errors, than the incremental models (AR1 and INV1). For example, rows one
and two of Column (I) show that the mean absolute error of the incremental model AR1-4qtr
(0.051) is larger than the mean absolute error of the historical model AR1-4qtr (0.037). This is also
the case for the peer-adjusted expectation models, where the historical expectations (AR2P and
INV2P) are better specified than the incremental models (AR1
P and INV1
P). For example, rows
five and six of Column (I) show that the absolute mean error of the incremental model with peer
information AR1P-4qtr (0.064) is larger than the mean error of the historical model with peer
information AR2P-4qtr (0.047). Moreover, the absolute error of the heuristic models is larger for
the models with peers than for the models without peers (e.g., AR1-4qtr = 0.051 and AR1P-4qtr
=0.064; AR2-4qtr = 0.037 and AR1P-4qtr =0.047).
Columns (III), (IV) and (V) of Table 13 show the mean, median and standard deviation of
signed errors for the samples without manipulation. Examining the models without peers, the
63
historical expectations for both accounts receivables and inventory (AR2 and INV2) are better
specified, exhibiting mean and median errors that are closer to zero, than the incremental models
(AR1 and INV1). This is also the case for peer-adjusted expectation models, where the historical
expectations (AR2P and INV2
P) are better specified than the incremental models (AR1
P and
INV1P). By calculating peer-adjusted expectations, the mean and median error is closer to zero
compared to the models without peers. Columns III and IV rows five eight are closer to zero than
rows one to four.
Columns (VI) and (VII) of Table 13 show the specification error of each model for the
samples without manipulation. Specification error is calculated as the percentage of the 250
random samples of 100 firms in which the mean of year-firm-level errors is statistically greater
and lower than zero at five percent level, rejecting the null of no manipulation. Column (VI) shows
that the models without peers (AR1 to INV2) reject the null of no manipulation more than five
percent of the samples on the negative side. For example, the first four rows of Column (VI) show
that the null of no manipulation is rejected too often by the models without peers, in 88.4 percent
of the of the samples by the receivables incremental model AR1-4qtr, and in 6.0 percent of the
samples by the receivables historical model AR2-4qtr. Similarly, the null is rejected in 80.4
percent of the samples by the inventory incremental model INV1-4qtr, and in 11.6 percent of the
samples by the inventory historical model INV2-4qtr. In contrast, rows five to eight of Columns
(VI) and (VII), show that the models with peer information reject the null of no manipulation in
less than five percent of the samples, providing evidence that using information from peer firms
improves the specification of the heuristic models.
2.5.3 Detection power of firm-level heuristic models
Columns (VIII), (IX) and (X) of Table 13 show summary statistics for the samples with
64
simulated directional manipulation, increasing accounts receivable and decreasing inventory by
one percent of total assets. The mean error in these samples with directional manipulation should
be close to 0.010 for the receivables models and close to -0.010 for the inventory models.
Examining the models without peers, the historical expectations for both accounts receivables and
inventory (AR2 and INV2) exhibit mean and median errors closer to the seeded errors than the
incremental models (AR1 and INV1). For example, rows one and two of Column (VIII) show that
the mean error of the incremental model AR1-4qtr (-0.034) is further from the expected 0.010 than
the mean error of the historical model AR1-4qtr (0.008). This is also the case for the peer-adjusted
expectation models, where the historical expectations (AR2P and INV2
P) are better specified than
the incremental models (AR1P and INV1
P). For example, rows five and six of Column (VII) show
that the absolute mean error of the incremental model with peer information AR1P-4qtr (-0.012) is
further from the expected 0.010 than the mean error of the historical model with peer information
AR1P-4qtr (0.010).
Column (XI) of Table 13 shows the detection power of each model for the samples with
simulated directional manipulation. Detection Power is calculated as the percentage of the 250
random samples of 100 firms in which the mean of year-firm-level errors is statistically greater or
lower than zero at five percent level in the direction of the manipulation, rejecting the null of no
manipulation. The first eight rows of Column (XI) show that using peer information does not
improve the detection power of the heuristic models. For example, rows one and five of Column
(VI) show that the null of no manipulation is not rejected by the receivables incremental model
AR1-4qtr, and that the null is only rejected in 0.4 percent of the samples by the receivables
incremental model using information from peer firms AR1P-4qtr. Similarly, rows two and six of
Column (VI) show that the null of no manipulation is rejected by the receivables historical model
AR2-4qtr in 39.2 percent of the samples, and by the receivables historical model using information
65
from peer firms AR2P-4qtr in 35.2 percent of the samples.
The combined analyses in Table 13 indicate that one-period incremental heuristic models
without peer information are misspecified (rejecting the null of no manipulation too often) and that
peer-adjusted heuristic models might correct this issue but take away the detection power of these
models. The simulation sample mean for the models without peers is negative too often (e.g.,
Column VI rows one and four is 88.4 and 80.4 percent) due to outliers in each simulation. The
expected increase in accounts receivable or inventory is too large compared to the actual increase
in these accounts for firms that experience a large increase in sales or cost of goods sold. A general
problem with the heuristic models is that they assume that accounts receivables or inventory
should experience a one-time change equal to the percentage change in sales or cost of goods sold
in this year-quarter compared to the same quarter in the previous year. The models based on
historical averages with peers perform better than the one-period models, detecting accounts
receivable manipulation in over 35 percent of the simulated samples and inventory manipulation in
over 45 percent of the samples (e.g., Column XI, rows six and eight, 35.2 and 45.2 percent).
Overall, without considering additional firm characteristics, more than one period, or additional
information from peer firms, account-level heuristic models seem to be generally unreliable.
2.5.4 Specification and detection power of firm-level time-series models
Table 14 presents the main analyses of statistical specification and power of the firm-level
time-series models. AR3 is the receivables model estimated in time-series, using the 16 quarters
prior to the year under audit and fitted to each quarter of the year under audit. INV3 is the
inventory model estimated in time-series, using the 16 quarters prior to the year under audit and
fitted to each quarter of the year under audit.
The first four rows of Table 14 show firm-year statistics calculated averaging errors across
66
all four quarters of the fiscal year for the accounts receivables and inventory models. The first two
rows (AR3-4qtr and INV3-4qtr) show the results for each model without peer information. The
following two rows (AR3P-4qtr and INV3
P-4qtr) show the results for each model including peer
data. Similarly the bottom four rows show the results averaging errors across the first two quarters
of the fiscal year, with and without peer information.
Columns (I) to (V) of Table 14 show the mean, median and standard deviation of absolute
and signed errors for the samples without manipulation. In general, the models with peer
information have very similar absolute and signed error statistics, compared to the models without
peer information. For example, rows one and four of Column (I) show that the mean absolute error
of the receivables model AR3-4qtr (0.041) is very close to the mean absolute error of the
receivables model with peer information AR3P-4qtr (0.042). Similarly rows one and four of
Column (III) show that the mean error of the receivables model AR3-4qtr (-0.002) is very close to
the mean error of the receivables model with peer information AR3P-4qtr (0.042).
Columns (VI) and (VII) of Table 14 show the specification error of each model for the
samples without manipulation. The models with peer information are better specified, showing
lower bias than the models without peer information. The second row of Column (VI) shows that
the inventory model without peers INV3-4qtr rejects the null of no manipulation more than five
percent of the samples on the negative side. In contrast, rows three and four of Columns (VI) and
(VII) show that neither the receivables nor the inventory model with peers (AR3P-4qtr and INV3
P-
4qtr) reject the null of no manipulation more than five percent of the samples.
Finally, Column (XI) of Table 14 shows the detection power of each model for the samples
with simulated directional manipulation. The models with peer information have a higher
detection power. For example, including peer information increases detection power from 16.4 to
67
22 percent for the four-quarter receivables model, and from 54.8 to 58.8 percent for the four-
quarter inventory model. Overall, the time-series models with perform much better than the
heuristic models, and using peer information increases the effectiveness of time-series models.
2.5.5 Specification and detection power of industry cross-sectional models
Table 15 presents the main analyses of statistical specification and power of the industry
cross-sectional models. AR4 is the receivables model estimated in cross-section by industry for
each quarter of the year under audit. INV4 is the receivables model estimated in cross-section by
industry for each quarter of the year under audit.
The first four rows of Table 15 show the statistics calculated averaging errors across all
four quarters of the fiscal year for the accounts receivables and inventory models. The top two
rows (AR4-4qtr and INV4-4qtr) show the results for each model without peer data. The following
two rows (AR4P-4qtr and INV4
P-4qtr) show the results for each model including peer data.
Similarly the bottom four rows show the results averaging errors across the first two quarters of
the fiscal year, with and without peer data.
Columns (I) to (V) of Table 15 show the mean, median and standard deviation of absolute
and signed errors for the samples without manipulation. The models with peer information have
lower absolute error statistics, but similar signed error mean, compared to the models without peer
information. For example, rows one and four of Column (I) show that the mean absolute error of
the receivables model AR4-4qtr (0.051) is higher than the mean absolute error of the receivables
model with peer information AR4P-4qtr (0.027). Similarly rows one and four of Column (III) show
that the mean error of the receivables model AR4-4qtr and the mean error of the receivables model
with peer information AR3P-4qtr are very close to zero. In addition, rows one and four of Column
(V) show that the errors from models with peer information have lower standard deviations,
68
suggesting that these models are more precise than the ones without peer information.
Columns (VI) and (VII) of Table 15 show the specification error of each model for the
samples without manipulation. All models, with and without peer information are well specified,
rejecting the null of no manipulation in less than five percent of the samples.
For example, including peer information increases detection power from 31.6 to 80
percent for the four-quarter receivables model, and from 36 to 91.2 percent for the four-quarter
inventory model. Finally, Column (XI) of Table 15 shows the detection power of each model for
the samples with simulated directional manipulation. The models with peer information have a
noticeable higher detection power. For example, including peer information increases detection
power from 31.6 to 80 percent for the four-quarter receivables model, and from 36 to 91.2 percent
for the four-quarter inventory model.37
The analyses in Table 15 indicate, first that industry-cross sectional models can be equally
or better specified than the time-series models; second, that the industry cross-sectional models
with peer information perform better than the heuristic and time-series models in predicting
accounts receivables, with and without manipulation; and third, that using peer information
increases the effectiveness of industry cross-sectional models.
The time-series and the cross-sectional models are conceptually different approaches to
identify potential manipulation. Time-series models use data from an estimation period during
which no systematic manipulation is expected to occur. In addition, time-series models may
provide unreliable estimates when the number of observations used in estimating the parameters is
small and are less sensitive to current changes in industry characteristics.
Conversely, cross-sectional models assume no systematic manipulation across firms in an
37
The power to detect seeded errors in these cross-sectional models stands in contrast to the 24.0 percent detection
power of the discretionary revenue model, the 7.2 percent detection power of the modified-Jones model, and the six
percent detection power of the performance-matched modified Jones model, documented in Stubben (2010 p.707).
69
industry, and implicitly assume that the model parameters are the same across firms. The abnormal
receivables or inventory estimated from these models can be interpreted as ―industry-relative‖
manipulation. The cross-sectional models have the advantage of requiring less historical data, but
these models do not generate firm-specific coefficients. A drawback of the cross-sectional models
is that they may indicate potential manipulation when a given firm has different receivables or
inventory policies than the average firm in the industry. By including information from an
economically-comparable peer for each firm as an additional predictor in the model, each firm is
not only benchmarked against the industry, but also against its peer. This results in noticeable
improvements in detection power. This finding may be of interest to earnings management
researchers looking for ways to improve the detection power of cross-sectional models similar to
the Jones (1991) model.
2.5.6 Idiosyncratic error in AAP expectation models
A limitation of the AAP expectation models examined in this study is their high
idiosyncratic error in a one-client, one-period audit setting.38
The simulation results discussed in
the previous Sections, using repeated samples of 100 firms each, cancel out idiosyncratic noise. In
practice, the auditor estimates the expectation model and observes a single forecast error
calculated by comparing actual versus expected account balances. Next, the auditor needs to use
judgement to determine the necessary follow-up audit procedures, for example, an auditor can use
the forecast error to allocate audit hours for subsequent substantive testing. A higher forecast error
should indicate to the auditor that additional effort in investigating and explaining differences is
required. In this context, a materiality threshold can be used in later stages of the audit to judge the
38
In order to assess how sensitive are the main results of this paper to idiosyncratic error in smaller samples, The
simulation procedure is repeated drawing 20 observations instead of 100 observations for each of the 250 repetitions.
The detection power of all models is lower than the one presented in Tables 13 to 15; however, there is still a
significant improvement in the effectiveness of the models by including information from peer firms.
70
significance of any errors uncovered by substantive procedures in combination with the evidence
from the preliminary AAP.
Audit firms may reduce their overall risk if they consistently apply a more effective
methodology for their portfolio of audit clients. The simulation results may be interpreted as a
portfolio approach to test the effectiveness of AAP models, with and without peers, to detect high
risk areas and direct the auditor’s attention towards potential misstatements.
2.6 ADDITIONAL ANALYSES
2.6.1 Increasing the number of peer firms
To verify whether there is incremental benefit from using more than one peer firm for each
audit client, in terms of accuracy, specification and power, the firm-level time-series and industry
cross-sectional models are also estimated for the sub-sample of audit clients that have at least five
peers. The lag accounts receivable or inventory, and the concurrent changes in these two accounts,
from up to five peers is included in the prediction models (jARi,t-4 and jΔSALESi,t, or jINVi,t-4 and
jΔCOGSi,t for j =1,..,5), and the specification and power are compared across models.
In general, including more than one peer firm does not improve the models significantly.
There are small improvements only for the industry cross-sectional models, in terms of absolute
error and detection power, from having up to four peers in the model; however, using more than
four peers reduces the effectiveness of the prediction models in all cases. Increasing the number of
peers improves the power and specification of the expectation models due to the reduction of
idiosyncratic noise introduced by selecting peers that are not perfectly comparable; however, as
the number of peer firms increases, the comparability between peers also decreases. These results
suggest that auditors may benefit from focusing on finding a small number of closely-matched
peer firms in performing AAP, and that there is a trade-off between the number of peers and the
71
comparability between peers.
2.6.2 Using industry averages instead of matched peer firms
The time-series and industry cross-sectional models may be estimated using the year-
quarter t industry cross-sectional average by two-digit SIC code, and year-quarter t-4 average of
accounts receivables or inventory scaled by total assets at the end of the quarter (IndustryARi,t-4
and IndustryΔSALESi,t, or IndustryINVi,t-4 and IndustryΔCOGSi,t). The models with industry
averages represent an improvement over the models without industry averages; however, the
models with peer firms outperform the models with industry averages.39
2.6.3 Changing the structure of peer-firm information in the expectation models
The time-series and industry cross-sectional models may be estimated using a different
structure for the peer information. First, including only accounts receivables or inventory from a
peer firm, in year-quarter t, scaled by total assets (jARi,t or jINVi,t):
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + β3jARi,t + εi,t (AR5P)
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + γ3jINVi,t + εi,t (INV5P)
and second, including only accounts receivables or inventory from a peer firm, in year-quarter t
and t-4, scaled by total assets (jARi,t and jINVi,t or jARi,t-4 and jINVi,t-4):
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + β3jARi,t + β4jARi,t-4 + εi,t (AR6P)
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + γ3jINVi,t + γ3jINVi,t-4 + εi,t (INV6P)
These variations result in similar results to those documented in Tables 13 to 15; however,
39
This model is similar to the Dechow and Sloan (1991) industry model in the earnings management literature. The
industry model assumes that variation in discretionary accruals is common across firms in the same industry.
72
the main specification used in Tables 13 to 15, including the lagged terms and concurrent changes
in related accounts from a peer firm, is more consistent with the specification of the model without
peers.
2.6.4 Simulating manipulation in opposite direction and varying the amount of manipulation
The main results presented in Tables 13 to 15 are based on directional manipulation equal
to one percent of total assets, simulating fictitious sales resulting in overstatement of accounts
receivables and sales, and cut-off error resulting in understatement of inventory and overstatement
of cost of goods sold. All models are also estimated using the opposite direction. Simulating erros
in the opposite direction produces qualitatively similar results, in terms of accuracy, specification
and power, to those documented in Tables 13 to 15. These results suggests that the models
examined here may be used to detect both under and overstatement of accounts receivables and
inventory.
A commonly used materiality threshold by auditors is half a percent of assets. All models
are also estimated simulating directional manipulation equal to half percent of total assets. This
produces qualitatively similar results, in terms of accuracy, specification and power, to those
documented in Tables 13 to 15.
2.7 CONCLUSION
Auditors conduct AAP in three steps, first they set an expectation for a particular financial
statement account or ratio, second they compare the difference between actual and expected
numbers, and finally they perform additional audit procedures to explain any unexpected
differences. In practice, the auditor can estimate a forecast error for a given audit client by
comparing actual account balances to the expectation generated using a prediction model. A higher
forecast error should indicate to the auditor that additional effort in investigating and explaining
73
differences is required. This Chapter compares the effectiveness of three main types of models to
estimate account-level expectations for analytical auditing purposes, with and without including
information from economically-comparable peer firms: heuristic, time-series, and industry cross-
sectional models.
The expectation models examined in this Chapter share some similarities with the models
proposed in extant earnings management studies. For example, discretionary accruals models are
estimated in time-series or in cross-section using year-end data for a sample of firms available to
the researcher; however, in the context of AAP, the auditor faces a single-client situation where
year-end financial statements for the client and its potential peers are usually not available in the
early phases of the audit process. The methodology examined in this Chapter is mostly applicable
during the planning phase of the audit, assuming that the auditor has access to two or three
quarters of publicly available financial information for a client and its potential peer firms.
This study documents that information from economically-comparable peer firms improves
the specification of all models, and improves the detection power of time-series and industry
cross-sectional models. Comparing between models, one-period incremental heuristic models are
generally unreliable, and industry cross-sectional models can be more effective than previously
proposed time-series models.
The simulation results in this study, seeding directional account-level errors equal to one
percent of total assets, show that the time-series models, including information from peer firms,
detect manipulation in accounts receivables in approximately 22 percent of the simulated samples,
and in inventory in approximately 59 percent of the simulated samples. In contrast, industry cross-
sectional models detect manipulation in accounts receivables in 80 percent of the simulated
samples, and in inventory in over 91 percent of the simulated samples. The cross-sectional models
have the advantage of requiring less historical data, but these models do not generate firm-specific
74
coefficients. These models may indicate that there is manipulation when a given firm has different
receivables or inventory policies than the average firm in the industry. By including information
from an economically-comparable peer for each firm as an additional predictor in the model, each
firm is not only benchmarked against the industry, but also against its peer. This results in
noticeable improvements in detection power. Finally, the results from including multiple peers in
the analysis suggest that auditors may benefit from focusing on finding a small number of closely-
matched peer firms in performing AAP. A limitation of the AAP expectation models examined in
this study is their high idiosyncratic error in a one-client, one-period audit setting and the auditor
has to exercise judgement in assessing the forecast error produced by each expectation model.
Finally, auditors may use the methodology suggested in this study to select peers in order to
benchmark other information such as the client’s strategy, managerial compensation packages,
corporate governance practices, etc.
In summary, the findings of this study may be helpful for auditors of public companies,
earnings management researchers, regulators, financial analysts and other stakeholders in selecting
expectation models and finding comparable peer firms to assess the reasonability of a company’s
financial information at the account level.
75
REFERENCES
Albuquerque, A. 2009. Peer firms in relative performance evaluation. Journal of Accounting and
Economics 48 (1): 69–89.
Altman, E. 1983. Corporate financial distress: A complete guide to predicting, avoiding, and
dealing with bankruptcy. Wiley, New York.
American Institute of Certified Public Accountants (AICPA). 1988. Analytical Procedures. AU 329
Statement of Auditing Standards No. 56. New York, NY: AICPA.
Angrist, J., and J. Hahn. 2004. When to control for covariates? Panel asymptotics for estimates of
treatment effects. Review of Economics and Statistics 86 (1): 58–72.
Ball, R., and L. Shivakumar. 2006. The role of accruals in asymmetrically timely gain and loss
recognition. Journal of Accounting Research 44 (2): 207–242.
Balsam, S., J. Krishnan, and J. Yang. 2003. Auditor industry specialization and earnings quality.
Auditing: A Journal of Practice and Theory 22 (2): 71–97.
Barber, B., and J. Lyon. 1997. Detecting long-run abnormal stock returns: the empirical power and
specification of test statistics. Journal of Financial Economics 43 (3): 341–372.
Beneish, M. 1999. The detection of earnings manipulation. Financial Analysts Journal 55 (5): 24–
36.
Bhojraj, S., C. Lee, and R. Sloan. 2002. Who is my peer? A valuation-based approach to the
selection of comparable firms. Journal of Accounting Research 40 (2): 407–439.
Blouin, J., B. Murray Grein, and B. Roundtree. 2007. An analysis of forced auditor change: The
case of former Arthur Andersen clients. The Accounting Review 82 (3): 621–650.
Boone, J., I. Khurana, and K. Raman. 2010. Do the Big 4 and the second-tier firms provide audits
of similar quality? Journal of Accounting and Public Policy (Forthcoming).
Brookhart, M., S. Schneeweiss, K. Rothman, R. Glynn, J. Avorn, and T. Sturmer. 2006. Variable
selection for propensity score models. American Journal of Epidemiology 163 (12): 1149–
1156.
Byrnes, N. ―Deloitte Consulting is Hiring.‖ Bloomberg BusinessWeek, February 2010.
http://www.businessweek.com/managing/content/feb2010/ca20100211_911815.htm
Callen, J., S. Robb, and D. Segal. 2008. Revenue manipulation and restatements by loss firms.
Auditing: A Journal of Practice and Theory 27 (2): 1–29.
———, and M. Morel. 2003. The Enron-Andersen debacle: Do equity markets react to auditor
reputation? Finance Letters: (October): 1–5.
Cahan, S., and W. Zhang. 2006. After Enron: Auditor conservatism and ex-Andersen clients. The
Accounting Review 81 (1): 49–82.
Carson, E., and N. Fargher. 2007. Note on audit fee premiums to client size and industry
specialization. Accounting and Finance 47 (1): 423 –446.
Caylor, R. 2009. Strategic revenue recognition to achieve earnings benchmarks. Journal of
76
Accounting and Public Policy 29 (1): 82–95.
Chan, L., J. Lakonishok, and B. Swaminathan. 2007. Industry classifications and return
comovement. Financial Analyst Journal 63 (6): 56–70.
Chaney, P., and K. Philipich. 2002. Shredded reputation: the cost of audit failure. Journal of
Accounting Research 40 (4): 1221–1245.
Chen, Y., and R. A. Leitch. 1999. An analysis of the relative power characteristics of analytical
procedures. Auditing: A Journal of Practice & Theory 18 (2): 35–69.
Cram, D., V. Karan, and I. Stuart. 2009. Three threats to validity of choice-based and matched
sample studies in accounting research. Contemporary Accounting Research 26 (2): 477–
516.
DeAngelo, L. 1986. Accounting numbers as market valuation substitute: A study of management
buyouts of public stockholders. The Accounting Review. 61 (3): 400–420.
Dechow, P., and R. Sloan. 1991. Executive incentives and the horizon problem: An empirical
investigation. Journal of Accounting and Economics 14 (1): 51–89.
———, R. Sloan, and A. Sweeney. 1995. Detecting earnings management. The Accounting Review
70 (2): 193–225.
———, W. Ge, C. Larson, and R. Sloan. 2010. Predicting material accounting misstatements.
Contemporary Accounting Research 28 (1): 17–82.
DeFond, M., J. Francis, and T.J. Wong. 2000. Auditor industry specialization and market
segmentation: Evidence from Hong Kong. Auditing: A Journal of Theory and Practice 19
(1): 49–66.
De Franco, G., S. P. Kothari, and R. Verdi. 2011. The benefits of financial statement comparability.
Journal of Accounting Research (Forthcoming)
Dehejia, R., and S. Wahba. 2002. Propensity score matching methods for non-experimental causal
studies. Review of Economics and Statistics 84 (1): 151–161.
Dunn, K., and B. Mayhew. 2004. Audit firm industry specialization and client disclosure quality.
Review of Accounting Studies 9 (1): 35–58.
Financial Accounting Standards Board (FASB). 1987. Statement of Cash Flows. Statement of
Financial Accounting Standards No. 95. Norwalk, CT: FASB.
Francis, J., K. Reichelt, and D. Wang. 2005. The pricing of national and city-specific reputations
for industry expertise in the U.S. audit market. The Accounting Review 80 (1): 113–136.
, C. Lennox, and Z. Wang. 2010. Selection models in accounting research. Working Paper,
University of Missouri-Columbia, Hong Kong University of Science and Technology, and
Nanyang Technological University.
Gauntt, J., and G. Glezen. ―Analytical auditing procedures.‖ The Internal Auditor, February 1997,
56–60.
Geiger, M., and D. Rama. 2003. Audit fees, nonaudit fees, and auditor reporting on stressed
companies. Auditing: A Journal of Practice & Theory 22 (2): 53–69.
Glazerman, S., D. Levy, and D. Myers. 2003. Nonexperimental versus experimental estimates of
77
earnings impacts. Annals of the American Academy of Political and Social Science 589 (1):
63–93.
Government Accountability Office (GAO). 2008. Audits of public companies. Continued
concentration in audit market for large public companies does not call for immediate
action. Washington, DC: GAO.
Hahn, J. 1998. On the role of the propensity score in the efficient estimation of average treatment
effects. Econometrica 66 (1): 315–332.
Healy, P. 1985. The effect of bonus schemes on accounting decisions. Journal of Accounting and
Economics 7 (1-3): 85–107.
Heckman, J. 1979. The sample bias as a specification error. Econometrica 47 (1): 153–62.
2005. The scientific model of causality. Sociological Methodology 35 (1):1–97.
, H. Ichimura, and P. Todd 1998. Matching as an econometric evaluation estimator. The Review
of Economic Studies 65 (2): 261–294.
Hirst, D., and L. Koonce. 1996. Audit analytical procedures: A field investigation. Contemporary
Accounting Research 13 (2): 457–486.
Ho, D. E., K. Imai, G. King, and E. Stuart. 2007. Matching as nonparametric preprocessing for
reducing model dependence in parametric causal inference. Political Analysis 15 (3): 199–
236.
Hoitash, R., A. Kogan, and M. Vasarhelyi. 2006. Peer-based approach for analytical procedures.
Auditing: A Journal of Practice and Theory 25 (2): 53–84.
Hribar, P., T. Kravet, and R. Wilson. 2009. A new measure of accounting quality. Working Paper.
University of Iowa and University of Texas at Dallas.
Hyndman, R., and A. Koehler. 2006. Another look at measures of forecast accuracy. International
Journal of Forecasting 22 (1): 679–688.
International Federation of Accountants (IFAC). 2009. Analytical procedures. International
Auditing Standard 520. New York, NY: IFAC.
Imbens, G. W. 2004. Nonparametric estimation of average treatment effects under exogeneity: A
review. Review of Economics and Statistics 86 (1): 4–29.
Jones, J., 1991. Earnings management during import relief investigation. Journal of Accounting
Research 29 (2): 193–228.
Kendall, M. 1938. A new measure of rank correlation. Biometrika 30 (1-2): 81–89.
Kinney, W. 1987. Attention-directing analytical review using accounting rations: A case study.
Auditing: A Journal of Practice & Theory 6 (2): 59–73.
Knechel, W., V. Naiker, and G. Pacheco. 2007a. Does auditor industry specialization matter?
evidence from market reaction to auditor switches. Auditing: A Journal of Practice and
Theory 26 (1): 19–45.
———, S. Salterio, and B. Ballou. 2007b. Auditing: Assurance & Risk, 3e. Mason, OH: Thomson
South-Western.
78
Kothari, S. P., A. Leone, and C. Wasley. 2005. Performance matched discretionary accrual
measures. Journal of Accounting & Economics 39 (1): 163–197.
Krishnan, G. V. 2003. Does big 6 auditor industry expertise constrain earnings management?
Accounting Horizons 17 (Supplement): 1–16.
Lawrence, A., M. Minutti-Meza, and P. Zhang. 2011. Can Big 4 versus non-Big 4 differences in
audit-quality proxies be attributed to client characteristics? The Accounting Review 86 (1):
259–286.
Leitch, R., and Y. Chen. 2003. The effectiveness of expectation models in recognizing error
patterns and generating and eliminating hypotheses while conducting analytical
procedures. Auditing: A Journal of Practice and Theory 22 (2): 147–170.
Lev, B. 1980. On the use of index models in analytical reviews by auditors. Journal of Accounting
Research (2): 524–550.
———, and R. Thiagarajan. 1993. Fundamental information analysis. Journal of Accounting
Research 31 (2): 190–215.
Lim, C.Y., and H.T. Tan. 2008. Non-audit service fees and audit quality: the impact of auditor
specialization. Journal of Accounting Research 46 (1): 199–246.
Low, K.Y. 2004. The effect of industry specialization on audit risk assessments and audit planning
decisions. The Accounting Review 79: 201–209.
Makridakis, S., S. Wheelwright, and R. Hyndman. 1998. Forecasting methods and applications,
3e. New York, NY: John Wiley & Sons, Inc.
Marquardt, C., and C. Wiedman. 2004. How are earnings managed? An examination of specific
accruals. Contemporary Accounting Research 21 (2): 461–491.
McNichols, M. 2002. Discussion of ―The quality of accruals and earnings: the role of accruals
estimation errors.‖ The Accounting Review 77 (Supplement): 61–69.
Melumad, N., and D. Nissim. 2008. Line-item analysis of earnings quality. In Foundations and
Trends in Accounting, edited by S. Reichelstein, 3 (2-3). Boston, MA.
Morgan, S., and D. Harding. 2006. Matching estimators of causal effects: Prospects and pitfalls in
theory and practice. Sociological Methods & Research 35 (1): 3–60.
Nagy, A. 2005 Mandatory audit firm turnover, financial reporting quality, and client bargaining
power: the case of Arthur Andersen. Accounting Horizons 19 (2): 51–68.
Neal, T. and R. Riley, Jr. 2004. Auditor industry specialist research design. Auditing: A Journal of
Practice and Theory 23 (2): 169–177.
Nelson, K., R. Price, and B. Roundtree. 2008. The market reaction to Arthur Andersen's role in the
Enron scandal: Loss of reputation or confounding effects? Journal of Accounting and
Economics 46 (2): 279–293.
Nelson, M., J. Elliot, and R. Tarpley. 2003. How are earnings managed? Examples from auditors.
Accounting Horizons (Supplement): 17–35.
Ng, C. ―Assurance issues.‖ CA Magazine, November 2008, 49–51.
Owhoso, V. E., W. F. Messier, and J. Lynch. 2002. Error detection by industry-specialized teams
79
during the sequential audit review. Journal of Accounting Research 40 (3): 883–900.
Public Oversight Board. 2000. The Panel on Audit Effectiveness Report and Recommendations.
Stamford, CT: Public Oversight Board. Available at: http://www.pobauditpanel.org/
PricewaterhouseCoopers (PwC). 2010. Audit and assurance services. Available at:
http://www.pwc.com/gx/en/audit-services/index.jhtml
Public Company Accounting Oversight Board (PCAOB). 2008. Firm inspection reports. New
York, NY: PCAOB. Available at: http://pcaobus.org/Inspections/Reports/Pages/default.aspx
Reichelt, K., and D. Wang. 2010. National and office-specific measures of auditor industry
expertise and effects on audit quality. Journal of Accounting Research 48 (3): 647–686.
Rosenbaum, P., and D. Rubin. 1983. The central role of the propensity score in observational
studies for causal effects. Biometrika 70 (1): 41–55.
2002. Observational Studies, 2nd Edition. Springer Verlag, New York, NY.
Rubin, D. 1973. The use of matched sample and regression adjustment to reduce bias. Biometrics
29 (1): 318–328.
1979. Using multivariate matched sampling and regression adjustment to control bias in
observational studies. Journal of the American Statistical Association 74 (1): 318–328.
2001. Using propensity scores to help design observational studies: application to the tobacco
litigation. Health Services & Outcomes Research Methodology 2 (1): 169–188.
2006. Matched sampling for causal inference. Cambridge University Press, Cambridge,
England.
, and N. Thomas 2000. Combining propensity score matching with additional adjustments for
prognostic covariates. Journal of the American Statistical Association 95: 573–585.
Solomon, I., M. Shields, and O. Whittington. 1999. What do industry-specialist auditors know?
Journal of Accounting Research 37 (1): 191–208.
Stuart, E. 2010. Matching methods for causal inference: a review and look forward. Statistical
Science 25(1): 1–21.
Stubben, S. 2010. Discretionary revenues as a measure of earnings management. The Accounting
Review 85 (2): 695–717.
Taylor, M. 2000. The effects of industry specialization on auditor’s inherent risk assessments and
confidence judgments. Contemporary Accounting Research 17 (4): 693–712.
Trompeter, G., and A. Wright. 2010. The world has changed –Have analytical procedure practices?
Contemporary Accounting Research 27 (2): 667–700.
Vandervelde, S., Y. Chen, and R. Leitch. 2008. Auditing: A Journal of Practice and Theory 27 (2):
79–107.
Wheeler, S., and K. Pany. 1990. Assessing the performance of analytical procedures: A best case
scenario. The Accounting Review 65 (3): 557–577.
Zhao, Z. 2004. Using matching to estimate treatment effects: data requirements, matching metrics,
and Monte Carlo evidence. Review of Economics and Statistics 86 (1): 91–107.
80
APPENDIX A –Variable Definitions –Chapter 1
ADA = absolute discretionary accruals estimated using the cross-sectional Jones (1991) model,
including ROA as per Kothari et al. (2005), estimated by industry and year;
ADA_FULL = absolute discretionary accruals estimated using the cross-sectional Jones (1991) model,
including ROA (Kothari et al. 2005), cash flows in periods t and t-1 scaled by total assets
(McNichols 2002), and a non-linear interaction term based on the sign of cash flows in
period t (Ball and Shivakumar, 2006);
ADREV = absolute discretionary revenue estimated using the cross-sectional Stubben (2010) model,
estimated by industry and year;
NLEAD1 = ―1‖ for auditors that have the largest market share in a given industry at the U.S. national
level and have more than 10 percent greater market share than the closest competitor, and
―0‖ otherwise;
CLEAD1 = ―1‖ for auditors that have the largest market share in a given industry at the U.S. city level,
where city is defined as a Metropolitan Statistical Area following the 2003 U.S. Census
Bureau MSA definitions, and have more than 10 percent greater market share than the
closest competitor, and ―0‖ otherwise.
BIG4 = ―1‖ if the client has a Big 4 auditor and ―0‖ otherwise;
LOG_MKT = natural logarithm of market value;
LEV = (total liabilities)/average total assets;
ROA = (net income)/average total assets;
ROAL = (net incomet-1) / average total assets t-1;
LOSS = indicator variable equal one if net income is negative, and ―0‖ otherwise;
CFO = (cash flow from operations)/average total assets;
BTM = (book value of equity)/market value of equity;
ABS(ACCRL) = absolute value of (total accrualst-1)/average total assetst-1;
GROWTH = sales growth calculated as (sales – salest-1)/salest-1;
ALTMAN = Altman’s (1983) scores;
STDEARN = standard deviation of income before extraordinary items in the past four years;
TENURE = ―1‖ if the client has kept the same auditor for three or more fiscal years, and ―0‖ otherwise;
YEAR = year fixed effects;
AA = ―1‖ for AA clients and ―0‖ otherwise;
ΔLEAD_AA interaction term between AA and changes between specialist auditors, where ΔLEAD= ―-1‖
for clients that switched in industries where AA was a specialist and the successor is not a
specialist auditor, ΔLEAD= ―1‖ for clients that switched in industries where AA was not a
specialist and the successor is a specialist auditor, and ΔLEAD= ―0‖ for all other cases;
ij = pair-wise difference between matched observations; and,
Δ one-year change in the level of each variable.
81
APPENDIX B – Summary of Discretionary Accruals and Discretionary Revenue Estimates
Jones (1991) discretionary accruals model including ROA:
ACi,t = α + β1ΔRi,t + β2PPEi,t + β3ROAi,t + εi,t (3)
Jones (1991) discretionary accruals model including ROA and other accrual drivers:
ACi,t = α + β1ΔRi,t + β2PPEi,t + β3ROAi,t + β4CFOi,t-1 + β5CFOi,t
+ β6CFOi,t+1 + β7Di,t + β8D×CFOi,t + εjt
(4)
Stubben (2010) discretionary revenue model:
ΔARi,t = α + β1ΔRi,t+ β2ΔRi,t×SIZEi,t + β3ΔRi,t×AGEi,t + β4ΔRi,t×AGE_SQi,t
+ β5ΔRi,t×GRR_Pi,t + β6ΔRi,t×GRR_Ni,t + β7ΔRi,t×GRMi,t
+ β8ΔRi,t×GRM_SQi,t + εjt (5)
where for each firm i, and fiscal year-end t:
ADA = absolute value of error term εi,t in Equation (3) ;
ADA_FULL = absolute value of error term εi,t in Equation (4);
ADREV = absolute value of error term εi,t in Equation (5);
AC = (cash flow from operations - income before extraordinary items)/average total assets;
ΔR = (revenuet - revenuet-1)/average total assets;
PPE = gross property, plant and equipment/average total assets;
ROA = (net income before extraordinary items)/average total assets;
CFO = (cash flow from operations)/average total assets;
D = ―1‖ if CFO is negative and ―0‖ otherwise;
ΔAR = change in accounts receivable reported in the cash flow statement;
SIZE = natural logarithm of total assets;
AGE = natural logarithm of the number of years since the firm has data in COMPUSTAT;
GRR_P = industry-median-adjusted revenue growth (=0 if negative);
GRR_N = industry-median-adjusted revenue growth (=0 if positive);
GRM = industry-median-adjusted gross margin; and
_SQ = square of variable.
I estimate each model by industry (defined as two-digit SIC code) and year, eliminating observations in
industries with less than 20 observations. All variables are winsorized at the 1 and 99 percent levels before
estimating each model.
82
APPENDIX C –Issues Identified by PCAOB during the Inspections of Big 4 Firms from 2003 to 2008
Involving the Application of Analytical Procedures
(SOURCE: http://pcaobus.org/Inspections/Reports/Pages/default.aspx)
PCAOB Report
Issues Identified by PCAOB
2007 Inspection of Deloitte
& Touche LLP
―The Firm failed to test the data it used in its analytical procedure regarding
additions to a significant intangible asset. ―
2008 Inspection of Ernst &
Young LLP
―The Firm's substantive procedures to test sales cut-off were limited to
analytical procedures that failed to provide the necessary level of assurance
because the Firm did not establish expectations for the procedures.‖
2007 Inspection of Ernst &
Young LLP
―While the Firm also performed analytical procedures in support of its
testing, these procedures lacked sufficient precision to detect
misstatements that might, individually or in the aggregate, be material.‖
2005 Inspection of Ernst &
Young LLP
―The Firm also failed to perform sufficient substantive audit procedures with
respect to revenue, including unbilled revenue. The Firm failed to test, other
than through analytical review of issuer-prepared reports, unbilled
revenue.‖
2004 Inspection of Ernst &
Young LLP
―The Firm's documented testing specific to this aspect of the banks' ALL was
limited to certain high level analytical procedures that were not
substantive tests.‖
2007 Inspection of KPMG
LLP
―The Firm's analytical procedures did not meet the requirements for
substantive tests, because the Firm's expectation was developed using data
from the same system from which the issuer's actual interest expense was
derived.‖
2005 Inspection of KPMG
LLP
―The Firm’s alternative procedures were limited to analytical procedures,
and the Firm did not obtain corroboration for management’s explanations of
significant differences from the Firm’s expectations.‖
2004 Inspection of KPMG
LLP
―The Firm performed a substantive analytical procedure using expectations
based on statistics the validity of which the Firm did not test.‖
2003 Limited Inspection of
KPMG LLP
―It appeared that appropriate audit procedures had been performed on the
majority of the revenue, but audit testing for the remainder of the total
revenue consisted only of an analytical review. The staff concluded that the
analytical procedure should have, at a minimum, been supported by tests
verifying the amounts of the assets under management supplied by the related
party.‖
83
APPENDIX C (continued)
2006 Inspection of
PricewaterhouseCoopers
LLP
―The Firm's planned substantive testing was limited to analytical procedures that
covered the entire balance. As executed, however, these procedures did not
qualify as substantive tests, as the Firm failed to establish precise
expectations, set a threshold that would identify a potential material
misstatement, and obtain corroboration of management's explanations of
certain significant differences.‖
2005 Inspection of
PricewaterhouseCoopers
LLP
―There was no evidence in the audit documentation, and no persuasive other
evidence, that the Firm had obtained corroboration of explanations provided
by management for significant unexpected fluctuations that the Firm had
identified when performing substantive analytical procedures over revenues.‖
―The Firm failed to perform sufficient procedures to test the valuation and
existence assertions related to certain portions of the inventory. The deficiencies
included the failure to establish a precise expectation, and to appropriately
investigate or evaluate significant differences, when performing substantive
analytical procedures.‖
―The Firm relied on substantive analytical procedures in its tests of revenue, but
failed to test the reliability and accuracy of the data it used to establish its
expectations and failed to obtain corroboration of the issuer's explanations of significant differences.‖
―The Firm limited its tests of the existence and accuracy of accounts receivable
to tests of controls and analytical procedures. The analytical procedures did not
qualify as substantive analytical procedures because there was no evidence in
the audit documentation, and no persuasive other evidence, that the Firm
had established precise enough expectations or obtained corroboration of
management's explanations of certain significant differences.‖
2004 Inspection of
PricewaterhouseCoopers
LLP
―In two audits, the Firm used analytical procedures as alternative procedures.
The Firm, however, did not perform certain required steps, such as
developing expectations and defining a significant difference or threshold,
when performing the analytical procedures.‖
―On two audits, the Firm failed to perform sufficient procedures on the activity
in accounts receivable between the interim date as of which the alternative
procedures were performed and year end. In both instances, the Firm only
performed high-level analytical procedures to test the activity during this
period.‖
2003 Limited
Inspection of
PricewaterhouseCoopers
LLP
―The engagement team did not compare the amounts reported by the third-party
inventory services company with the issuer's records for any locations not visited
by the engagement team. The staff expressed its view that such a procedure
would have been appropriate as a test of the issuer's controls. In response, the
engagement team expressed its view that its testing of controls over the
physical inventory, combined with certain substantive analytical procedures
over inventories, was adequate.‖
84
APPENDIX D –Variable Definitions –Chapter 2
ARi,t = accounts receivables in year-quarter t, scaled by total assets;
ARi,t-4 = accounts receivables in year-quarter t-4, scaled by total assets;
CHSALESi,t = ratio of the current quarter sales to the sales in the same quarter of the previous year;
ΔSALESt = sales in year-quarter t minus sales in year-quarter t-4, scaled by total assets;
INVi,t = inventory in year-quarter t, scaled by total assets;
INVi,t-4 = inventory in year-quarter t-4, scaled by total assets;
CHCOGSi,t = ratio of the current quarter cost of goods sold to the cost of goods sold in the same
quarter of the previous year;
ΔCOGSi,t = cost of goods sold in year-quarter t minus cost of goods sold in year-quarter t-4, scaled
by total assets.;
jARi,t = accounts receivables from closest peer firm in year-quarter t, scaled by total assets;
jARi,t-4 = accounts receivables from closest peer firm in year-quarter t-4, scaled by total assets;
jINVi,t = inventory from closest peer firm in year-quarter t, scaled by total assets;
jINVi,t-4 = inventory from closest peer firm in year-quarter t-4, scaled by total assets;
PEER-R2 = R
2 of regression of firm i’s stock returns on peer firm j’s stock returns over 48 months
ending in the year prior to the audit; and,
PEER-TAU = Tau correlation coefficient between firm i’s stock returns and peer firm j’s stock returns,
estimated over 48 months ending in the year prior to the audit.
85
APPENDIX E –Summary of Expectation Models
Heuristic Expectation Models Without Peers
E[ARi,t] = ARi,t-4*CHSALESi,t (AR1)
𝐸[𝐴𝑅𝑖,𝑡] = ∑𝐴𝑅𝑖,𝑡−4−𝑘
12
12
𝑘=1
(AR2)
E[INVi,t] = INVi,t-4*CHCOGSi,t (INV1)
𝐸[𝐼𝑁𝑉𝑖,𝑡] = ∑𝐼𝑁𝑉𝑖,𝑡−4−𝑘
12
12
𝑘=1
(INV2)
Heuristic Expectation Models With Peers
Unexpected ARi,t = [ARi,t –ARi,t-4*CHSALESi,t]– [jARi,t – jARi,t-4*jCHSALESi,t] (AR1P)
𝑈𝑛𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝐴𝑅𝑖,𝑡 = [𝐴𝑅𝑖,𝑡 − ∑𝐴𝑅𝑖,𝑡−4−𝑘
12
12
𝑘=1
] − [𝑗𝐴𝑅𝑖,𝑡 − ∑𝑗𝐴𝑅𝑖,𝑡−4−𝑘
12
12
𝑘=1
] (AR2P)
Unexpected INVi,t = [INVi,t –INVi,t-4*CHCOGSi,t]– [jINVi,t – jINVi,t-4*jCHCOGSi,t] (AR1P)
𝑈𝑛𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝐼𝑁𝑉𝑖,𝑡 = [𝐼𝑁𝑉𝑖,𝑡 − ∑𝐼𝑁𝑉𝑖,𝑡−4−𝑘
12
12
𝑘=1
] − [𝑗𝐼𝑁𝑉𝑖,𝑡 − ∑𝑗𝐼𝑁𝑉𝑖,𝑡−4−𝑘
12
12
𝑘=1
] (AR2P)
Time-Series and Industry Cross-Sectional Expectation Models Without Peers
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + εi,t (AR3 – AR4)
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + εi,t (INV3 – INV4)
Time-Series and Industry Cross-Sectional Expectation Models With Peers
ARi,t = β0 + β1ARi,t-4 + β2ΔSALESi,t + β3jARi,t-4 + β4jΔSALESi,t + εi,t (AR3P – AR4
P)
INVi,t = γ0 + γ1INVi,t-4 + γ2ΔCOGSi,t + γ3jINVi,t-4 + γ4jΔCOGSi,t + εi,t (INV3P – INV4
P)
In addition, all multivariate models with peer-firm information are estimated with the following variations:
1) Including only accounts receivables or inventory from a peer firm, in year-quarter t, scaled by total assets
(jARi,t or jINVi,t );
2) Including only accounts receivables or inventory from a peer firm, in year-quarter t and t-4, scaled by total
assets (jARi,t and jINVi,t or jARi,t-4 and jINVi,t-4);
3) Including information from up to five peers (ΔjARi,t-4 and jΔSALESi,t or ΔjINVi,t-4 and jΔCOGSi,t for j=1,
…,5);
4) Using the quarterly industry cross-sectional average and year-quarter t-4 average of accounts receivables or
inventory, scaled by total assets at the end of the quarter (IndustryARi,t-4 and IndustryΔSALESi,t or
IndustryΔCOGSi,t and IndustryINVi,t-4)
86
APPENDIX F –Summary of Simulation Procedure
The results shown in Tables 3, 4, and 5 and are calculated using the following steps:
1) Estimate models each quarter using firm-level data or industry cross-sectional data for all companies in
the full sample;
2) Calculate quarter firm-level errors by fitting model parameters to quarterly data, with and without
simulated errors;
3) Average quarter firm-level errors for the first two, three, or four, quarters of the current year to compute
a year firm-level error measure used as indication of potential error or manipulation;
4) Draw a sub-sample of 100 firm year-level errors at random from the full sample;
5) Obtain the mean, median, and standard deviation of errors for the sub-sample of 100 firm year-level
errors;
a. When there is no manipulation, the prediction model is misspecified if the mean of the 100 firm
year-level errors is statistically different than zero at the five percent level, positive or negative.
b. When there is manipulation, the prediction model has power if the mean of the 100 firm year-
level errors is statistically different than zero at the five percent level, in the same direction as
the simulated error.
6) Repeat steps one to five 250 times (i.e., draw 250 samples of 100 observations);
7) Obtain the overall mean, median, and standard deviation of firm year-level errors across the 250
samples; and,
8) Compute the percentage of the 250 samples were the mean of firm year-level errors is statistically
different than zero, with and without manipulation.
87
FIGURE 1- LEAD1 Simulation Results from 1,000 replications
Assigning Clients to Five Auditors at Random –Full Sample
FIGURE 2- LEAD1 Simulation Results from 1,000 replications
Assigning Clients to Five Auditors at Random –Sample of Industries without Specialist
01
02
03
0
Fre
que
ncy
-.006 -.004 -.002 0 .002
Coefficient on NLEAD1 VariableNegative in 97.6% of Iterations
Mean = -0.0020
Simulation Results NLEAD1Discretionary Accruals Multivariate Regression
All Industries0
10
20
30
40
Fre
que
ncy
-.006 -.004 -.002 0 .002 .004
Coefficient on NLEAD1 VariableNegative in 69.3% of Iterations
Mean = -0.0009
Simulation Results NLEAD1Discretionary Accruals Multivariate Regression
Industries Without Specialist
88
TABLE 1 –Discretionary Accruals Analyses
PANEL A: Descriptive Statistics Ranking Auditors by Industry Market Share at National Level
Market Total Total assets Market value ROA GROWTH ABS(ACCRL) ADA
share Number Mean Mean Mean Mean Mean Mean
rank of clients Median Median Median Median Median Median
1 14,759 3,444 3,671 -0.006 0.077 0.136 0.055
467 434 0.037 0.061 0.062 0.036
2 12,947 2,552 2,697 -0.005 0.082 0.133 0.057
366 316 0.037 0.066 0.062 0.037
3 11,659 1,941 2,151 -0.018 0.079 0.143 0.061
252 222 0.035 0.066 0.067 0.041
4 10,519 1,433 1,660 -0.036 0.070 0.160 0.065
198 177 0.029 0.063 0.068 0.043
5 7,254 676 886 -0.045 0.074 0.189 0.070
99 82 0.028 0.066 0.070 0.046
6 5,543 394 454 -0.059 0.066 0.202 0.076
61 46 0.022 0.060 0.074 0.049
7 3,234 177 171 -0.075 0.054 0.343 0.081
24 18 0.013 0.048 0.083 0.055
8 1,921 75 82 -0.080 0.074 0.417 0.087
19 15 0.009 0.049 0.090 0.057
89
PANEL B: Full Sample Partition by Industry Specialization at National Level (NLEAD1)
Full sample NLEAD1 Matched sample a
I II III IV V VI VII All Obs. NLEAD1=1 NLEAD1=0 Univariate NLEAD1=1 NLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
ADA 0.068 0.053 0.070 -0.017*** 0.054 0.054 -0.001 0.077 0.060 0.079 (-19.22) 0.063 0.065 (-0.74)
Total assets 1,723 3,676 1,486 2,190*** 3,790 3,253 537*** 4,873 7,280 4,435 (-38.68) 8,950 7,714 (3.36)
Market value 1,823 3,851 1,644 2,207*** 4,879 4,101 778** 6,202 9,066 5,711 (-30.53) 18,857 16,180 (2.32)
BIG4 0.793 0.981 0.771 0.210*** 0.975 0.894 0.081*** 0.405 0.137 0.421 (44.86) 0.156 0.308 (17.37)
LOG_MKT 4.870 5.990 4.734 1.256*** 6.038 5.890 0.148*** 2.535 2.510 2.505 (42.70) 2.330 2.291 (3.36)
LEV 0.256 0.282 0.253 0.029*** 0.278 0.269 0.009** 0.251 0.238 0.253 (9.84) 0.236 0.226 (2.06)
ROAL -0.035 0.005 -0.040 0.045*** 0.005 -0.004 0.009* 0.263 0.195 0.270 (14.52) 0.223 0.293 (1.82)
ROA -0.046 -0.004 -0.051 0.046*** -0.002 -0.010 0.008* 0.254 0.187 0.260 (15.63) 0.199 0.247 (1.78)
LOSS 0.364 0.285 0.373 -0.088*** 0.288 0.297 -0.009 0.481 0.451 0.484 (-15.66) 0.453 0.457 (-1.03)
CFO 0.029 0.061 0.025 0.036*** 0.061 0.054 0.007* 0.201 0.158 0.205 (15.17) 0.163 0.229 (1.88)
BTM 0.444 0.419 0.447 -0.028*** 0.410 0.445 -0.035*** 0.429 0.369 0.435 (-5.59) 0.742 0.435 (-3.04)
ABS(ACCRL) 0.145 0.102 0.151 -0.049*** 0.122 0.126 -0.004 0.308 0.212 0.318 (-13.43) 0.793 0.799 (-0.24)
GROWTH 0.071 0.078 0.070 0.008** 0.084 0.085 -0.002 0.318 0.266 0.323 (2.13) 0.378 0.282 (-0.25)
ALTMAN 3.452 3.889 3.398 0.491*** 4.263 3.914 0.349** 8.279 6.341 8.483 (5.05) 10.250 7.783 (2.01)
STDEARN 37.24 62.17 34.22 27.95*** 63.14 55.03 8.11*** 94.23 122.20 89.78 (25.39) 153.40 134.60 (2.94)
TENURE 0.993 0.992 0.993 -0.001 0.994 0.994 0.001 0.084 0.090 0.083 (1.15) 0.076 0.076 (0.00)
N. Obs. 75,188 8,147 67,041 75,188 5,479 5,479 10,958 % Obs. 100.00% 10.84% 89.16% 50.00% 50.00%
90
PANEL C: Full Sample Partition by Industry Specialization at City Level (CLEAD1)
Full sample CLEAD1 Matched sample a
I II III IV V VI VII All Obs. CLEAD1=1 CLEAD1=0 Univariate CLEAD1=1 CLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
ADA 0.075 0.057 0.085 -0.028*** 0.056 0.059 -0.002* 0.089 0.070 0.097 (-22.68) 0.070 0.072 (-1.74)
Total assets 1,979 4,128 877 3,251*** 3,380 2,673 707*** 5,621 8,238 3,072 (33.44) 8,108 6,203 (4.89)
Market value 2,175 4,399 1,035 3,364*** 3,799 2,892 907*** 7,127 10,450 4,129 (35.01) 15,671 12,132 (3.23)
BIG4 0.718 0.940 0.605 0.335*** 0.963 0.855 0.108*** 0.450 0.239 0.489 (57.43) 0.189 0.352 (19.09)
LOG_MKT 5.164 6.367 4.548 1.819*** 6.407 6.228 0.179*** 2.522 2.382 2.365 (55.43) 1.969 1.922 (4.58)
LEV 0.242 0.259 0.233 0.025*** 0.249 0.238 0.011** 0.284 0.247 0.300 (6.45) 0.245 0.249 (2.17)
ROAL -0.079 -0.014 -0.112 0.098*** -0.014 -0.038 0.024*** 0.342 0.239 0.380 (20.98) 0.237 0.279 (4.65)
ROA -0.088 -0.022 -0.123 0.101*** -0.022 -0.042 0.020*** 0.321 0.223 0.356 (23.04) 0.221 0.244 (4.25)
LOSS 0.425 0.318 0.480 -0.162*** 0.319 0.364 -0.044*** 0.494 0.466 0.500 (-23.98) 0.466 0.481 (4.67)
CFO 0.003 0.053 -0.022 0.075*** 0.051 0.033 0.017*** 0.255 0.182 0.282 (21.58) 0.191 0.204 (4.40)
BTM 0.417 0.388 0.432 -0.044*** 0.409 0.425 -0.016** 0.461 0.380 0.498 (-7.01) 0.394 0.383 (2.06)
ABS(ACCRL) 0.125 0.092 0.142 -0.050*** 0.086 0.089 -0.003 0.171 0.118 0.191 (-21.21) 0.102 0.109 (1.34)
GROWTH 0.052 0.067 0.045 0.022*** 0.070 0.063 0.008 0.310 0.245 0.338 (5.04) 0.351 0.307 (1.19)
ALTMAN 2.306 3.630 1.627 2.003*** 4.433 3.968 0.465* 11.550 7.623 13.060 (12.57) 15.260 9.347 (1.83)
STDEARN 51.42 88.53 32.40 56.13*** 71.68 64.13 7.55*** 123.10 165.20 88.66 (33.75) 165.40 140.30 (2.46)
TENURE 0.997 0.996 0.998 -0.002*** 0.997 0.999 -0.002* 0.055 0.064 0.050 (-2.09) 0.053 0.035 (1.79)
N. Obs.b 23,307 7,897 15,410 23,307 4,979 4,979 9,958
% Obs. 100.00% 33.88% 66.12% 50.00% 50.00%
91
This table presents the descriptive statistics of the data used in the discretionary accruals analyses: Panel A shows descriptive statistics ranking the top eight auditors by market share
at industry-national level; Panel B shows descriptive statistics for the full sample and for a partition by industry specialization at national level (NLEAD1); and Panel C shows
descriptive statistics for the full sample and for a partition by industry specialization at city level (CLEAD1). a Clients of specialist and non-specialist auditors are matched based on
industry, size, and stock returns covariance. Variable definitions are included in Appendix A. *, **, *** indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-
tailed tests. b The sample size is reduced to those years with auditor city data in audit analytics matched to the U.S. Census Bureau 2003 list of Metropolitan Statistical Areas.
92
TABLE 2 – Discretionary Accruals Analyses: Pooled Multivariate Tests
Full and Matched Samples
Dependent variable = ADA
Full samples Matched samples a
I II III IV
NLEAD1 CLEAD1 NLEAD1 CLEAD1
Predicted Estimate Estimate Estimate Estimate
sign (t-statistic) (t-statistic) (t-statistic) (t-statistic)
NLEAD1 ( - ) -0.0037*** 0.0006
(-4.39) (0.40)
CLEAD1 ( - ) -0.0026** 0.0009
(-2.18) (0.56)
BIG4 -0.0082*** -0.0104*** -0.0029 -0.0050
(-7.59) (-5.87) (-0.73) (-1.42)
LOG_MKT -0.0060*** -0.0062*** -0.0050*** -0.0053***
(-29.02) (-15.30) (-9.21) (-8.71)
LEV -0.0357*** -0.0306*** -0.0288*** -0.0237***
(-17.30) (-8.80) (-5.20) (-4.66)
ROAL 0.0343*** 0.0178** 0.0199 0.0034
(6.40) (2.35) (1.25) (0.16)
ROA -0.1116*** -0.0985*** -0.1003*** -0.1264***
(-16.17) (-9.08) (-3.91) (-4.04)
LOSS -0.0031*** -0.0044*** 0.0061** -0.0010
(-3.54) (-2.89) (2.40) (-0.34)
CFO -0.0133** 0.0076 0.0054 0.0411
(-2.39) (0.77) (0.22) (1.59)
BTM -0.0242*** -0.0194*** -0.0077 -0.0152***
(-21.40) (-9.99) (-1.52) (-5.26)
ABS(ACCRL) 0.0269*** 0.0764*** 0.0013 0.0893***
(16.74) (11.97) (0.92) (6.45)
GROWTH 0.0106*** 0.0086*** 0.0104*** 0.0051
(7.95) (3.16) (2.68) (1.59)
ALTMAN -0.0003*** -0.0004*** 0.0001 -0.0001*
(-3.76) (-3.73) (0.98) (-1.83)
STDEARN 0.0001*** 0.0001*** 0.0001** 0.0001**
(6.51) (3.89) (2.14) (2.48)
TENURE -0.0046 0.0006 0.0057 0.0148**
(-1.44) (0.09) (1.02) (2.44)
Intercept 0.0979*** 0.1215*** 0.0842*** 0.0839***
(23.88) (16.22) (8.46) (9.95)
Year F.E. Included Included Included Included
N. Obs. 75,188 23,307 10,958 9,958
Adj-R2 0.21 0.25 0.16 0.20
This table presents the pooled discretionary accruals analyses for the full sample and matched samples, using NLEAD1and CLEAD1 as
definitions of auditor industry specialization. a Clients of specialist and non-specialist auditors are matched based on industry, size, and
stock returns covariance. Variable definitions are included in Appendix A. *, **, *** indicate significance at the 0.10, 0.05 and 0.01
levels, respectively, using two-tailed tests. T-statistics and p-values are calculated using clustered standard errors by firm. For brevity,
the year-specific intercepts are not reported.
93
TABLE 3 – Discretionary Accruals Analyses: Multivariate Pair-Wise Differences Tests
Dependent variable = ADAij
Matched samples a
I II
NLEAD1 CLEAD1
Predicted Estimate Estimate
sign (t-statistic) (t-statistic)
BIG4ij -0.0008 -0.0001
(-0.19) (-0.03)
LOG_MKTij 0.0052*** 0.0008
(3.68) (0.51)
LEVij 0.0003 -0.0021
(0.07) (-0.28)
ROALij 0.0148 0.0094
(1.12) (0.51)
ROAij -0.0901*** -0.1189***
(-3.92) (-3.84)
LOSSij 0.0033 -0.0021
(1.23) (-0.64)
CFOij 0.0210 0.0506*
(1.03) (1.75)
BTMij -0.0034 -0.0082*
(-1.34) (-1.90)
ABS(ACCRL)ij 0.0009 0.0659***
(0.19) (3.92)
GROWTHij 0.0075** -0.0011
(2.24) (-0.32)
ALTMANij -0.0003** -0.0003*
(-2.14) (-1.87)
STDEARNij 0.0001* 0.0001***
(1.73) (3.07)
TENUREij 0.0068 0.0199*
(0.77) (1.81)
Intercept ( - ) -0.0013 -0.0014
(-1.12) (-1.02)
Year F.E. Included Included
N. Obs. 5,479 4,979
Adj-R2 0.05 0.08
This table presents the pair-wise differences in discretionary accruals for the matched samples, using NLEAD1 and CLEAD1 as
definitions of auditor industry specialization. a Clients of specialist and non-specialist auditors are matched based on industry, size,
and stock returns covariance. Variable definitions are included in Appendix A and ij denotes the pair-wise difference between
matched observations. *, **, *** indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. T-
statistics and p-values are calculated using clustered standard errors by firm. For brevity, the year-specific intercepts are not
reported.
94
TABLE 4 – Discretionary Revenue Analyses – PANEL A: Full Sample Partition by Industry Specialization at National Level (NLEAD1)
Full Sample NLEAD1 Matched sample a
I II III IV V VI VII All Obs. NLEAD1=1 NLEAD1=0 Univariate NLEAD1=1 NLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
ADREV 0.031 0.024 0.032 -0.008*** 0.024 0.023 0.001 0.038 0.031 0.039 (-17.45) 0.031 0.031 (0.84)
Total assets 1,667 3,594 1,432 2,162*** 3,714 3,174 540*** 4,721 7,121 4,280 (38.01) 8,774 7,501 (3.32)
Market value 1,798 3,695 1,566 2,129*** 4,697 3,836 861*** 5,855 8,572 5,385 (30.05) 18,141 15,028 (2.60)
BIG4 0.797 0.980 0.775 0.205*** 0.973 0.893 0.081*** 0.402 0.141 0.418 (42.33) 0.161 0.310 (16.41)
LOG_MKT 4.878 5.979 4.744 1.235*** 6.024 5.870 0.154*** 2.505 2.501 2.472 (40.97) 2.321 2.284 (3.35)
LEV 0.256 0.285 0.253 0.032*** 0.281 0.268 0.013*** 0.247 0.238 0.248 (10.74) 0.239 0.224 (2.76)
ROAL -0.025 0.011 -0.029 0.040*** 0.012 0.003 0.009* 0.241 0.179 0.247 (13.81) 0.210 0.255 (1.91)
ROA -0.036 0.002 -0.040 0.043*** 0.004 - 0.003 0.007* 0.234 0.170 0.240 (15.04) 0.187 0.192 (1.87)
LOSS 0.353 0.277 0.362 -0.086*** 0.280 0.292 - 0.012 0.478 0.447 0.481 (-14.72) 0.449 0.455 (-1.32)
CFO 0.038 0.067 0.034 0.033*** 0.066 0.061 0.006* 0.183 0.142 0.187 (14.80) 0.149 0.164 (1.85)
BTM 0.446 0.419 0.450 -0.031*** 0.409 0.449 - 0.040*** 0.419 0.362 0.425 (-6.13) 0.762 0.432 (-3.25)
ABS(ACCRL) 0.141 0.101 0.146 -0.045*** 0.123 0.121 0.003 0.293 0.205 0.302 (-12.51) 0.823 0.688 (0.18)
GROWTH 0.074 0.081 0.074 0.007* 0.084 0.086 - 0.003 0.314 0.265 0.320 (1.84) 0.382 0.283 (-0.37)
ALTMAN 3.461 3.770 3.424 0.346*** 4.183 3.872 0.311* 6.884 5.418 7.042 (4.13) 9.530 6.660 (1.91)
STDEARN 36.09 60.83 33.08 27.75*** 62.68 53.50 9.18*** 90.68 118.80 86.14 (25.24) 153.70 130.20 (3.24)
TENURE 0.994 0.993 0.995 -0.001 0.994 0.997 - 0.003* 0.075 0.082 0.074 (-1.40) 0.076 0.056 (-1.94)
N. Obs. 69,512 7,558 61,954 69,512 5,053 5,053 10,106 % Obs. 100.00% 10.87% 89.13% 50.00% 50.00%
95
PANEL B: Full Sample Partition by Industry Specialization at City Level (CLEAD1)
Full sample CLEAD1 Matched sample a
I II III IV V VI VII All Obs. CLEAD1=1 CLEAD1=0 Univariate CLEAD1=1 CLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
ADREV 0.030 0.022 0.034 -0.012*** 0.023 0.023 - 0.001 0.038 0.029 0.041 (-21.91) 0.029 0.030 (-0.70)
Total assets 2,000 4,152 887 3,265*** 3,370 2,664 706*** 5,646 8,254 3,082 (42.19) 8,148 6,205 (4.72)
Market value 2,186 4,378 1,052 3,326*** 3,735 2,910 825*** 7,108 10,359 4,171 (33.67) 15,773 12,397 (2.82)
BIG4 0.726 0.943 0.613 0.330*** 0.963 0.850 0.112*** 0.446 0.231 0.487 (55.46) 0.190 0.357 (19.05)
LOG_MKT 5.210 6.398 4.596 1.802*** 6.398 6.215 0.183*** 2.495 2.355 2.339 (53.93) 1.966 1.933 (4.56)
LEV 0.241 0.260 0.231 0.029*** 0.249 0.239 0.010* 0.274 0.243 0.289 (7.51) 0.245 0.249 (1.87)
ROAL -0.062 -0.004 -0.092 0.088*** - 0.006 - 0.029 0.023*** 0.308 0.215 0.342 (20.24) 0.221 0.258 (4.58)
ROA -0.072 -0.013 -0.103 0.091*** - 0.015 - 0.034 0.020*** 0.290 0.201 0.323 (22.16) 0.202 0.227 (4.39)
LOSS 0.413 0.309 0.467 -0.158*** 0.314 0.356 - 0.042*** 0.492 0.462 0.499 (-22.70) 0.464 0.479 (-4.33)
CFO 0.017 0.061 -0.006 0.068*** 0.057 0.039 0.018*** 0.229 0.162 0.253 (20.95) 0.172 0.187 (4.80)
BTM 0.421 0.389 0.438 -0.050*** 0.413 0.429 - 0.017** 0.451 0.372 0.487 (-7.71) 0.395 0.387 (-2.05)
ABS(ACCRL) 0.120 0.090 0.135 -0.045*** 0.086 0.089 - 0.003 0.157 0.109 0.175 (-20.50) 0.101 0.107 (-1.49)
GROWTH 0.056 0.068 0.050 0.018*** 0.071 0.064 0.006 0.301 0.241 0.328 (4.24) 0.358 0.305 (0.95)
ALTMAN 2.502 3.559 1.956 1.603*** 4.142 3.832 0.310* 9.449 6.434 10.640 (11.94) 9.529 8.560 (1.66)
STDEARN 52.45 89.94 33.05 56.89*** 72.30 64.15 8.15** 124.90 167.10 90.00 (32.74) 166.50 141.60 (2.55)
TENURE 0.997 0.996 0.998 -0.001* 0.997 0.999 - 0.002*** 0.052 0.060 0.047 (-1.89) 0.055 0.025 (-2.67)
N. Obs.b 21,914 7,471 14,443 21,914 4,695 4,695 9,390
% Obs. 100.00% 34.09% 65.91% 50.00% 50.00%
96
This table presents the descriptive statistics of the data used in the discretionary revenue analyses: Panel A shows descriptive statistics for the full sample and for a partition by
industry specialization at national level (NLEAD1); and, Panel B shows descriptive statistics for the full sample and for a partition by industry specialization at city level (CLEAD1). a
Clients of specialist and non-specialist auditors are matched based on industry, size, and stock returns covariance. Variable definitions are included in Appendix A. *, **, *** indicate
significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. b The sample size is reduced to those years with auditor city data in audit analytics matched to the
U.S. Census Bureau 2003 list of Metropolitan Statistical Areas.
97
TABLE 5 – Discretionary Revenue Analyses: Pooled Multivariate Tests
Full and Matched Samples
Dependent Variable = ADREV
Full samples Matched samples a
I II III IV
NLEAD1 CLEAD1 NLEAD1 CLEAD1
Predicted Estimate Estimate Estimate Estimate
sign (t-statistic) (t-statistic) (t-statistic) (t-statistic)
NLEAD1 ( - ) -0.0013** 0.0011
(-2.41) (1.32)
CLEAD1 ( - ) -0.0020*** 0.0004
(-3.23) (0.50)
BIG4 -0.0041*** -0.0062*** -0.0001 -0.0023
(-5.92) (-6.23) (-0.05) (-1.32)
LOG_MKT -0.0042*** -0.0036*** -0.0036*** -0.0039***
(-32.07) (-16.42) (-12.43) (-11.52)
LEV -0.0208*** -0.0138*** -0.0153*** -0.0131***
(-17.91) (-7.51) (-6.57) (-5.83)
ROAL 0.0115*** 0.0073** 0.0089* 0.0171***
(4.12) (2.16) (1.93) (3.79)
ROA -0.0058* -0.0040 -0.0085 -0.0138**
(-1.86) (-0.97) (-1.37) (-2.28)
LOSS 0.0024*** 0.0005 0.0020* -0.0001
(5.07) (0.62) (1.77) (-0.06)
CFO -0.0116*** -0.0032 -0.0091* 0.0004
(-5.84) (-1.08) (-1.71) (0.08)
BTM -0.0113*** -0.0071*** -0.0027 -0.0037**
(-16.78) (-6.83) (-1.30) (-2.24)
ABS(ACCRL) 0.0098*** 0.0212*** 0.0009 0.0125***
(12.08) (7.70) (1.27) (2.92)
GROWTH 0.0105*** 0.0086*** 0.0087*** 0.0027
(14.04) (5.92) (3.67) (1.17)
ALTMAN -0.0002*** -0.0002*** 0.0010 0.0010
(-5.26) (-4.25) (0.93) (0.86)
STDEARN 0.0001*** 0.0001* 0.0001*** 0.0001***
(6.13) (1.75) (4.82) (4.06)
TENURE -0.0038** 0.0093*** -0.0002 0.0075**
(-2.24) (4.93) (-0.06) (2.39)
Intercept 0.0590*** 0.0537*** 0.0478*** 0.0517***
(26.39) (20.19) (10.33) (11.40)
Year F.E. Included Included Included Included
N. Obs. 69,512 21,914 10,106 9,390
Adj-R2 0.12 0.13 0.13 0.09
This table presents the pooled discretionary revenue analyses of the full sample and matched samples, using NLEAD1and CLEAD1 as
definitions of auditor industry specialization. a Clients of specialist and non-specialist auditors are matched based on industry, size, and
stock returns covariance. Variable definitions are included in Appendix A. *, **, *** indicate significance at the 0.10, 0.05 and 0.01
levels, respectively, using two-tailed tests. T-statistics and p-values are calculated using clustered standard errors by firm. For brevity,
the year-specific intercepts are not reported.
98
TABLE 6 – Discretionary Revenue Analyses: Multivariate Pair-Wise Differences Tests
Dependent variable = ADREVij
Matched samples a
I II
NLEAD1 CLEAD1
Predicted Estimate Estimate
Sign (t-statistic) (t-statistic)
BIG4ij 0.0011 -0.0010
(0.59) (-0.56)
LOG_MKTij -0.0001 -0.0022***
(-0.14) (-3.00)
LEVij -0.0060** -0.0005
(-2.20) (-0.19)
ROALij -0.0024 0.0207***
(-0.55) (4.06)
ROAij -0.0002 -0.0141*
(-0.03) (-1.85)
LOSSij 0.0029** 0.0003
(2.26) (0.25)
CFOij 0.0074 0.0043
(1.15) (0.62)
BTMij -0.0020 -0.0023
(-1.05) (-1.23)
ABS(ACCRL)ij 0.0027* 0.0150***
(1.75) (2.88)
GROWTHij 0.0082*** 0.0025
(2.79) (0.90)
ALTMANij -0.0002** -0.0001
(-2.30) (-1.38)
STDEARNij 0.0010 0.0010
(1.11) (1.43)
TENUREij -0.0014 -0.0025
(-0.44) (-0.60)
Intercept ( - ) 0.0005 -0.0002
(0.67) (-0.26)
Year F.E. Included Included
N. Obs. 5,053 4,695
Adj-R2 0.02 0.02
This table presents the pair-wise differences in discretionary revenue for the matched samples, using NLEAD1 and CLEAD1 as
definitions of auditor industry specialization. a Clients of specialist and non-specialist auditors are matched based on industry, size,
and stock returns covariance. Variable definitions are included in Appendix A and ij denotes the pair-wise difference between
matched observations. *, **, *** indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. T-
statistics and p-values are calculated using clustered standard errors by firm. For brevity, the year-specific intercepts are not
reported.
99
TABLE 7 – Going-Concern Analyses – PANEL A: Full Sample Partition by Industry Specialization at National level (NLEAD1)
Full sample NLEAD1 Matched sample a
I II III IV V VI VII All Obs. NLEAD1=1 NLEAD1=0 Univariate NLEAD1=1 NLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
GCONCERN 0.095 0.043 0.102 -0.059*** 0.029 0.026 0.003 0.293 0.202 0.302 (-12.55) 0.167 0.159 (0.60)
Total assets 2,608 5,957 2,138 3,819*** 6,143 5,294 849* 7,773 11,814 6,896 (30.75) 17,042 15,291 (1.87)
Market value 2,785 6,071 2,324 3,747*** 7,077 5,803 1,274** 9,159 13,777 8,206 (25.50) 24,418 19,232 (2.07)
BIG4 0.725 1.000 0.687 0.313*** 1.000 0.894 0.106*** 0.446 0.000 0.464 (44.54) 0.000 0.307 (17.31)
LOG_MKT 5.211 6.638 5.011 1.627*** 6.699 6.552 0.147** 2.634 2.412 2.602 (39.01) 2.178 2.167 (2.41)
LEV 0.246 0.257 0.244 0.013*** 0.251 0.243 0.007 0.279 0.228 0.285 (2.95) 0.230 0.215 (1.20)
ROAL -0.069 -0.005 -0.078 0.072*** 0.002 - 0.010 0.013 0.335 0.223 0.347 (13.37) 0.591 0.273 (0.97)
ROA -0.077 -0.012 -0.086 0.074*** - 0.006 - 0.016 0.010 0.310 0.209 0.321 (14.76) 0.481 0.224 (0.99)
LOSS 0.405 0.297 0.420 -0.122*** 0.302 0.312 - 0.010 0.491 0.457 0.494 (-15.43) 0.459 0.463 (-0.79)
CFO 0.010 0.056 0.004 0.052*** 0.053 0.047 0.005 0.248 0.178 0.255 (12.99) 0.183 0.201 (0.99)
BTM 0.432 0.407 0.436 -0.029*** 0.384 0.414 - 0.029*** 0.470 0.377 0.482 (-3.84) 0.348 0.351 (-2.99)
ABS(ACCRL) 0.123 0.083 0.128 -0.045*** 0.091 0.082 0.010 0.173 0.106 0.180 (-16.01) 0.559 0.096 (0.86)
GROWTH 0.055 0.063 0.054 0.009* 0.063 0.076 - 0.013 0.307 0.232 0.317 (1.89) 0.353 0.265 (-1.52)
ALTMAN 2.472 3.916 2.270 1.646*** 4.764 4.167 0.597** 11.530 7.556 11.970 (8.83) 12.720 8.230 (1.98)
STDEARN 78.75 137.40 70.53 66.87*** 138.20 122.20 16.00 238.80 307.40 226.30 (17.36) 440.00 387.20 (1.37)
TENURE 0.985 0.984 0.985 -0.001 0.987 0.987 - 0.001 0.122 0.126 0.121 (-0.63) 0.113 0.112 (0.12)
N. Obs. 35,406 4,351 31,055 35,406 2,539 2,539 5,078 % Obs. 100.00% 12.29% 87.71% 50.00% 50.00%
100
PANEL B: Full-Sample Partition by Industry Specialization at City Level (CLEAD1)
Full sample CLEAD1 Matched sample a
I II III IV V VI VII All Obs. CLEAD1=1 CLEAD1=0 Univariate CLEAD1=1 CLEAD1=0 Univariate
Mean Mean Mean Estimate Mean Mean Estimate
Std Dev Std Dev Std Dev (t-statistic) Std Dev Std Dev (t-statistic)
GCONCERN 0.097 0.051 0.120 -0.069*** 0.027 0.031 - 0.004 0.295 0.220 0.325 (-17.02) 0.161 0.173 (-1.26)
Total assets 1,998 4,172 880 3,292*** 3,236 2,577 659*** 5,728 8,406 3,107 (43.23) 8,499 6,539 (4.32)
Market value 2,193 4,440 1,037 3,403*** 3,123 2,672 451** 7,227 10,605 4,159 (34.97) 9,236 11,035 (2.21)
BIG4 0.719 0.940 0.605 0.335*** 0.963 0.854 0.109*** 0.450 0.238 0.489 (57.54) 0.189 0.353 (19.09)
LOG_MKT 5.168 6.368 4.550 1.818*** 6.380 6.208 0.172*** 2.521 2.383 2.361 (55.55) 1.940 1.904 (4.45)
LEV 0.241 0.256 0.233 0.023*** 0.246 0.237 0.009* 0.282 0.244 0.300 (5.92) 0.242 0.247 (1.90)
ROAL -0.080 -0.015 -0.113 0.098*** - 0.016 - 0.039 0.023*** 0.345 0.243 0.383 (20.67) 0.240 0.282 (4.43)
ROA -0.089 -0.023 -0.123 0.100*** - 0.024 - 0.043 0.019*** 0.322 0.226 0.357 (22.73) 0.223 0.246 (4.09)
LOSS 0.425 0.317 0.480 -0.162*** 0.320 0.365 - 0.044*** 0.494 0.466 0.500 (-24.05) 0.467 0.481 (-4.64)
CFO 0.003 0.052 -0.022 0.074*** 0.049 0.032 0.017*** 0.255 0.184 0.281 (21.34) 0.194 0.205 (4.16)
BTM 0.419 0.390 0.434 -0.044*** 0.409 0.425 - 0.016** 0.461 0.379 0.497 (-6.90) 0.395 0.384 (-2.02)
ABS(ACCRL) 0.125 0.093 0.142 -0.049*** 0.087 0.091 - 0.004* 0.173 0.121 0.193 (-20.65) 0.105 0.130 (-1.86)
GROWTH 0.053 0.067 0.045 0.021*** 0.071 0.063 0.008 0.311 0.248 0.338 (5.01) 0.353 0.313 (1.22)
ALTMAN 2.360 3.647 1.698 1.949*** 4.451 3.981 0.470* 11.330 7.569 12.790 (12.49) 15.290 9.369 (1.84)
STDEARN 64.53 113.30 39.41 73.89***
***
101.10 76.72 24.38** 189.30 255.80 136.80 (28.75) 718.60 271.00 (2.24)
TENURE 0.997 0.996 0.998 -0.002** 0.997 0.999 - 0.002* 0.054 0.063 0.049 (-2.18) 0.053 0.035 (-1.79)
N. Obs.b 23,349 7,934 15,415 23,349 4,951 4,951 9,902
% Obs. 100.00% 33.98% 66.02% 50.00% 50.00%
101
This table presents the descriptive statistics of the data used in the analysis of propensity to issue a going-concern opinion: Panel A shows descriptive statistics for the full sample and
for a partition by industry specialization at national level (NLEAD1); and Panel B shows descriptive statistics for the full sample and for a partition by industry specialization at city
level (CLEAD1). a Clients of specialist and non-specialist auditors are matched based on industry, size, and stock returns covariance. Variable definitions are included in Appendix A.
*, **, *** indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. b The sample size is reduced to those years with auditor city data in audit analytics
matched to the U.S. Census Bureau 2003 list of Metropolitan Statistical Areas.
102
TABLE 8 – Going-Concern Analyses: Pooled Multivariate Tests
Full and Matched Samples
Dependent variable = GCONCERN
Full samples Matched samples a
I II III IV
NLEAD1 CLEAD1 NLEAD1 CLEAD1
Predicted Estimate Estimate Estimate Estimate
sign (t-statistic) (t-statistic) (t-statistic) (t-statistic)
NLEAD1 ( + ) 0.0932 0.0384
(0.66) (0.15)
CLEAD1 ( + ) 0.2722*** 0.0563
(2.58) (0.27)
BIG4 -0.1295 -0.2110** 0.5010 0.0300
(-1.58) (-2.14) (1.04) (0.11)
LOG_MKT -0.5703*** -0.6265*** -0.6554*** -0.8479***
(-24.75) (-21.70) (-7.12) (-10.25)
LEV 0.3023** 0.2278 -0.3286 -0.3550
(2.36) (1.53) (-0.36) (-0.86)
ROAL -0.2932** -0.2698 1.0041 0.5848
(-1.99) (-1.54) (1.40) (1.17)
ROA -0.9298*** -0.8906*** -2.2854* -2.1210***
(-4.83) (-3.90) (-1.94) (-2.85)
LOSS 1.0772*** 1.0299*** 1.8303*** 1.7608***
(12.53) (9.64) (3.30) (5.53)
CFO -0.5781*** -0.8187*** -0.0617 -0.3220
(-3.74) (-4.63) (-0.07) (-0.58)
BTM -0.8838*** -0.8464*** -1.4887** -1.3189***
(-10.69) (-8.89) (-2.00) (-4.22)
ABS(ACCRL) 0.9087*** 0.8791*** -0.2848 0.2768
(7.13) (5.71) (-0.41) (0.75)
GROWTH -0.2110*** -0.1917** -0.2216 -0.2634*
(-2.73) (-2.07) (-1.32) (-1.68)
ALTMAN -0.0183*** -0.0204*** -0.0851*** -0.0240
(-5.90) (-5.64) (-2.89) (-1.47)
STDEARN 0.0012*** 0.0015*** 0.0008 0.0001*
(6.91) (5.66) (1.24) (1.87)
TENURE -0.0962 -0.8863 0.0001 0.0001
(-0.33) (-1.27) (0.01) (0.01)
Intercept -0.2901 -15.0146*** 3.6700 2.6239**
(-0.53) (-12.69) (0.79) (2.52)
Year F.E. Included Included Included Included
Industry F.E. Included Included Included Included
N. Obs. 35,406 22,961 5,078 9,902
Pseudo-R2 0.48 0.49 0.40 0.46
This table presents the pooled analyses of propensity to issue a going-concern opinion for the full sample and matched samples, using
NLEAD1and CLEAD1 as definitions of auditor industry specialization. a Clients of specialist and non-specialist auditors are matched
based on industry, size, and stock returns covariance. Variable definitions are included in Appendix A. *, **, *** indicate significance
at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. T-statistics and p-values are calculated using clustered standard
errors by firm. For brevity, the year and industry-specific intercepts are not reported.
103
TABLE 9 – Going-Concern Analyses: Conditional Logistic Regression Tests
Dependent variable = GCONCERN
Matched samples a
I II
NLEAD1 NLEAD1
Predicted Estimate Estimate
sign (t-statistic) (t-statistic)
BIG4 0.4226 0.2801
(0.57) (0.45)
LOG_MKT -0.7570*** -0.7444***
(-2.77) (-3.62)
LEV -1.5445 0.0954
(-1.30) (0.11)
ROAL -1.4976 -0.0618
(-0.96) (-0.06)
ROA -0.5987 -1.6951
(-0.29) (-1.32)
LOSS 1.7702*** 1.7337***
(2.87) (2.74)
CFO -0.4037 -1.2199
(-0.17) (-0.91)
BTM -0.9670 -1.0829***
(-1.44) (-2.85)
ABS(ACCRL) -2.5354 -0.8835
(-1.20) (-0.83)
GROWTH 0.1051 -0.7950
(0.13) (-1.50)
ALTMAN -0.1130 -0.0970**
(-1.29) (-2.29)
STDEARN 0.0019 0.0028
(0.74) (1.57)
Intercept ( + ) 0.1758 -0.3481
(0.48) (-1.26)
N. Obs. 264 466
Pseudo-R2 0.58 0.64
This table presents the conditional logistic regression analyses of propensity to issue a going-concern opinion for the matched pairs
with intra-pair variation in going-concern opinions, using NLEAD1 and CLEAD1 as definitions of auditor industry specialization. . a Clients of specialist and non-specialist auditors are matched based on industry, size, and stock returns covariance. Variable
definitions are included in Appendix A. *, **, *** indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-
tailed tests. T-statistics and p-values are calculated using clustered standard errors by firm. Tenure, year, and industry-specific
intercepts are not included because there is insufficient intra-pair variation in these variables.
104
TABLE 10 – Clients that Switched from Arthur Andersen 2001–2002
PANEL A: Pre-Post Switch Analyses for Arthur-Andersen Clients
Dependent variable = ΔADA Dependent variable = ΔADREV
I II III IV
Predicted Estimate Estimate Estimate Estimate
sign (t-statistic) (t-statistic) (t-statistic) (t-statistic)
ΔNLEAD1 ( - ) -0.0081 -0.0057
(-0.79) (-1.16)
ΔCLEAD1 ( - ) 0.0115 -0.0011
(1.48) (-0.29)
ΔBIG4 0.0280* 0.0219 0.0055 0.0051
(1.77) (1.37) (0.73) (0.66)
ΔLOG_MKT -0.0020 -0.0012 0.0037 0.0044
(-0.33) (-0.20) (1.27) (1.50)
ΔLEV -0.1179*** -0.1194*** -0.0446*** -0.0435***
(-3.90) (-3.96) (-3.06) (-2.98)
ΔROAL -0.0343 -0.0361 0.0013 -0.0011
(-1.07) (-1.14) (0.08) (-0.07)
ΔROA -0.0520 -0.0511 -0.0385** -0.0366*
(-1.30) (-1.28) (-1.99) (-1.90)
ΔLOSS -0.0060 -0.0058 -0.0076* -0.0080*
(-0.63) (-0.61) (-1.66) (-1.73)
ΔCFO 0.0297 0.0294 -0.0127 -0.0134
(0.88) (0.87) (-0.78) (-0.82)
ΔBTM 0.0042 0.0044 0.0025* 0.0023
(1.37) (1.44) (1.73) (1.59)
ΔABS(ACCRL) -0.0661*** -0.0660*** -0.0481*** -0.0484***
(-3.13) (-3.14) (-4.73) (-4.75)
ΔGROWTH 0.0141 0.0139 0.0034 0.0034
(1.45) (1.43) (0.73) (0.72)
ΔALTMAN 0.0010 0.0010 0.0005 0.0004
(1.26) (1.30) (1.27) (1.14)
ΔSTDEARN 0.0001 0.0001 0.0001 0.0001
(0.30) (0.12) (0.94) (1.03)
Intercept -0.0036 -0.0053 -0.0055** -0.0056**
(-0.77) (-1.12) (-2.41) (-2.47)
N. Obs. 393 393 393 393
R2 0.11 0.12 0.13 0.12
This table presents the pre-post switch discretionary accruals and revenue manipulation analyses for the sample of Arthur Andersen
Clients, using NLEAD1 and CLEAD1 as definitions of auditor industry specialization. Variable definitions are included in Appendix A
and Δ denotes the difference between 2002 and 2001 for each observation. *, **, *** indicate significance at the 0.10, 0.05 and 0.01
levels, respectively, using two-tailed tests. T-statistics and p-values are calculated using clustered standard errors by firm.
105
PANEL B: Pre-Post Switch Analyses for Arthur-Andersen Clients and Matched Control Group
Dependent variable = ΔADA Dependent variable = ΔADREV
I II III IV
Predicted Estimate Estimate Estimate Estimate
sign (t-statistic) (t-statistic) (t-statistic) (t-statistic)
AA 0.0076 0.0051 0.0008 -0.0001
(0.95) (0.65) (0.25) (-0.03)
ΔNLEAD1 0.0105 0.0039
(0.63) (0.59)
AA_ΔNLEAD1 ( - ) -0.0131 -0.0119
(-0.59) (-1.34)
ΔCLEAD1 -0.0235 0.0050
(-1.52) (0.81)
AA_ΔCLEAD1 ( - ) 0.0324* -0.0028
(1.76) (-0.38)
ΔBIG4 0.0160 0.0114 0.0106 0.0091
(0.75) (0.53) (1.24) (1.05)
ΔLOG_MKT -0.0083 -0.0080 0.0042* 0.0048*
(-1.32) (-1.30) (1.68) (1.94)
ΔLEV -0.0838** -0.0850** -0.0253* -0.0255*
(-2.45) (-2.49) (-1.86) (-1.87)
ΔROAL -0.0123 -0.0120 0.0097 0.0086
(-0.32) (-0.31) (0.63) (0.56)
ΔROA -0.0599 -0.0573 -0.0193 -0.0189
(-1.43) (-1.37) (-1.16) (-1.13)
ΔLOSS -0.0187** -0.0177** -0.0046 -0.0043
(-2.12) (-2.03) (-1.30) (-1.22)
ΔCFO 0.0252 0.0243 0.0144 0.0135
(0.79) (0.76) (1.13) (1.05)
ΔBTM 0.0047 0.0051 0.0002 0.0002
(1.10) (1.20) (0.12) (0.13)
ΔABS(ACCRL) -0.1841*** -0.1886*** -0.0157 -0.0152
(-5.83) (-5.96) (-1.25) (-1.20)
ΔGROWTH -0.0006 -0.0007 0.0009 0.0010
(-0.12) (-0.16) (0.51) (0.53)
ΔALTMAN 0.0007 0.0007 -0.0001 -0.0001
(1.05) (1.12) (-0.28) (-0.37)
ΔSTDEARN -0.0001 0.0001 0.0001 0.0001
(-1.12) (0.96) (0.33) (0.58)
Intercept -0.0073 -0.0059 -0.0065*** -0.0065***
(-1.24) (-1.00) (-2.78) (-2.78)
N. Obs. 574 574 574 574
R2 0.11 0.11 0.04 0.03
This table presents the pre-post switch discretionary accruals and revenue manipulation analyses for the sample of Arthur-Andersen
Clients and a control group of comparable clients, using NLEAD1 and CLEAD1 as definitions of auditor industry specialization.
Variable definitions are included in Appendix A and Δ denotes the difference between 2002 and 2001 for each observation. *, **, ***
indicate significance at the 0.10, 0.05 and 0.01 levels, respectively, using two-tailed tests. T-statistics and p-values are calculated using
clustered standard errors by firm.
106
Table 11 –Heuristic Expectation Models
Descriptive Statistics
PANEL A: Full Sample
Variable Mean Median 25th
Perc. 75th
Perc. Std. Dev.
ARt 0.196 0.165 0.087 0.257 0.155
ARt-4 0.197 0.167 0.088 0.260 0.155
INVt 0.165 0.130 0.041 0.244 0.150
INVt-4 0.167 0.132 0.043 0.246 0.150
SALESt 0.324 0.282 0.171 0.415 0.250
COGSt 0.228 0.181 0.096 0.295 0.214
CHSALESt 1.114 1.066 0.947 1.206 0.373
CHCOGSt 1.123 1.068 0.936 1.223 0.407
ATt 4,691 218.7 41.71 1,242 39,875
N. Obs. 53,215
Panel B: Sub-Sample with Peer firmsa
Variable Mean Median 25th
Perc. 75th
Perc. Std. Dev.
ARt 0.193 0.164 0.089 0.251 0.150
ARt-4 0.195 0.166 0.090 0.253 0.151
INVt 0.163 0.132 0.046 0.238 0.144
INVt-4 0.165 0.133 0.047 0.240 0.144
SALESt 0.318 0.279 0.173 0.406 0.228
COGSt 0.222 0.178 0.096 0.286 0.202
CHSALESt 1.106 1.068 0.960 1.194 0.324
CHCOGSt 1.117 1.070 0.950 1.212 0.365
ATt 5,145 335.4 72.44 1,755 38,487
jARt 0.192 0.161 0.087 0.249 0.151
jARt-4 0.193 0.163 0.087 0.252 0.151
jINVt 0.165 0.132 0.045 0.240 0.146
jINVt-4 0.166 0.134 0.046 0.242 0.146
jSALESt 0.315 0.276 0.168 0.407 0.227
jCOGSt 0.219 0.176 0.093 0.287 0.198
jCHSALESt 1.108 1.068 0.956 1.199 0.333
jCHCOGSt 1.120 1.071 0.947 1.217 0.375
jATt 4,484 299.4 63.01 1,532 33,319
PEER-R2 0.160 0.118 0.056 0.218 0.149
PEER-TAU 0.172 0.207 0.060 0.311 0.216
N. Obs. 33,152
This table presents the descriptive statistics of the data used in the analyses of heuristic expectation models: Panel A shows descriptive
statistics for the full sample; Panel B shows descriptive statistics for the sample of companies with a comparable peer firm. a Peer firms
are selected using a pair-wise match based on industry, size, and stock returns covariance. Variable definitions are included in Appendix
D.
107
Table 12 –Heuristic Expectation Models
Correlation Table
Sub-sample with Peer Firms a
Variable ARt ARt-4 INVt INVt-4 COGSt SALESt CHSALESt CHCOGSt jARt jARt-4 jINVt jINVt-4 jCOGSt jSALESt jCHSALESt
ARt-4 0.952
<0.001
INVt 0.030 0.029
<0.001 <0.001
INVt-4 0.031 0.030 0.958
<0.001 <0.001 <0.001
COGSt 0.135 0.115 0.374 0.371
<0.001 <0.001 <0.001 <0.001
SALESt 0.126 0.103 0.427 0.425 0.933
<0.001 <0.001 <0.001 <0.001 <0.001
CHSALESt 0.016 -0.043 -0.035 -0.033 0.034 0.064
0.004 <0.001 <0.001 <0.001 <0.001 <0.001
CHCOGSt 0.044 -0.002 -0.036 -0.040 0.042 0.024 0.749
<0.001 0.785 <0.001 <0.001 <0.001 <0.001 <0.001
jARt 0.595 0.595 0.031 0.032 -0.042 -0.055 -0.004 0.030
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.463 <0.001
jARt-4 0.591 0.594 0.034 0.035 -0.039 -0.053 -0.014 0.021 0.951
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.009 <0.001 <0.001
jINVt 0.027 0.027 0.514 0.517 0.287 0.336 -0.005 -0.016 0.030 0.028
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.390 0.005 <0.001 <0.001
jINVt-4 0.029 0.031 0.517 0.522 0.287 0.336 -0.007 -0.018 0.032 0.032 0.957
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.178 0.001 <0.001 <0.001 <0.001
jCOGSt -0.056 -0.053 0.337 0.342 0.408 0.457 -0.003 -0.020 0.139 0.117 0.430 0.426
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.572 0.000 <0.001 <0.001 <0.001 <0.001
jSALESt -0.043 -0.040 0.289 0.292 0.394 0.415 -0.005 -0.018 0.145 0.127 0.380 0.376 0.940
<0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.403 0.001 <0.001 <0.001 <0.001 <0.001 <0.001
jCHSALESt -0.006 -0.015 0.003 0.001 0.002 0.005 0.095 0.076 0.013 - 0.048 -0.029 -0.024 0.067 0.037
0.287 0.006 0.576 0.802 0.653 0.322 <0.001 <0.001 0.015 <0.001 <0.001 <0.001 <0.001 <0.001
jCHCOGSt 0.016 0.008 -0.012 -0.016 -0.018 -0.018 0.081 0.089 0.036 - 0.012 -0.030 -0.030 0.030 0.037 0.750
0.003 0.152 0.024 0.004 0.001 0.001 <0.001 <0.001 <0.001 0.029 <0.001 <0.001 <0.001 <0.001 <0.001
This table presents the Pearson correlation between variables and their t-statistic used in the analyses of heuristic expectation models for the sample of companies with a closest peer
firm. a Peer firms are selected using a pair-wise match based on industry, size, and stock returns covariance. Variable definitions are included in Appendix D.
108
Table 13 –Heuristic Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
Heuristic Specification Tests Power Tests
Models Samples Without Manipulation Samples with Directional Manipulation (1% Total Assets)
Increasing Receivables and Decreasing Inventory
Abs. Forecast Error Forecast Error Specification Error
% Forecast Error Detection Power
I II III IV V VI VII VIII IX X XI
Mean Median Mean Median Std. Dev. (-) Sig (+) Sig Mean Median Std. Dev. Sig
No peers
AR1-4qtr 0.051 0.050 -0.022 -0.021 0.008 88.4% 0.0% -0.034 -0.032 0.013 0.0% AR2-4qtr 0.037 0.036 -0.002 -0.002 0.005 6.0% 0.0% 0.008 0.008 0.005 39.2% INV1-4qtr 0.051 0.050 -0.020 -0.020 0.008 80.4% 0.0% -0.017 -0.017 0.009 64.8% INV2-4qtr 0.031 0.031 -0.003 -0.003 0.005 11.6% 0.8% -0.013 -0.013 0.005 82.0%
With Peers AR1
P-4qtr 0.064 0.064 -0.000 0.000 0.010 1.6% 0.2% -0.012 -0.011 0.013 0.4%
AR2P -4qtr 0.047 0.047 -0.000 -0.000 0.006 0.2% 1.6% 0.010 0.010 0.006 35.2%
INV1P-4qtr 0.067 0.067 0.000 0.001 0.009 2.4% 1.6% 0.003 0.003 0.010 1.6%
INV2P-4qtr 0.042 0.042 -0.001 -0.001 0.006 2.4% 0.2% -0.011 -0.011 0.006 45.2%
No peers AR1-2qtr 0.051 0.050 -0.024 -0.022 0.009 80.0% 0.0% -0.036 -0.034 0.013 0.0% AR2-2qtr 0.034 0.033 -0.003 -0.003 0.005 7.2% 0.4% 0.007 0.007 0.005 35.2% INV1-2qtr 0.052 0.051 -0.021 -0.021 0.010 68.0% 0.0% -0.017 -0.017 0.011 49.6% INV2-2qtr 0.029 0.029 -0.002 -0.002 0.005 6.0% 1.2% -0.012 -0.012 0.005 72.4%
With Peers AR1
P-2qtr 0.064 0.063 -0.000 -0.001 0.011 2.4% 0.2% -0.012 -0.011 0.015 0.0%
AR2P-2qtr 0.044 0.044 -0.000 -0.001 0.006 2.8% 2.4% 0.010 0.009 0.006 36.0%
INV1P-2qtr 0.067 0.067 0.000 0.001 0.011 0.4% 1.6% 0.003 0.004 0.011 0.2%
INV2P-2qtr 0.039 0.038 -0.001 -0.001 0.006 2.4% 1.6% -0.011 -0.011 0.006 46.0%
This table presents the analyses of the heuristic expectation models. Peer firms are selected using a pair-wise matching approach based on industry, size, and stock returns covariance.
The results on this table are calculated following the steps outlined in Appendix F. Specification Error is the percentage of 250 random samples of 100 firms in which the mean of
year-firm-level errors is statistically greater and lower than zero at five percent level. Detection Power is the percentage of 250 random samples of 100 firms in which the mean of
year-firm-level errors is statistically greater or lower than zero at five percent level in the direction of the manipulation. AR1 is the receivables model using the incremental approach
and AR2 is the receivables model using the historical approach. INV1 is the inventory model using the incremental approach and INV2 is the inventory model using the historical
approach. The top four rows (AR1-4qtr to INV2-4qtr) show the results for each model without peers, using the average error of four quarters for each firm. The following four rows
(AR1P-4qtr to INV2P-4qtr) show the results for each model with one peer for each audit client, using the average error of four quarters for each firm. Similarly the bottom eight rows
show the results, with and without peers, using the average error of two quarters for each audit client. Variables and model definitions are included in Appendices D and E.
109
Table 14 –Time Series Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
Time-Series Specification Tests Power Tests
Models Samples Without Manipulation Samples with Directional Manipulation (1% Total Assets)
Increasing Receivables and Decreasing Inventory
Abs. Forecast Error Forecast Error Specification Error
% Forecast Error Detection Power
I II III IV V VI VII VIII IX X XI
Mean Median Mean Median Std. Dev. (-) Sig (+) Sig Mean Median Std. Dev. Sig
No peers
AR3-4qtr 0.041 0.038 -0.002 -0.002 0.013 4.8% 0.4% 0.006 0.006 0.013 16.4% INV3-4qtr 0.039 0.036 -0.003 -0.004 0.015 6.4% 0.4% -0.013 -0.013 0.015 54.8%
With Peers AR3
P-4qtr 0.042 0.039 -0.001 -0.001 0.012 3.6% 1.2% 0.007 0.007 0.012 22.0%
INV3P-4qtr 0.042 0.036 -0.007 -0.003 0.040 3.6% 0.8% -0.016 -0.013 0.040 58.8%
No peers AR3-2qtr 0.035 0.034 -0.002 -0.002 0.009 4.4% 0.0% 0.005 0.005 0.009 15.2% INV3-2qtr 0.032 0.031 -0.003 -0.003 0.009 6.0% 0.8% -0.012 -0.012 0.009 56.4%
With peers AR3
P-2qtr 0.035 0.033 -0.001 -0.001 0.010 3.6% 1.2% 0.007 0.007 0.010 17.6%
INV3P-2qtr 0.035 0.030 -0.003 -0.002 0.030 4.0% 0.0% -0.013 -0.012 0.030 58.4%
This table presents the analyses of the time-series expectation models. Peer firms are selected using a pair-wise matching approach based on industry, size, and stock returns
covariance. The results on this table are calculated following the steps outlined in Appendix F. Specification Error is the percentage of 250 random samples of 100 firms in which the
mean of year-firm-level errors is statistically greater and lower than zero at five percent level. Detection Power is the percentage of 250 random samples of 100 firms in which the
mean of year-firm-level errors is statistically greater or lower than zero at five percent level in the direction of the manipulation. AR3 is the receivables model estimated in time-series,
using the 16 quarters prior to the year under audit and fitted to each quarter of the year under audit. INV3 is the inventory model estimated in time-series, using the 16 quarters prior to
the year under audit and fitted to each quarter of the year under audit. The top two rows (AR3-4qtr and INV3-4qtr) show the results for each model without peers, using the average
error of four quarters for each firm. The following two rows (AR3P-4qtr and INV3P-4qtr) show the results for each model with one peer for each audit client, using the average error of
four quarters for each firm. Similarly the bottom four rows show the results, with and without peers, using the average error of two quarters for each audit client. Variables and model
definitions are included in Appendices D and E.
110
Table 15 –Industry Cross-Sectional Expectation Models
Simulation Results –Tests of Model Specification and Detection Power
Cross- Specification Tests Power Tests
Sectional Samples Without Manipulation Samples with Directional Manipulation (1% Total Assets)
Models
Increasing Receivables and Decreasing Inventory
Abs. Forecast Error Forecast Error Specification Error Forecast Error Detection Power
I II III IV V VI VII VIII IX X XI
Mean Median Mean Median Std. Dev. (-) Sig (+) Sig Mean Median Std. Dev. Sig
No peers
AR4-4qtr 0.051 0.051 -0.000 -0.000 0.007 3.2% 1.2% 0.010 0.010 0.007 31.6% INV4-4qtr 0.045 0.044 -0.000 -0.000 0.006 3.6% 1.2% -0.010 -0.010 0.006 36.0%
With Peers AR4
P-4qtr 0.027 0.027 -0.000 -0.000 0.003 4.8% 0.2% 0.009 0.009 0.003 80.0%
INV4P-4qtr 0.024 0.024 -0.000 -0.000 0.003 4.0% 0.8% -0.010 -0.010 0.003 91.2%
No peers AR4-2qtr 0.050 0.050 -0.000 0.000 0.008 4.4% 1.2% 0.010 0.010 0.008 28.0% INV4-2qtr 0.044 0.044 -0.000 0.000 0.007 3.6% 0.4% -0.010 -0.010 0.007 34.4%
With peers AR4
P-2qtr 0.027 0.027 -0.000 0.000 0.004 2.4% 1.2% 0.009 0.009 0.004 65.6%
INV4P-2qtr 0.024 0.024 -0.000 -0.000 0.003 3.2% 0.8% -0.010 -0.010 0.003 78.8%
This table presents the analyses of the industry cross-sectional expectation models. Peer firms are selected using a pair-wise matching approach based on industry, size, and stock
returns covariance. The results on this table are calculated following the steps outlined in Appendix F. Specification Error is the percentage of 250 random samples of 100 firms in
which the mean of year-firm-level errors is statistically greater and lower than zero at five percent level. Detection Power is the percentage of 250 random samples of 100 firms in
which the mean of year-firm-level errors is statistically greater or lower than zero at five percent level in the direction of the manipulation. AR4 is the receivables model estimated in
cross-section by industry for each quarter of the year under audit. INV4 is the receivables model estimated in cross-section by industry for each quarter of the year under audit. The top
two rows (AR4-4qtr and INV4-4qtr) show the results for each model without peers, using the average error of four quarters for each firm. The following two rows (AR4P-4qtr and
INV4P-4qtr) show the results for each model with one peer for each audit client, using the average error of four quarters for each firm. Similarly the bottom four rows show the results,
with and without peers, using the average error of two quarters for each audit client. Variables and model definitions are included in Appendices D and E.