TAX NON-COMPLIANCE AND CORPORATE GOVERNANCE:
A COMPARATIVE STUDY*
Lindsay M. Tedds
Department of Economics, University of Manitoba
503 Fletcher Argue Building
Winnipeg, Manitoba, Canada, R3T 5V5
Email: tedds@cc.umanitoba.ca
Phone: (204) 474-9275
This version: May 2006
Abstract
It is generally accepted that taxes and tax evasion are intrinsically linked. While there exists an
extensive knowledge base regarding tax compliance by individuals, very little is actually known
about tax compliance by businesses due to a lack of detailed and readily available data. This
paper uses a unique and recently available dataset that contains detailed information on over
10,000 firms (broadly defined) from around the world to investigate factors, particularly
corporate governance, that affect business tax compliance. The empirical strategy employed
exploits the nature of the dependent variable, which is interval coded, and uses interval
regression which provides an asymptotically more efficient estimator than the ordered probit,
provided that the classical linear model assumptions hold. These assumptions are investigated
using standard diagnostic tests that have been modified for the interval regression model.
Evidence is presented that shows that firms in all regions around the world engage in tax noncompliance, but that there is substantial variation across and within regions. The legal
organization of a business is found to affect tax compliance, as does size, ownership, competition
and audit controls but industry and age has no effect. Not surprisingly, high taxes and
government corruption result in reduced compliance while access to financing, organized crime,
political instability and the fairness of the legal system do not affect compliance levels.
Keywords: Underground Economy, Tax Non-compliance, Firm Characteristics, Interval
Regression
JEL Classification: C24, D21, O17
*
I would like to thank Tom Crossley, Mike Veall and David Bjerk, all of McMaster University, David Giles,
University of Victoria, Wayne Simpson, University of Manitoba, seminar participants at the University of Victoria,
and participants at the 2005 Tax Policy: Perspectives from Law and Accounting conference for their helpful
comments and invaluable guidance, Hassan Azziz for his assistance with the literature review, and Adian McFarlane
for his diligent research support. The author would also like to gratefully acknowledge the financial support from the
Social Sciences and Humanities Research Council (#752-2004-1096 and #501-2002-0107) and the University of
Manitoba. I am solely responsible for any remaining errors and omissions.
TAX NON-COMPLIANCE AND CORPORATE GOVERNANCE:
A COMPARTATIVE STUDY
1
Introduction
It is generally accepted that taxes and tax evasion are intrinsically linked; one cannot exist
without the other. As a result of a great deal of theoretical, experimental, and empirical research
conducted over the last twenty years, there exists an extensive knowledge base regarding tax
evasion by individuals. However, research regarding tax evasion1 by businesses is, by comparison,
surprisingly modest. This is startling, given the importance of businesses and their decisions not
only in economic models but also in tax systems and the economy as a whole.
There is some evidence to suggest that there is cause for concern, that a substantial share of
business income goes unreported to the relevant tax authorities. The United States Internal
Revenue Service (IRS) routinely estimates the total amount of under-reported income and
overstated deductions, and calculates the total loss of tax revenue, or the “tax gap”. The latest data
from the IRS (2004) regarding the “tax gap” related to business activities are for the 2001 tax year.
These estimates, denoted in U.S. dollars, indicate that: (1) the corporate tax gap amounted to $29.9
billion, of which corporations with over $10 million in assets contributed $25.0 billion; (2) the tax
gap associated with business income earned by individuals amounted to $81.2 billion; (3) the selfemployed evaded $61.2 billion in employment tax; and (4) corporations underpaid the amount of
taxes due based on reported income by $2.3 billion. In total, businesses evaded $174.6 billion in
taxes, which amounted to almost 10% of total taxes paid voluntarily. This is not an insignificant
1
The focus of this paper is on illegal tax evasion and not legal tax avoidance. Tax evasion or tax non-compliance
refers to income tax that is legally owed but is not reported or paid whereas tax avoidance refers to legal actions taken
to reduce tax liability. Tax avoidance includes such activities as “…purchasing tax-exempt bonds, which is certainly
legal, not at all nefarious, but also certainly done for tax reasons.” (Slemrod 2004, 4)
2
amount and yet should be considered a lower bound estimate because: it is based on twenty year
old compliance rates; it does not include businesses that do not file tax returns (also known as nonfilers, the hard-to-tax, ghosts or informal businesses); and/or it does not consider firms that are
engaged in illegal activities. While these data are for the United States, it is not unreasonable to
assume that firms evade taxes in every country around the world.
Given that there is some evidence which supports the notion that businesses engage in tax
non-compliance, several questions arise and require investigation. First, do businesses around the
world engage in tax evasion or is it confined to a few countries or regions? Second, does the legal
status of the business (e.g. sole proprietorship, partnership, corporation, etc.) affect the incidence
and/or intensity of tax evasion? If so, then it may be possible to effect changes in the legal system
in order to increase tax compliance. Third, do businesses that engage in tax non-compliance share
common and observable characteristics, or is there too wide a variety of shapes and sizes to permit
a useful generalization about them? If it were possible to define a typical evader, the tax authority
could target their auditing activities more accurately. Finally, while there is considerable
agreement internationally about the factors that trigger tax non-compliance (e.g. the tax burden, the
degree of regulation, the level of enforcement, confidence in government, labour force
characteristics, and morality), how do these features influence the intensity of non-compliance?
With this information, policy makers could effect changes to increase the amount of tax revenues
collected from businesses.
One of the main constraints to investigating and attempting to provide answers to these and
related questions is the lack of data. Previously, the only data sources available are from tax audits.
However, these sources are only available for a very small number of select countries, the data is
costly to collect, and access to the data is limited. In addition, tax audit data is subject to both
3
selection and measurement error bias. The selection effect occurs because firms are generally not
randomly audited. Rather, firms are usually selected for an audit because they have a particular
characteristic (e.g. U.S. businesses with more than $10 million in assets are audited annually) or
because their tax filing documents raise a “red-flag” with the tax authority. The measurement error
bias occurs because not only are audits unlikely to completely and accurately measure tax noncompliance but also the final non-compliance report is usually based on negotiations between the
firm and the tax authority as well as various legal appeals.
More recently, however, an alternative data source has become available that is conducive
to investigating issues related to business tax non-compliance. The World Business Environment
Survey (WBES) was launched in 1999 by the World Bank’s Investment Climate and Institute
Units. The survey was administered to more than 10,000 firms in eighty countries in late 1999 and
early 2000, and provides information regarding a firm’s characteristics and well as responses to
multiple questions on the investment climate and business environment. The survey also includes a
question regarding the extent and intensity to which firms fail to report income to the tax authority.
Rather than reporting the exact amount of income that went unreported to the tax authority, firms
were provided with seven possible intervals of varying size that were grouped according to the
percentage of sales that are reported for tax purposes. When a quantitative outcome is grouped
into known intervals on a continuous scale the data is said to be interval-coded. Asymptotically
consistent estimates can be obtained using interval regression. The advantage of interval regression
is that, provided certain key assumptions are met, the coefficients on the exogenous variables can
be interpreted as if ordinary least squares were applied to a continuous dependent variable.
Overall, the findings reveal that firms which are sole proprietorships and privately owned
corporations report a smaller percentage of their sales to the tax authority. The previous literature
4
has suggested that firms in the service and construction sector should be less compliant but this
study finds that industry sector has no effect on compliance. There is no consensus in the previous
literature about the relationship between firm size and under-reporting, but it is found,
unambiguously, that small firms are less and large firms are more compliant. Foreign owned firms,
government owned firms, and firms that have audited financial statements are also more compliant,
as was also found by other researchers. Not surprisingly, high taxes and government corruption
result in lower compliance and the magnitude of these effects are quite large. Access to capital,
political instability, organized crime, inflation, and the legal system are found to have no
statistically significant effect on tax non-compliance.
The paper begins with a brief review of the relevant tax non-compliance literature. This
literature makes it clear that a businesses legal organization likely affects the decision to evade
taxes but other factors, such as firm size, have ambiguous effects or, at the very least, vary across
countries. Therefore, it seems worthwhile to conduct a worldwide study of firms’ tax compliance
behaviour. The WBES is then described and the rationale for the empirical techniques is outlined.
The results are then summarized and the paper ends with some concluding comments.
2
Literature Review
In this section, attention is focused on two critical aspects of the literature regarding tax
evasion. First, a review of the development of the theoretical literature and the associated
predictions is provided, commencing with the classical model of an individual’s decision to evade
taxes and how the model has been modified for firm behaviour. Second, there have been a few
empirical papers that explore firm tax non-compliance, and this literature is summarized along with
the key findings. This literature will help shape the empirical model utilized in this paper.
5
2.1 Theoretical Studies
As previously noted, extensive literature exists on tax evasion by individuals and the
seminal contribution was provided by Allingham and Sandmo (1972).2 Their model was sparse but
surprisingly robust in modeling an individual’s decision to evade taxes and, if so, how much to
evade. The model leads to four propositions about the incidence of tax evasion: (1) the rate of
return to evasion is positively related to the incidence of tax evasion; (2) individuals with higher
risk aversion tend to evade less; (3) individuals with higher personal income tend to evade more;
and (4) and compliance is positively related to the probability of being audited and the size of the
penalty if caught.3 The first and third propositions leads one to conclude that, in the context of
business tax non-compliance, the self-employed, sole-proprietors and other small businesses would
be less likely to evade taxes than large businesses. The fourth proposition leads to the opposite
conclusion given that large businesses and corporations are normally subjected to audits at a higher
rate than are smaller businesses.4 However, the true effect of these propositions is muddied when
the choice is not only the level of evasion but also the level of output, as with most businesses, and
when there is no direct link between the owner’s income, risk preference, audit probability, and
evasion decisions. This is particularly true for businesses where evasion decisions are made by
managers and not the owner(s).
2
The Allingham-Sandmo (1972) model has been extended in a number of dimensions over the last thirty years (e.g.
Watson 1985, Trandel and Snow 1999, Mookherjee and P’ng 1989, Border and Sobel 1987, Scotchmer 1987, Cremer,
Marchand and Pestieau 1990, and Sanchez and Sobel (1993)) and this literature is nicely surveyed by Andreoni, Erard,
and Feinstein (1998) and Slemrod and Yitzhaki (2002).
3
The relationship between tax rates and tax evasion, however, is uncertain. A higher tax rate implies less income
which, according to the model, implies less evasion. On the other hand, an increase in the tax rate implies that the
“return” from evasion increases as well, assuming a constant penalty, which implies more evasion. As a result, the two
forces are of opposite tendency which makes the net result uncertain.
4
For example, in the United States corporations with more than $10 million in assets face an audit rate of nearly 100
percent.
6
Given the above discussion, the reasons driving many businesses to evade taxes is likely
different from that of individuals and, hence, should be modeled differently. There is a much
smaller pool of literature that addresses tax non-compliance by businesses.5 The majority of this
work utilizes the basic framework of the Allingham-Sandmo (1972) model, but the key variation in
these models is that they explore how the tax rate, probability of detection, and penalty rate affect
the two choices of evasion and output.6 In these models, it is generally found that: (1) reported
sales decrease as the tax rate increases; (2) a rise in taxes increases the market price of the good, but
by less than the amount of the tax, since some of the tax increase is absorbed through increased
evasion; and (3) an increased probability of detection or penalty increases the proportion of sales
declared and the market price of the good. As opposed to the propositions of the AllinghamSandmo (1972) model, there is an unambiguous positive relationship between the tax rate and tax
evasion. Further, it should be noted that these results hold, regardless of whether the market, in
which the firm is operating, is competitive or monopolistic.
The preceding literature, however, continues to assume that an individual is at the centre of
the tax decision. That is, the above noted firm models assume that the firm owner makes the tax
reporting decision. This assumption, however, likely only applies to the self-employed and other
small, privately-held businesses. This suggests that the outcomes predicted by the AllinghamSandmo (1972) model may not apply to other types of businesses, notably those businesses where
financial decisions, including those related to taxes, are not made by the owner/shareholder but,
rather, by their agents. With this in mind, Chen and Chu (2002) extend the standard model to
include a firm that hires a risk-averse manager who is offered some form of compensation (e.g.
stock options, bonuses, etc.) as an incentive to engage in tax evasion on behalf of the firm. As
5
6
Cowell (2004) provides an excellent review of this literature.
There is no agreement in this literature if the firm in these models should be modeled as risk-averse or risk neutral.
7
opposed to the finding of the Allingham-Sandmo (1972) model, that an individual will only evade
taxes if the rate of return from evasion is greater than the rate of return from compliance, Chen and
Chu’s model implies that a firm will evade tax only when the expected profit from evasion is
significantly greater. This is because tax evasion by a business actually involves the interaction of
many agents and is much more complicated than individual income tax evasion.
The principal-agent model proposed by Chen and Chu, however, does not consider the role
of penalties in the tax evasion decision. To address this shortcoming Crocker and Slemrod (2005)
propose a model of “…corporate tax evasion in the context of the contractual relationship between
the shareholders of a firm and the chief financial officer (CFO), who determines the firm’s
deductions from taxable corporate income.” (p. 1594) The model implies that corporate tax
evasion is reduced when penalties are imposed on the CFO directly, as opposed to the shareholder,
and that tax evasion increases if the CFO’s compensation contract optimally adjusts to offset the
penalties imposed when evasion is detected.
2.2 Previous Empirical Studies of Business Tax Non-Compliance
Unfortunately, the empirical analysis of business tax evasion is not extensive, mainly due to
a lack of data. There have, however, been a small number of empirical studies that will be
summarized here.
One of the first empirical examinations into business tax non-compliance focused on the
self-employed.7 The self-employed are commonly believed to have lower compliance rates than
wage and salary earners, primarily due to a lack of third-party reporting and withholding. Smith et
7
A self-employed person can register their business in a variety of forms, including a corporation, which, based on the
theoretical literature, likely affects their tax compliance behaviour. Unfortunately, only Apel (1994) explored the
relationship between tax non-compliance and the legal form of the business, largely due to the lack of information on
the latter in the data employed by these studies.
8
al. (1986) obtain estimates that indicate that in 1982 the self-employed in Great Britain understated
their income in the range of 30 to 36%. Pissarides and Weber (1989), using the same data, improve
on Smith et al.’s approach and find that the self-employed under-reported their income by up to
90%. This modified approach was subsequently applied by: (1) Apel (1994) who estimates that the
self-employed in Sweden under-reported their income by 25% in 1988, though this figure rises to
35% for the self-employed who own firms that are unincorporated (which supports the proposition
that the legal organization of the business does affect the tax evasion decision); (2) Mirus and
Smith (1997) concentrate their analysis on Canada and obtain an estimate of 12.5% for the year
1990; and (3) Schuetze (2002), who also applies the approach to Canada, finds that the selfemployed under-reported their income by between 11 and 23% over the period from 1969 to 1992
and that the construction and service occupations are more likely to be involved in tax noncompliance. Lyssiotou et al. (2004) propose further modifications and conclude that the selfemployed in Great Britain in 1993 under-reported their income by 118% if they were in blue collar
occupations and 64% if they were in white collar occupations. Finally, Tedds (2005) introduces a
nonparametric framework to the approach and found that the gap between true and reported selfemployment income is larger for households at the lower end of the self-employment income
distribution, a result which runs contrary to the theoretical prediction of the Allingham-Sandmo
(1972) model.
Smith and Adams (1987) examine the extent of tax non-compliance by informal suppliers in
six “at risk” sectors8 using results from a survey commissioned by the U.S. Internal Revenue
Service (IRS) on the expenditures made by consumers on goods and services provided by these
suppliers. For the 1992 tax year, the authors find that unreported income by these informal
8
They were: home repairs and additions, domestic services, auto repair, music lessons, appliance repairs, cosmetic
services, and catering.
9
suppliers amounted to approximately US$59.6 billion and that informal suppliers tend to report
only about 20% of their net business income.
Several studies have investigated business tax non-compliance using data from tax audits.
Probably the most comprehensive tax audit dataset in the world is the U.S. IRS Tax Compliance
Measurement Program (TCMP).9 Rice (1992) used TCMP data from 1980 to investigate tax
compliance by small corporations (defined as corporations with assets between $1 and $10
million). His findings suggest four key results. First, compliance is higher among publicly traded
corporations, which he attributes to the requirement that publicly traded corporations must disclose
more information to the public about their operations.10 Second, high profit companies are more
likely to under-report their income, and corporations whose profits are below the industry mean
tend to resort to non-compliance, perhaps as a means of limiting costs. Third, the marginal tax rate
is negatively associated with compliance. Finally, firm size and tax non-compliance are positively
related, and corporations engaging in tax non-compliance appear to be geographically bundled.
The latter finding is interesting and may be evidence that the compliance behaviour of others
directly affects ones own compliance behaviour. Joulfaian (2000), using TCMP data from 1987,
found that non-compliant corporations are three time more likely to be managed by executives who
have evaded personal taxes. TCMP data has also been used to explore tax evasion by the selfemployed and these studies include Christian (1994), Erard (1992), and Joulfaian and Rider (1998).
Hanlon et al. (2005) are the first to the use operational data from the Voluntary Compliance
Baseline Measurement (VCBLM) program compiled by the Large and Mid-Sized Business
9
The TCMP features data from a random sample of individual and small corporate income tax returns filed in a given
year that were subject to intensive audits by experienced examiners. For certain groups, such as non-filers and
proprietors who tend to not report a significant amount of their income, the results from special research studies are
used to supplement TCMP data. Finally, data for large corporations are obtained from routine operational audits.
10
Tannenbaum (1993), however, disagrees with this statement and instead argues that higher compliance could be the
result of managers in these corporations having greater independence from the owners.
10
(LMSB) Research Division of the IRS to examine corporate tax non-compliance, as measured by
deficiencies identified during audit. This data is more up to date, containing audit files up to and
including 2002, than the TCMP, which ended in 1988. They find that business tax non-compliance
relative to scale is a convex function, with medium-sized businesses (within the population of firms
with assets exceeding $10 million) having the lowest rate of non-compliance. Private firms,
multinationals and firms with incentivized executive compensation schemes are associated with
higher non-compliance (providing empirical evidence for the predictions from the model proposed
by Crocker and Slemrod (2005)) while foreign controlled firms are more compliant. In general,
firms in the manufacturing industry, trade, transportation and warehousing industry, and education,
healthcare and warehousing industry are less compliant. They also find no association between
compliance and effective tax rates.
Giles (2000) discusses some of the factors that determine the probability of non-compliance
among a very large population of New Zealand businesses that were audited by the Inland Revenue
Department in that country between 1993 and 1995. Contrary to Rice (1992) and Hanlon et al.
(2005), Giles finds that an increase in the scale of the business, regardless of how this is measured,
unambiguously raises the probability of compliance, once other characteristics are controlled for.
That is, businesses that are relatively small, in terms of sales revenue or before-tax profit, are more
likely to evade taxes than are large corporations, all other things being equal. Businesses in the
construction, wholesale trade, retail trade, accommodation, and cafes and restaurants sectors
exhibited below-average compliance rates over the study period. He also considers several other
characteristics that were found to be important in reducing the tax compliance rate among New
Zealand businesses. Relatively inefficient businesses tended to be less compliant than more
efficient ones, where efficiency was defined as either “return on net assets”, or “activity ratio”
11
(sales as a percentage of net assets). In addition, businesses which were registered off-shore were
generally more compliant than their on-shore counterparts, a finding that is shared by Hanlon et al.
(2005). Finally, it was found that in general, an aggressive use of legitimate tax-minimization
instruments (i.e. tax avoidance behaviour such as the deduction of interest and depreciation costs,
and the writing-off of bad debts) tended to be associated with increased compliance.
Many countries use tax holidays to attract foreign investment by providing a limited period
of tax exemptions and reductions for qualified investors. Chan and Mo (2000) examine the effect
of tax holidays on foreign investors’ tax non-compliance behaviour in China. They analyzed 583
tax audit cases, made available by the Chinese tax authorities, on corporate tax non-compliance by
foreign investors. Their results indicate that the corporate taxpayers tax holiday position
significantly affects non-compliance, notably: (1) companies in the pre-holiday position are least
compliant; (2) companies are most compliant in the tax exemption period that has a zero tax rate
and a heavy penalty for evasion; (3) domestic market-oriented companies have a higher rate of noncompliance than their export-oriented counterparts; and (4) wholly foreign-owned and
manufacturing-oriented companies have higher compliance than joint ventures and service-oriented
companies.
Only two studies appear to exist which use data on firm-reported tax non-compliance.
Using firm level data from a 1997 survey of private manufacturing firms in Poland, Romania,
Russia, Slovakia, and the Ukraine, Johnson et al. (2000) investigate the relationship between
government corruption, criminal activities, and firm tax compliance. The dependent variable
(percentage of sales that are unreported to the tax authority) is similar to that which is used in this
12
paper11 (discussed below), except that it is a continuous variable (rather than grouped data as in the
WBES). The authors found a positive and significant relationship between under-reporting of sales
and bribing of corrupt officials, but no relationship between under-reporting of sales and protection
payments to the mafia, tax payments, or efficiency of the legal system. Finally, firm tax noncompliance was greater in Russia and the Ukraine than in the other countries included in the study.
Batra et al. (2003) were the first to use the WBES to investigate the determinants of underreporting by firms. The WBES asks each firm to provide an estimate of the percentage of sales
revenues that firms like their own report to the tax authority. As indicated previously and discussed
in more detail later in this paper, seven answers were possible: all, 90-99%, 80-89%, 70-79%, 6069%, 50-59%, and less than 50%. The authors define seven binary variables based on these seven
intervals and each binary variable takes a value of one if the firm selected the particular interval of
interest and zero otherwise. They then estimate seven different OLS regressions which include
variables for firm characteristics, rule of law, business constraints and country fixed effects.
Overall they find that: (1) “…small or medium-size firms that produce for the domestic market
(non-exporters), lack foreign investment, and are located in large cities (but not necessarily in the
capital) tend to engage more in unofficial activity” (p. 76); (2) the prevalence, though not the
unpredictability, of corruption also significantly affects non-compliance; and (3) “…a firm’s age,
sector or mode of ownership do not influence [a firm’s] under-reporting of revenue.” (p. 78).
There are several reasons to be concerned about these reported results. First, the number of
observations in each of the seven regressions ranged from a low of 3,802 to a high of 4,781 from a
total number of observations of over 10,000 available in the WBES. The authors do not discuss the
11
The question regarding the percentage of sales not reported was prefaced by “It is thought that many firms in your
industry, in order to survive and grow, may need to misreport their operational and financial results. Please estimate
the degree of under-reporting by firms in your area of activity.” This is similar wording to that which appears in
question regarding tax non-compliance in the WBES.
13
reasons for the attrition but it is likely the result of missing observations and non-response for both
the dependent and independent variable. As attrition of this magnitude can lead to biased results, it
is important that it be investigated further. Second, the authors do not link their choice of
explanatory variables to the existing theoretical or empirical literature regarding firm tax noncompliance. As a result, some potential controls are overlooked and some controls may be misspecified. Third, the equations across categories are not all identically specified and it is not clear
if the results should be compared across regressions, given the different specifications. Finally,
Batra et al. (2003) do not exploit the nature of the dependent variable. The firm’s response
regarding under-reporting behaviour is grouped into categories. When a quantitative outcome is
grouped into known intervals on a continuous scale, the data are said to be “interval-coded”.12
However, Batra et al. (2003) define the dependent variable as a binary outcome for each category
and estimate the resulting equations by Ordinary Least Squares (OLS). The considerable statistical
limitations of such a linear probability model are well known.13 There is an estimation technique
that has been developed specifically for interval-coded data. This estimation procedure is known as
“interval regression” and is undertaken using maximum likelihood techniques.
This paper addresses these shortcomings. In particular, it uses the interval regression
technique and the extent to which various covariates affect the estimation results is also explored.
In addition, this paper uses information from existing theoretical and empirical literature of firm tax
non-compliance to build the empirical model explored in this paper. Finally, the issue of missing
observations or non-response is explored.
12
When the interval thresholds are unknown and have to estimated along with the other parameters, an ordered
logit/probit is the preferred estimation framework.
13
In the WBES, over 40% of the firms surveyed indicate that 100% of their sales are reported to the tax authority and
between 6-14% of firms appear in one of the other categories. This means that the linear probability models estimated
by Batra et al. will be either zero or one inflated. This results in a large number of predicted probabilities falling
outside the unit interval.
14
3
Data
In 1998, the World Bank Group launched its World Business Environment Survey (WBES).
The WBES used many of the same questions from the enterprise survey conducted for the 1997
World Development Report14 (World Bank 1997) but expanded the number of businesses and
countries surveyed and the questions/issues covered. During late 1998 and early 2000, face to face
interviews15 were conducted with either managers or owners of just over 10,000 firms in eighty
countries (plus the West Bank and Gaza).
The purpose of the survey was to assess and compare the business environment in a large
number of countries. To achieve this goal, the survey gathered information regarding the firm’s
characteristics, such as size and ownership structure, as well as responses to multiple questions on
the investment climate and the local business environment as shaped by domestic economic policy,
governance, regulatory, infrastructural, and financial impediments, as well as assessments of public
service quality. A more detailed description of the survey can be found in the Appendix and in
Batra et al. (2003).
Of interest to this study, the survey included a question regarding the extent and intensity to
which firms fail to report income to the tax authority which permits exploring the links between tax
non-compliance and various firm, business, governance, and political characteristics. The main
advantages of the WBES over audit data include the fact that: (1) audit data are not widely
available, unlike the WBES, which covers eighty countries; (2) audit data only include firms that
are selected (for diverse reasons) or caught by the tax authorities, while the WBES is a broad
14
Unfortunately, this dataset contains neither detailed information on characteristics of the firms nor the key variable of
interest contained in the WBES dataset. As a result, it cannot be appended to the WBES dataset.
15
With the exception of Africa where interviews were predominantly conducted by mail.
15
sample of firms; and (3) the WBES includes additional information that is not included in tax
audits. In particular, factors that firms perceive as business obstacles, such as taxes and regulations,
which may effect a firm’s decision to under-report, are included.
The WBES does not come without disadvantages and these will be discussed below. One
disadvantage shared by both the WBES and audit data is that neither data source includes “ghosts”
(Erard and Ho 2001); firms that operate solely in cash and avoid normal business obstacles and
regulations. The effect of omitting these firms from the estimates will bias estimates of tax noncompliance downwards and this bias will be larger for those countries that have a larger percentage
of unregistered firms.
3.1 Dependent Variable
As was indicated above, the intent of this paper is to explore the relationships between firm
characteristics and tax compliance using firm-level data collected from around the world. How can
the WBES be used to investigate firm tax compliance? The WBES asks each firm the question
What percentage of total sales would you estimate the typical firm in your area of activity
reports for tax purposes? Possible answers: (1) less than 50%; (2) 50-59%; (3) 60-69%;
(4) 70-79%; (5) 80-89% (6) 90-99%; and (7) all (100%).16
The variable is thus not expressed in actual quantities as with Johnson et al. (2000). This
represents something of a constraint but its key advantage is that the interval-coded nature of the
responses is likely to reduce a respondent’s reporting bias. The large number of intervals, most of
them of equal size, is beneficial as it provides information on the scale of under-reporting rather
16
In the WBES, the dependent variable is actually coded as follows: What percentage of total sales would you estimate
the typical firm in your area of activity reports for tax purposes? Possible answers: (1) all (100%); (2) 90-99%;
(3) 80-89%; (4) 70-79%; (5) 60-69%; (6) 50-59%; and (7) less than 50%. The order of the responses is revered solely
for expositional clarity.
16
than simply its incidence. Given that the incidence of under-reporting is relatively high, as will be
shown below, information on the scale of under-reporting is extremely useful.17
Another disadvantage is that firms are asked about their perceived behaviour of the “typical
firm” rather than the behaviour of the firm being interviewed, similar to the wording contained in
the survey used though not discussed by Johnson et al. (2000), and this may result in response bias.
However, there are several things to consider. It is well known that survey questions on sensitive
topics, such as tax non-compliance, are subject to tremendous levels of both response and nonresponse bias (Jones and Darroch Forrest 1990; Lemmen et al. 1988; Mensch and Kandel 1988;
and Rogers and Turner 1991). One obvious solution is for the surveyor to provide confidentiality
assurances to the respondent but these rarely address all of the concerns of the respondent and,
hence, do not yield the desired effect (Rasinski et al. 1999). Another solution is to ask hypothetical
questions but such questions usually do not produce clear and consistent data representing real
activities when they are ask about situations the respondent has never encountered first hand. By
far the most popular solution is to use proxy-responses, where the individual responding to the
survey provides information about another individual. This is a technique that had become popular
in household surveys where a household member provides responses for household members who
are absent. In the case of the WBES, the firm is not asked about the behaviour of a particular firm
but of a typical firm in the same area of activity to that of the firm being surveyed. This question is
a mix between a hypothetical situation and a proxy response. Based on the phrasing of the
question, respondents are likely to base their response, at least in part if not in whole, on their own
reporting behaviour, which would minimize the bias. Second, firms will not feel stigmatized by the
17
Further, the interval coding of the data along with the lack of data on the dollar amount of sales for the firms
surveyed in the WBES means that the tax gap cannot be calculated. The tax gap is the difference between the taxes
paid and the taxes that should have been paid. Such a calculation is useful in determining the amount of tax revenues
lost to non-compliance.
17
interviewer and/or fear possible repercussions when responding to the question which again will
minimize bias. Finally, firms are likely to base their own behaviour on the perceptions that they
have regarding the behaviour of other firms in their area of activity, especially those that they
compete with. If they perceive firms in their area activity to be engaging in tax non-compliance
this will influence their reporting decision.
The distribution of answers to the sales reporting question over the entire sample is given in
Table 1a. The first column of the table indicates that just over 18% of the firms in the sample did
not respond to the question of interest. It is quite common in survey data to have missing data of
this nature. If the data is missing for unknown reasons and the missing data is unrelated to the
completeness of the other observations then it is reasonable to exclude the missing observations
from the analysis, though efficiency is sacrificed. If, however, the missing data are systematically
related to the phenomenon being modeled then there will be a sample selection problem. For
example, if firms whose response is missing for the sales reporting question are more likely to be
firms who under-report their income, then the missing observations represent more than just
missing information. Unfortunately, there is insufficient evidence to definitively ascertain the
reasons for the missing observations in this particular case. Rather, the relationship between these
missing observations and firm characteristics that will be used as explanatory variables in our
model were explored. If the missing observations are randomly distributed amongst these
explanatory variables then this provides some evidence that the missing information may not cause
systematic bias in the estimates. There appeared to be no significant systematic relationship
between nonresponse to the sales reporting question and the explanatory variables , with one
exception noted below.
18
Of the firms that responded to the sales reporting question, Table 1b shows that 60% of
firms worldwide indicate that the typical firm fails to report their sales in full to the tax authority
and, of those firms, over 19% of them fail to report more than half their sales. This shows that
business tax compliance is a significant issue. Not surprisingly, there appears to be some
differences in perceived tax non-compliance across regions, as is shown in Figure 1. In particular,
in OECD countries only approximately 40% of firms are perceived to under-report their sales to the
tax authority and of those, approximately 50% fail to report only up to 10% of their sales. Further,
compared to other regions, firms in Latin America and Asia perceive that significantly more firms
fail to report less than 50% of their sales. In addition, there are significant differences in perceived
firm tax compliance across countries. Given these differences, the effect of country specific
dummy variables on the results will be investigated in the empirical analysis.
There are, however, some differences in the distribution of the missing observations
discussed above with respect to region and country (Table 2 lists the countries covered by the
survey, sorted by region). Only 11% of firms located in the former Soviet Union, approximately
15% of firms in the OECD and Latin America, approximately 20% in Asian and Africa and the
Middle East, and 27% in Transition Europe failed to respond to the sales reporting question. The
missing observations appear to be randomly allocated amongst the countries in Latin America and
Africa and the Middle East. In Transition Europe, no firm located in Albania responded to the
question, indicating that survey was not carried out in a similar fashion to that in other countries
rather than a response bias. As a result, Albania is excluded from the analysis (163 observations
leaving 9,869). Higher non-response also occurs in Malaysia, India and Thailand in Asia and
Russia, Georgia, Kazakhstan, and the Ukraine in the Former Soviet Union but in all cases the nonresponse rate is not dramatically higher from that experienced in other countries in the regions. In
19
the OECD, France, Canada and the United States all have a higher response rate than the other
countries in the group but again the difference is not alarming.
3.2 Explanatory Variables
As was outlined above, previous empirical work using audit data found relationships
between tax non-compliance and various firm characteristics and these relationships will also be
explored in this paper. The firm characteristics that will be included as explanatory variables
include dummy variables for: industry sector, size, age, ownership, export status, whether the
firm’s financial statements are audited, legal ownership, and number of competitors. In the
secondary analysis, the effect of various perceived business obstacles on tax compliance will be
investigated, including access to capital, taxes, and regulations, organized crime, government
corruption, inflation, and exchange rates. As the sample size decreases due to non-response with
the inclusion of these latter explanatory variables, they are only included in a secondary regression.
3.2.1
Legal Status
The previous literature makes it clear that tax non-compliance likely varies across
governance models. The WBES provides an opportunity to explore the relationship between the
legal structure of the business and tax non-compliance. Firms in the WBES are categorized as
being either: sole proprietorships (18%), partnerships (17%), cooperatives (3%), private
corporations (28%), public corporations (9.6%), or other (15%).18 Dummy variables for business
type need to be created if this characteristic is to be included in the analysis. Few firms in the
18
Fewer than 9% (836) of the firms in the sample did not respond to this question.
20
sample indicated that their business type is a cooperative. Consequently, these firms will need to
be included in one of the other categories. As cooperatives are accountable to their members,
cooperatives are more comparable to public corporations than any of the other business structure.
Cooperatives are, therefore, included with public corporations. The remaining governance
categories are treated individually. Based on the literature, it is expected that publicly traded
corporations will be the most compliant. In most countries, publicly traded corporations have a
greater probability of being audited and are subject to public disclosure requirements and
independent financial auditing, which tends to expose any under-reporting behaviour to the
authorities. As audit levels and detection probability are greater, compliance should be higher.
3.2.2
Sector
Many studies have found that tax non-compliance varies widely across industry sectors. In
the WBES, firms are categorized into one of five possible broadly defined industry sectors.19 They
are: services (33%), agriculture (39%), construction (6%), manufacturing (9%), and other (3%).
Dummy variables for industry sector need to be created if this characteristic is to be included in the
analysis. These industry categories in the WBES are likely too broadly defined to obtain any
significant relationship between industry sector and tax non-compliance. For example, the service
sector will include restaurants and hair salons as well as law and accounting firms. As a result, no
significant relationship may be found between industry sectors and under-reporting.
19
Exactly 9% (888) of the firms in the sample did not respond to this question.
21
3.2.3
Size
Rice (1992), Giles (2000), and Hanlon et al. (2005) each investigated the relationship
between a firm’s size and tax non-compliance. Rice (1992) defined a firm’s size according to the
dollar value of assets it held and found a positive relationship between firm size and tax noncompliance, whereas Giles (2000) used a firm’s sales revenues and before tax-profits, and found a
negative relationship between firm size and tax non-compliance. Using the same definition as
Rice, Hanlon et al. (2005) found a convex relationship between firm size and non-compliance.
None of these measures of firm size are available in the WBES. Instead, this paper investigates a
firm’s size as defined in terms of the number of employees.20 A small firm is one with fewer than
50 employees (40%), a medium-sized firm has between 50 and 500 employees (40%), and a large
firm has over 500 employees (20%). The relationship between firm size and tax compliance
remains unclear but it is not unreasonable to assume a negative relationship between firm size, as
defined in the WBES, and tax non-compliance.
3.2.4
Competition
While not explicitly investigated in the previous literature, it is possible that there is a
positive relationship between the number of competitors in a given market and tax non-compliance.
For example, firms in highly competitive markets may resort to tax non-compliance in order to
reduce costs, set a lower price for their goods, and/or increase profits. The WBES asks firms how
many competitors they face in their market.21 Possible answers are: none (11%), between one and
three (37%), and more than three (48%).
20
21
Only ¼ of 1% (25) of firms failed to respond to this question.
Fewer than 4% of firms in the sample did not respond to this question.
22
3.2.5
Age
The WBES includes information on the firm’s age22. Specifically, if the firm is less than
five years old (26%), between five and fifteen years old (35%), or greater than fifteen years old
(39%). The effect of a firm’s age on tax compliance is ambiguous. For example, younger firms
may be less compliant because they may have more competitors, may be struggling to turn a profit,
and/or may view tax evasion as a way to cut costs. On the other hand, young firm may be more
reliant on external financing, which could cause them to be more compliant.
3.2.6
Other Characteristics
Other firm characteristics include indicators of whether the firm is foreign owned (18%),
government owned (12%), an exporter (34%), and whether the firm subjects its financial statements
to audits (58%).23 All of these indicators are expected to be positively related to compliance as
found in the previous literature.
3.2.7
Perceived Obstacles
In addition to recording information regarding the firm’s characteristics, the WBES also
asks a number of questions about the firm’s perceptions of various constraints in the business
environment, which likely influence operational decisions including tax compliance. While many
of these variables have a low response rate, several have a relatively higher response rate.24 In
particular, firms are asked if the following items present an obstacle or are a constraint to
conducting business: access to capital, inflation, exchange rate, organized crime, political
22
All firms responded to this question.
Non-response among these questions if very low (fewer than 5%).
24
Taking account of these explanatory variables reduces the number of available observations to 5.393.
23
23
instability, taxes, regulations, and corruption. If these issues are perceived by the firm as affecting
their ability to conduct business, then they may result in the firm engaging in tax non-compliance in
order to reduce costs and be more competitive.
Andreoni (1992) argues that “…individuals facing binding borrowing constraints may use
tax evasion to transfer resources from the future to the present. Even if a person finds tax evasion
undesirable in the absence of borrowing constraints, it could become desirable if a borrowing
constraint is binding. Tax evasion, therefore, may be a high-risk substitute for a loan.” (Andreoni
1992, 35-36). Johnson et al. (2000) note that under-reporting is likely increased if the economy has
a poorly developed banking sector. Hence, if access to capital is an obstacle, the firm may
underreport.
As noted by Giles and Tedds (2002), price inflation may influence under-reporting in two
ways. First, if tax rates are not indexed then the tax burden rises simply due to “bracket creep”.
Second, inflation generates uncertainty and “…businesses may hedge against this uncertainty by
engaging in (more) underground activity.” (p. 11) They also note that the exchange rate can
influence under-reporting. Exchange rates can be an obstacle to doing business if the exchange rate
is unpredictable and can increase the input costs faced by businesses.
Various governmental factors can also influence under-reporting. Political instability
constrains a countries ability to collect taxes and also reduces the efficiency of collection, thereby
increasing non-compliance. It is generally acknowledged in the tax non-compliance literature that
the presence of corporate taxes provides incentives for firms to hide output to reduce their tax
burden (Schneider and Enste 2000). The degree of regulation is often cited as a factor that
influences people to engage in underground activity as regulations reduce a firms’ freedom of
choice (Brehm 1966; Deregulation Commission 1991; Monoplkommission 1998; Schneider and
24
Enste 2000). Johnson et al. (1998a) find a positive correlation between an index of regulations and
the underground economy and argue that governments should focus on reducing the density of
regulations.
In the WBES, firms are asked if: (1) it is common to pay some “irregular” additional
payments to government officials; and if (2) laws and regulations that affect the firm are interpreted
inconsistently by the government or courts. If corruption is common, then among other effects, it
will increase the cost of business, reduce morality, and reduce a firm’s confidence in government;
all of which are likely to have a negative relationship with tax compliance. Johnson et al. (2000)
and Johnson et al. (1998b) find a positive and significant relationship between extortion by
government officials and under-reporting. The relationship between tax compliance and
inconsistency in the application of laws and regulations is more ambiguous. If a firm can
individually garner the favour of the government(s) and/or courts in the interpretation and
application of laws and regulations, then this may reduce tax non-compliance. On the other hand,
if the firm does not benefit from this inconsistency, then it may resort to tax non-compliance.
3.3
Data Summary Statistics
Since no obvious explanation is available for the missing observation on the sales reporting
questions, these observations (1,716) were dropped from the data set leaving 8,153 observations.
Table 3 provides summary statistics for the key variables described above.
4
Empirical Framework
The survey question which forms the basis for the dependent variable and is described
above refers to categories. As a result, a firm’s reporting behaviour is not directly observed.
25
Rather, firms are categorized on the basis of the percentage of sales that go unreported. When a
quantitative outcome is grouped into known intervals on a continuous scale, the data is said to be
“interval-coded”. An ordered probit (or logit) is ideal when the dependent variable is discrete,
ordinal in nature, and when the categories or thresholds are unknown.25 In this case, the thresholds
are estimated along with the model’s coefficients and the variance of the error term is normalized to
be one. It is possible, however, to modify the ordered probit model so that the thresholds are fixed
at their known values and only the coefficients and the error variance are estimated. This
estimation procedure is known as “interval regression” and is undertaken using maximum
likelihood techniques. The key advantage of interval regression over the ordered probit is that it
provides an asymptotically more efficient estimator as it uses the known threshold information and
involves estimating fewer parameters. It is also preferred to OLS, as OLS on the grouped
dependent variable model is inconsistent.26
The following is the general setup of the model and is based on the discussion contained in
Stewart (1983). The responses for the dependent variable are coded 1, 2, 3, 4, 5, 6, and 7 to capture
seven distinct sales under-reporting categories. Let yi denote the observable ordinal variable
coded in this way for the ith firm and let y i* denote the underlying variable that captures the sales
under-reporting of the ith firm. This can be expressed as a linear function of a vector of explanatory
variables xi using the following relationship:
25
That is, for an ordered logit/probit model, while it can be said that two is greater than one, it cannot be determined if
the difference between two and one is somehow twice as important as the difference between one and zero. The latter
is not true of interval coded data.
26
Another popular alternative is to define an artificial dependent variable that takes the value of the midpoint of the
reported interval and then estimate the model by OLS. The resulting estimates are inconsistent but can be transformed
using a minimum χ2 method and the transformed estimates are consistent. However, the transformation requires
equally strong distributional assumptions, usually that of normality and homoskedasticity, to that required of the
interval regression model.
26
yi* = xi' β + ui , ui ~ N (0, σ 2 ).
(1)
It is assumed that y i* is related to the observable ordinal variable yi as follows:
y i = 1 if y i* ≤ 49%
y i = 2 if 49% < y i* ≤ 59%
y i = 3 if 59% < y i* ≤ 69%
y i = 4 if 69% < y i* ≤ 79%
(2)
y i = 5 if 79% < y i* ≤ 89%
y i = 6 if 89% < y i* ≤ 99%
y i = 7 if y i* > 99%
Based on the above, the seven possible components of the general log likelihood function
for the ith individual is expressed as:
Li = I [ y i = 1] × log e {Φ (
Φ(
49 − xi' β
σ
log e {Φ (
49 − xi' β
σ
) − Φ(
a 0 − xi' β
)} + I [ y i = 3] × log e {Φ (
79 − xi' β
) − Φ(
69 − xi' β
σ
69 − xi' β
σ
)} + I [ y i = 2] × log e {Φ (
) − Φ(
59 − xi' β
σ
)} + I [ y i = 5] × log e {Φ (
59 − xi' β
σ
)−
)} + I [ y i = 4] ×
89 − xi' β
σ
σ
σ
'
'
'
79 − xi β
99 − xi β
89 − xi β
Φ(
)} + I [ y i = 6] × log e {Φ (
) − Φ(
)} +
σ
σ
σ
a K − xi' β
99 − xi' β
I [ y i = 7] × log e {Φ (
) − Φ(
)}
σ
σ
)−
(3)
where Φ denotes the cumulative distribution function of the standard normal, Φ (−∞ ) = 0 ,
Φ (∞ ) = 1 , log e (·) denotes the natural logarithmic operator, and I[·] is an indicator function that
takes the value of one when the statement in the square brackets is true and zero when it is false.
27
The relevant part of the log-likelihood is then triggered by the indicator function for whether the
individual falls within one of the seven intervals in question. The maximum likelihood procedure
now involves the estimation of the β parameter vector and the ancillary standard error parameter σ.
The question then becomes how the end intervals should be treated. That is, what are the
values for a0 and aK noted in equation 3? Stewart (1983) treats the first and last intervals as open
ended. In this paper, it is difficult to argue that a firm would report less than 0% of their sales.
Instead, it could be argued that the first interval should begin at either 1 or 0%: 1% since it can be
argued that firms that report 0% of their sales would likely be operating completely as “ghosts” and
would not have been selected for an interview since there would be no formal record of the firm. It
could also be argued that the last interval should be treated as close-ended at 100%. However,
treating the last interval as open ended accounts for possible sales over-reporting. Rice (1992)
found that about 6% of all corporations overstate their taxable income to some extent. Firms may
over-report due to a misinterpretation of tax laws, to avoid a tax audit, to secure financing, or
particularly for public corporations, to appear more competitive. In the WBES, firms that overreport will likely be included in the full compliance category. The effect of choosing different
values for a0 and aK on the resulting estimates will be explored. It should be noted that the
sensitivity of the results and predicted values to the treatment of the end intervals is often
overlooked in the literature yet can be of clear importance to the economic validity of the resulting
estimates.
Unlike the situation with the ordered probit estimation, the estimated coefficients from an
interval regression are interpretable as if y i* is observed for each i and estimated E ( y * | x) = xβ by
OLS. That is, the estimated coefficients can be interpreted as the marginal effects (i.e. the change
in percentage of sales reported given a change in the independent variable, holding all else
28
constant). It should be noted that the estimates contained in the β parameter vector are only
interpretable in this way due to the assumption that y* given x satisfies the classical linear model
assumptions. If these assumptions do not hold then the interval regression estimator of β would be
inconsistent. As a result, it is important to test the key assumptions of functional form,
homoskedasticity, and normality, all of which are used to derive the log likelihood function noted
in equation (3). Unfortunately, these assumptions, despite their importance, are seldom tested in
applied research. Equally, these assumptions, particularly homoskedasticity and normality, are
rarely met when using micro data.
4.2 Diagnostic Tests
Chesher and Irish (1987) outline diagnostic tests for (pseudo) functional form, normality
and homoskedasticity for the ordered probit model that are easily modified for the interval
regression model. These tests are all score (or Lagrange Multiplier) tests for which the test
statistics take the form
ξ = 1' F ( F ' F ) −1 F '1
(4)
where 1 is an n-dimensional vector of ones, and F is a matrix with row order n where each row
contains the score contributions for all the parameters of the model. ξ can be easily calculated as n
times the non-centered R2 from a regression of 1 on the columns of F.
The construction of the F matrices for these tests, which are described below, are based on
computations of the pseudo-residuals. Usually, residuals are defined as the difference between the
observed and estimated values of the dependent variable. However, the estimated values of the
dependent variable obtained in the interval regression have no counterpart in the data. Rather,
29
pseudo-errors need to be computed. An expression for the pseudo-errors is obtained by
differentiating equation (3) with respect to σ and denoted for the ith individual as:
φ(
ui =
a ( j −1)i − xi' β
σ [Φ (
σ
a ji − xi' β
σ
) − φ(
) − Φ(
a ji − xi' β
σ
)
a( j −1) i − xi' β
σ
(5)
)]
where φ (⋅) denotes the probability density function for the standard normal, and aj-1 and aj denote
the known interval parameters for individual i (e.g. if yi=2 then aj-1=49 and aj=59). The pseudoresiduals, ei, are obtained by replacing the unknown parameters in (5) with their maximum
likelihood estimates.
For the homoskedasticity and non-normality tests, higher-order moment residuals are
required, specified as:
M τi =
(
a( j −1)i − xi' β
σ
)τ φ (
σ [Φ (
a ( j −1)i − xi' β
σ
a ji − xi' β
σ
)−(
) − Φ(
a ji − xi' β
σ
)τ φ (
a ( j −1)i − xi' β
σ
a ji − xi' β
σ
)
(6)
)]
The higher-order moment residuals are obtained by replacing the unknown parameters in (6) with
their maximum likelihood estimates. The first four moment residuals are required for the desired
tests and are defined as follows:
ei1 = ei
e 2 = Mˆ
i
1i
e = 2e + Mˆ 2i
e 4 = 3e 2 + Mˆ
3
i
1
i
i
i
(7)
3i
The F matrix, or score contributions, is obtained by multiplying the pseudo-residuals by the various
auxiliary variables in question.
30
It should be noted that if either or both of the assumptions of normality and
homoskedasticity are rejected, then the Huber (1967) “sandwich” estimator of the variance can be
used in place of the conventional Maximum Likelihood variance estimator. This estimator is
expressed as:
Var ( βˆ ) = [ I ( βˆ )]−1 ( xi' ei2 xi )[ I ( βˆ )]−1
(5)
where I ( βˆ ) is the information matrix for the βˆ vector, computed at the maximum likelihood
estimates. However, unlike with linear regression, in an interval regression model is it not possible
to correctly specify E ( y * | x) but misspecify Var(y*|x) which implies that the validity of this
correction is highly questionable in the interval regression model and, hence, is not used in this
paper.
While it is worthwhile considering the results of the aforementioned tests, care should be
exercised when interpreting the associated results. Orme (1990) has questioned the use of such
score tests in the context of a simple binary probit and demonstrated their poor finite sample
properties in this setting. In particular, he notes that there is upward size-distortion. Assuming that
these findings extend to the interval regression model, the diagnostic tests may indicate that the
model does not satisfy the classical linear model assumptions when in fact it does. However, the
sample size for the data used in this paper is large enough that this concern likely does not apply.
4.2.1
Pseudo-Functional Form Test
The (pseudo) functional form test is a modified version of the RESET test (Ramsey 1969).
F is given as
F = (e1 x, e1 yˆ *2 ,..., e1 yˆ *K , e 2 )
(8)
31
where x includes a column of 1’s if the grouped model contains an intercept and yˆ *K is the Kth
power of model’s predicted standardized index, yˆ =
*
x ' β̂
σ
.
That test statistic ξ is distributed
as χ 2 ( K − 1) . If the null hypothesis is rejected, this is evidence of model misspecification that may
be rectified by considering additional variables and/or alternative functional forms (such as a semilog model).
4.2.2
Test for Homoskedasticity
For the test of homoskedasticity, F is given as
F = (e1 x, e 2 d i ) 27
(9)
The test statistic ξ is distributed as χ 2 (q ) where q is equal to the number of variables that are
interacted with e2. If the form of the heteroskedasticity is unknown, then d i = xx' . This is simply
White’s test for heteroskedasticity of unknown form, modified for the interval regression model.
The test is sensitive to model misspecification so if the test for (pseudo) functional form is rejected,
it is also likely that the test for homoskedasticity will also be rejected.
As indicated previously, the estimated coefficients are inconsistent in the presence of
uncorrected heteroskedasticity but it is possible to address this inconsistency by modeling the
heteroskedasticity. The preferred choice for the functional form for the heteroskedasticity is a
variation of Harvey’s (1976) multiplicative heteroskedasticity, denoted as Var(ui)=exp(diγ)2 where
di is a matrix of independent variables, including a column of one’s, that are the source of the
27
When an intercept is estimated so that x always contains a unit element, e2 is redundant in the test for
homoskedasticity.
32
heteroskedasticity and γ is a coefficient vector. The expression exp(diγ) replaces the σ noted in
equation (3).28
4.2.3
Non-Normality Test
Finally, F in the usual χ 2 (2) test for zero skewness and/or excess kurtosis is given by
F = (e1 x, e 2 , e 3 , e 4 )
(10)
It is anticipated that the assumption of normality is likely violated in this model. Table 1b clearly
shows that the distribution of sales reporting is negatively skewed. Normality may be more closely
approximated by taking the natural logarithm of 100 minus the dependent variable, which is simply
the natural logarithm of the percentage of sales not reported to the tax authority. Such a
transformation requires that the dependent variable take only strictly positive values and care
exercised in interpreting the resulting estimates. This is a transformation that can also be explored
if the model fails the functional form test. It should also be acknowledged that the presence of
uncorrected heteroskedasticity can also lead to a rejection of the null of normality. Correcting for
the heteroskedasticity, as noted above, can often achieve normality in the (pseudo) residuals.
5
Results
Estimation was undertaken using STATA 9.2 and the interval regression model was
estimated using the ‘intreg’ command. The command needs two variables, denoted y1 and y2, to
define the dependent variable. In particular, y1 and y2 are used to hold the endpoints of the interval.
28
In STATA, the heteroskedastic corrected interval regression model is estimated using the ‘het’ option on the ‘intreg’
command.
33
Table 4 provides the concordances between yi, the associated interval, y1, and y2. The model for
the analysis is:
y i* = α + β 1 PROPi + β 2 PARTNERi + β 3OTHERBUS i + β 4 PRIVCORPi + β 5 SERVICEi +
β 6 AGRi + β 7 CONSTRUC i + β 8 OTHERINDi + β 9 SMALLi + β10 LARGE i +
β 11 LESS 5 i + β 12 OVER15 i + β 13 NOCOMPETi + β14 MORE 3COMPETi + β 15 FOREIGN i +
(11)
β 16 GOVOWN i + β 17 EXPORTi + β 18 AUDITi + β 19 FIN i + β 20 INSTABILE i +
β 21ORGCRIMEi + β 22TAX i + β 23 CORRUPi + β 24 INFLATi + β 25 EXCHANGE i +
β 26 LAWS i +
106
∑ β COUNTRY +u ,
K = 27
j
i
i = 1....N
Table 3 provides a description of the above noted variables and Table 2 denotes the
countries available in the data. The model is estimated both with and without the perception
variables. This produces a total of two possible models to consider. As indicated above, it is
necessary to specify the values of the two end intervals. The values of -•, 0, and 1 are considered
for the values of a0, and 100 and • are considered for the values of aK. The results of these models,
twelve in total, are presented in Table 5. Models 1 through 6 do not include the perceptions
variables while Models 7 through 12 do include the perception variables. The treatment of the end
points is noted in the heading of column. For example, Model 1 and 7 treat the first and last
intervals as open ended.
In all models the constant represents the reporting behaviour of the baseline firm. The
baseline firm is a public corporation, in the manufacturing industry, has between 50 and 500
employees, is between five and fifteen years of age, is a domestic, non-government owned, nonexporter, has between one and three competitors and does not have it financial statements audited.
According to Model 1 through 3, the baseline firm reports just over 90% of its sales to the tax
authority and this falls to just over 80% for Model 4 through 6. For models 7 through 9, this jumps
34
to an implausibly high estimate of approximately 110% but falls to a more realistic number of
approximately 91% for Models 10 through 12.
Models 1 through 3 and 6 through 9 all treat the last interval as open ended and it is
important to investigate the implications that this has on the predicted values, ŷ * . The mean,
standard deviation, minimum, maximum and number of predicted values in excess of 100% are
also reported in Table 5. In Models 1 through 3 and 6 through 9, the maximum predicted value is
in excess of 140% and over 25% of the predicted values exceed 100%. These numbers seem
economically implausible, particularly that over 25% of firms report in excess of 100% of their
sales to the tax authority. Hence, Models 1 through 3 and 6 through 9 are ruled out.
This leaves Models 4-6 and 10-12 for consideration. In these models, the maximum
predicted value is approximately 105% and approximately 1% of the predicted values exceed
100%. These figures seem to be economically plausible and accord with the findings of Rice
(1992) who found that some firms do over-report to some extent. A comparison of the estimates
across Models 4-6 and 10-12 show little variation. Various goodness of fit measures are also
presented in Table 5 for the relevant models. Larger (less negative) log-likelihood values are
indicative of a better fit. The log-likelihood values can be compared across the models if they have
the same samples and dependent variable. The R-square, for technical reasons, cannot be
computed in the same way in interval regressions as it is in OLS regression. Various pseudo Rsquare measures, however, have been proposed, but there is no generally accepted measure. Veall
and Zimmermann (1996) recommend the measure of McKelvey and Zavoina (1975), which is
reported in Table 5.29 Again, a comparison of the goodness of fit estimates across the models show
little variation. Given the similarities across Models 4-6 and 10-12, it is difficult to eliminate
29
The McKelvey and Zavoina (1975) R-square is computed in STATA with the 'fitstat' command.
35
further models based solely on the goodness of fit measures. Since Models 5 and 11 constrain the
unobserved dependent variable to range between 1 and 100, which seems economically reasonable,
Models 5 and 11 are selected for further consideration.
The diagnostic test results for the preferred models, Models 5 and 11, are presented in Table
5. For model 5, the null of correct functional form is not rejected at a 1% significance level but
rejected at all others. All other tests are rejected at the usual significance levels. This is concerning
since it implies that the estimates are inconsistent. As mentioned above, the dependent variable is
negatively skewed. Since the preferred models each constrain the dependent variable to be strictly
positive, a possible transformation to consider is to transform the dependent variable into the
percentage of sales not reported to the tax authority, calculated as 100% minus the threshold
parameters. This newly formed variable is positively skewed and can be smoothed by taking the
natural logarithm.30 The results from these models are presented in Table 6. The coefficients
cannot be directly compared to those reported in Table 5 due to the transformation of the dependent
variable. That is, the coefficients are no longer marginal effects but 100 multiplied by the
estimated coefficient represents the percentage change.31 Finally, as the dependent variable is now
the natural logarithm of the percentage of sales not reported to the tax authority, the signs on the
estimated coefficient should be the opposite of those in Table 5, where the dependent variable was
the percentage of sales reported to the tax authority.
30
Since taking 100 minus the observed threshold parameters results in a value of 0 for those firms that report 100% of
their sales, a small constant is added to transformation such that the newly formed variable is strictly positive
everywhere.
31
Kennedy (1981), however, notes that in a semi-log model the coefficient on a dummy variable must be appropriately
transformed. That is, if bi is the estimated coefficient on a dummy variable and V(bi) is the estimated variance of bi
then
g i = 100(exp(bi −
V (bi )
) − 1)
2
(12)
gives an estimate of the percentage impact of the dummy variable on the dependent variable. The effect of this
transformation on the resulting coefficients was explored. As the resulting change to the coefficients reported in Table
6 were marginal, the results are not reported but are available from the author upon request.
36
The columns Model 5a and Model 11a in Table 6 present the results for the preferred
Models from Table 5 using the transformed dependent variable. For both models, the null of
correct functional form is not rejected at all the usual significance levels, which represents
significant improvement over the models presented in Table 5, but all other tests are rejected at the
usual significance levels. Figure 2 presents plots of the pseudo residuals from these two models
which support the finding of non-normality and it is not clear if any further transformations would
result in normality. In addition, further transformations would make it difficult to interpret the
estimated coefficients. It is, however, possible to correct for the heteroskedasticity, as noted above,
if a form for it is specified. It is not unreasonable to assume that much of the heteroskedasticity
results from countries and business type. Applications of the test for homoskedasticity of known
form across subsets of the variables confirm this suspicion. Models 5a and 11a are re-estimated to
account for the known form of the heteroskedasticity and the results are also presented in Table 5
under the column heading Model 5b and 11b. These results will now be discussed.
The results for Model 5b will be discussed first. Sole proprietorships and private
corporations are the least compliant form business of all the business models included in this study.
This result accords with those found by Rice (1992), that found that public corporations were more
compliant than other types of firms, and Hanlon et al. (2005) who found that private corporations
were less compliant. Small firms, and firms with more than three competitors are also found to be
the least compliant while large firms and firms with no competitors are more compliant. Giles
(2000) reports a similar result with respect to size, whereas Rice (2002) found firm size and tax
non-compliance were positively related and Hanlon et al. (2005) found that compliance was a
convex function over firm size. The results with respect to competition can be interpreted in
several ways. First, firms in highly competitive markets may resort to tax non-compliance to
37
reduce costs and allow the firm to set a lower price for their goods and/or increase the firm’s
profits. Second, firms base their behaviour on the behaviour of firms that are in the same area of
activity as themselves. This lends credence to using the firms response regarding the reporting
behaviour of a typical firm in their area of activity as a proxy for the firms own behaviour. Finally,
being an established firm, greater than fifteen years of age slight increases compliance and being
foreign or government owned, and having financial statements audited all lead to significantly
higher compliance. Both Giles (2000) and Chan and Mo (2000) found that foreign owned firms are
more compliant, and the results in Table 6 provide further support for this result. The relationship
between internal audit controls of the firm and tax compliance has not been investigated in previous
empirical work.
Numerically, the percentage of sales not reported to the tax authority is: 22.5 percent higher
for sole proprietorships and 9.4 percent higher for private corporations; 17.9 percent higher for
small firms and 13.1 percent lower for large firms; 7.9 percent lower for firms greater than fifteen
years of age; 17.5 percent lower for firms with no competitors but 11.6 percent higher for firms
with more than three competitors; 17.3 percent, 11.7 percent and 19.2 percent lower for foreign
owned firms, government owned firms, and firms with audited financial statements, respectively.
Chan and Mo (2000) report that export oriented firms are more compliant but this characteristic is
found to be insignificant. Previous empirical studies all found a significant relationship between
industry and tax non-compliance whereas no significant effect is found in this specification. As
discussed above, the categories included in the WBES data set are likely too broadly defined to
yield any significant results.
The results for Model 11b, which includes the perception variables, are similar to those
reported for Model 5b with most being only slightly larger. The key exceptions are: the coefficient
38
on sole proprietorships, which is over 1.5 times higher at 34 percent; the coefficient on the service
industry which is now statistically significant and suggests that the percentage of sales not reported
by firms in the service industry is 8.6% lower; the coefficient for firms greater than fifteen years of
age is no longer statistically significant while the coefficient for firms less than five years of age is
now statistically significant and suggests that they are more compliant, reporting 5.7% more of
their sales; the coefficient on more than three competitors is no longer statistically significant; and
the coefficient for foreign owned firms is also no longer statistically significant.
With respect to the perception variables, firms that perceive high taxes and regulations, and
government corruption as obstacles to doing businesses report significantly less of their sales to the
tax authority. The results presented in Table 6 indicate that government corruption has the single
largest effect, resulting in the percentage of sales not reported to the tax authority being a whopping
53.4 percent higher and taxes at 20.2 percent higher. In comparison, Johnson et al. (2000) also
found a positive relationship between non-compliance and government corruption but failed to find
a relationship between compliance and tax payments. Exchange rates have a positive effect on
compliance, which is the only coefficient that fails to follow a priori expectations. All other
variables are statistically insignificant.
7
Conclusion
Very little is actually known about firm tax compliance due to a lack of detailed and readily
available data. The purpose of this paper was to use a unique and recently available dataset that
contained information on firms from around the world to investigate some of the factors that effect
business tax compliance. This is one of the first studies to examine firm tax compliance using
39
worldwide data. The majority of previous empirical studies were confined to examining firms
within a particular country using tax audit data.
The empirical strategy employed in this paper exploits the nature of the dependent variable,
which is interval coded, and uses interval regression which provides an asymptotically more
efficient estimator than the ordered probit. In addition, the estimated coefficients from an interval
regression are interpretable as marginal effects provided that the model satisfies the assumptions of
correct functional form, homoskedasticity, and normality. If these assumptions do not hold then
the interval regression estimator of β is inconsistent. These assumptions are investigated using
standard diagnostic tests that have been modified for the interval regression model, a step
frequently ignored in applied research. The test results support the use of the semi-log model that
is corrected for heteroskedasticity, the origin of which is found to result at the country level and
from the business type.
Overall, evidence is presented that shows that firms in all regions around the world engage
in tax non-compliance, but that there is substantial variation within regions. The detailed results
indicated that the legal organization of a business does affect tax compliance, with sole
proprietorships and private corporations being the least compliant business type. Firm size is also
an important indicator of tax compliance, with small firms reporting less and large firms more of
their sales to the tax authority. Firms with no competitors, government owned firms and firms with
audited financial statements are significantly more compliant. Industry and firms that export is
found to have no significant effect on compliance and the effect of firms age on compliance varies
with model specification. Whether or not firms with more than three competitors and firms that are
foreign owned are more compliant depends on the specification of the model as is also the case
with firm age. Not surprisingly, taxes and government corruption are positively related to tax non40
compliance and the magnitude is quite large in both cases. Access to capital, political instability,
organized crime, inflation, and the legal system are found to have no statistically significant effect
on tax non-compliance.
The findings do suggest a role for public policy, as well as actions to be considered by the
tax authority and items that require further study. First, the findings suggest that administrations
interested in reducing business tax non-compliance should consider reducing business taxes,
minimizing the number of regulations, and reducing, if not eliminating, government corruption.
Admittedly, taking action of these issues is complex and involves more than just the tax authority.
Second, tax authorities should consider auditing sole proprietorships, private corporations, small
firms at a higher rate and requiring all firms to have their financial statement audited by a third
party. Since firms with no competitors are found to be significantly more compliant and firms in
highly competitive sectors are found to be less compliant, governments will have to weigh the
economic efficiency gains from encouraging market competition over the clear lost in tax revenues
that this causes. Finally, based on this study, it is not entirely clear why large firms, firms that are
foreign owned, and how exchange rates lead to higher tax compliance. Further exploration into
these relationships appears to be a worthwhile venture.
It is also worthy to consider the empirical strategy employed in this paper. The dependent
variable in this model is a fractional variable, an index, that takes values between zero and one.
The estimation strategy employed in this paper assumes that the marginal effects are the same
throughout the index. Papke and Wooldridge (2005) have extended a fractional response model
that accounts for responses occurring at the zero and one corners and “…ensures that the fitted
values are between zero and one and that the marginal effect are diminishing as the index argument
41
increases.” (p. 2) This and other fractional response models have not yet been extended to intervalcoded data but the consideration of such an extension is warranted.
42
TABLES
Table 1a: Univariate Frequencies of Percentage of Sales Reported to Tax Authorities,
including missing observations
Missing <50%
5060708090100%
59%
69%
79%
89%
99%
Frequency
1,879
Percent
18.73%
Observations
10,032
936
694
501
703
916
1,096
3,307
9.33%
6.92%
4.99%
7.01%
9.31%
10.93%
32.96%
Table 1b: Univariate Frequencies of Percentage of Sales Reported to Tax Authorities,
excluding missing observations
<50%
Frequency
Percent
50-59% 60-69% 70-79% 80-89% 90-99%
100%
936
694
501
703
916
1,096
3,307
11.48%
8.51%
6.14%
8.62%
11.24%
13.44%
40.56%
Observations 8,153
43
Table 2: Countries Surveyed, Categorized by Region, and Number of Observations in Each
Country
Country
Observations
Africa and Middle East
Botswana
Cameroon
Côte d'Ivoire
Egypt
Ethiopia
Ghana
Kenya
Madagascar
Malawi
Namibia
Nigeria
Senegal
South Africa
Tanzania
Uganda
West Bank and Gaza
Zambia
Zimbabwe
Total
Asia
Bangladesh
Cambodia
China
India
Indonesia
Malaysia
Pakistan
Philippines
Singapore
Thailand
Total
OECD
Canada
France
Germany
Italy
Portugal
Spain
Sweden
United Kingdom
United States1
Total
Notes:
1
Country
87
45
66
102
77
85
89
84
42
66
73
85
99
66
106
68
64
102
1,456
39
239
85
160
72
50
80
90
88
369
1,272
95
89
77
81
84
89
82
79
94
770
Observations
Transition Europe
Bosnia and Herzegovina
Bulgaria
Croatia
Czech Republic
Estonia
Hungary
Lithuania
Poland
Romania
Slovak Republic
Slovenia
Turkey
Total
Former Soviet Union
Armenia
Azerbaijan
Belarus
Georgia
Kazakhstan
Kyrgyzstan
Moldova
Russia
Ukraine
Uzbekistan
Total
Latin American and Caribbean
Argentina
Belize
Bolivia
Brazil
Chile
Colombia
Costa Rica
Dominican Republic
Ecuador
El Salvador
Guatemala
Haiti
Honduras
Mexico
Nicaragua
Panama
Peru
Trinidad and Tobago
Uruguay
Venezuela
Total
104
97
112
106
126
121
29
208
125
24
123
124
1,299
111
114
116
99
102
124
116
481
188
118
1,569
83
33
87
172
94
92
80
90
85
95
86
95
85
80
86
80
102
93
87
82
1,787
Denotes the omitted category in estimation.
44
Table 3: Data Summary
Variable
% of Sales
Reported
Sole Prop.
Partnerships
Private Corp
Public Corp. 1
Other Business
Manufacturing1
Service
Other
Agriculture
Construction
Small
Medium1
Large
<5
5-151
>15
No Competitors
1-31
>3
Foreign Owned
Gov. Owned
Exporter
Fin. Statements
Audited
Obs.
Acronym
Standard
Deviation
Minimum
Maximum
8153
REPORT
2.978
Legal Organization of Company
8153
PROP
0.180
8153
PARTNER
0.173
8153
PRIVCORP
0.280
8153
PUBCORP
0.130
8153
OTHER
0.154
Industry Sector
7418
MANUFAC
0.370
7418
SERVICE
0.424
7418
OTH
0.035
7418
AGR
0.075
7418
CONSTRUC
0.096
Firm Size
8139
SMALL
0.393
8139
MED
0.412
8139
LARGE
0.194
Firm Age
8153
LESS5
0.257
8153
5TO15
0.354
8153
OVER15
0.389
Number of Competitors
7907
NOCOMPET
0.111
7907
1TO3
0.399
7907
MORE3COMPET
0.490
Other
8153
FOREIGN
0.183
8153
GOVOWN
0.120
8153
EXPORT
0.339
8153
AUDIT
0.586
Secondary Parameters – Perception
7526
FIN
0.811
2.166
1
7
0.384
0.378
0.449
0.336
0.361
0
0
0
0
0
1
1
1
1
1
0.483
0.494
0.185
0.264
0.294
0
0
1
1
0
0
1
1
0.488
0.492
0.396
0
0
0
1
1
1
0.437
0.478
0.487
0
0
0
1
1
1
0.314
0.490
0.500
0
0
0
1
1
1
0.387
0.325
0.473
0
0
0
1
1
1
0.493
0
1
0.391
0
1
0.842
0.365
0
1
0.631
0.482
0
1
0.892
0.586
0.846
0.741
0.311
0.497
0.361
0.438
0
0
0
0
1
1
1
1
0.478
0.493
0
1
Financing
Political
7376
INSTABILE
Instability
Organized
Crime
7196
ORGCRIME
High Taxes &
Regulations
7660
TAX
Corruption
7485
CORRUP
Inflation
7425
INFLAT
Exchange Rate
7259
EXCHANGE
Laws & Regs
In consistent
8029
LAWS
Notes: 1Denotes the omitted category in estimation.
Mean
45
Table 4: Definition of the Dependent Variable for the Interval Regression Model
yi
1
2
3
4
5
6
7
Associated
Interval
(ao, 49]
(49, 59]
(59, 69]
(69, 79]
(79, 89]
(89, 99]
(99, aK]
y1
y2
ao
49
59
69
79
89
99
49
59
69
79
89
99
aK
Notes: § If: ao=-•, then y1=.; ao=0, then y1=0; and ao=1, then y1=1. Equally, if aK=•, then y2=.; and ak=1000, then
y1=100.
46
Table 5: Estimation Results
Variable
Model 1
-•<Yi*<•
Model 2
0<Yi*<•
Model 3
1<Yi*<•
Model 40£Yi*£100
Model 5
1£Yi*£100
Model 6
-•<Yi*£100
Model 7
-•<Yi*<•
Model 8
0<Yi*<•
Model 9
1<Yi*<•
Model 100£Yi*£100
Model 11
1£Yi*£100
Model 12
-•<Yi*£100
Constant
90.279
(3.709)*
-5.922
(1.746)*
-2.417
(1.644)
-1.044
(1.830)
-3.100
(1.423)**
-1.101
(2.466)
1.371
(0.976)
1.196
(1.692)
-1.357
(1.487)
-4.156
(1.031)*
3.454
(1.210)*
1.100
(1.045)
1.281
(1.050)
3.962
(1.475)*
-3.433
(1.222)*
4.846
(1.143)*
2.904
(1.510)***
90.315
(3.587)*
-5.718
(1.684)*
-2.368
(1.590)
-1.027
(1.767)
-3.037)
(1.377)**
-1.033
(2.387)
1.298
(0.941)
1.186
(1.634)
-1.309
(1.436)
-4.045
(0.995)*
3.345
(1.169)*
1.025
(1.007)
1.219
(1.013)
3.800
(1.422)*
-3.373
(1.181)*
4.684
(1.104)*
2.871
(1.459)**
90.318
(3.580)*
-5.707
(1.681)*
-2.366
(1.587)
-1.026
(1.764)
-3.033
(1.375)**
-1.030
(2.382)
1.295
(0.939)
1.186
(1.631)
-1.307
(1.433)
-4.039
(0.993)*
3.340
(1.167)*
1.021
(1.005)
1.216
(1.011)
3.794
(1.420)*
-2.369
(1.179)*
4.676
(1.102)*
2.869
(1.456)**
82.156
(2.067)*
-3.139
(0.989)*
-0.938
(0.920)
-0.221
(1.014)
82.158
(2.067)*
-3.138
(0.988)*
-0.938
(0.919)
-0.221
(1.013)
82.143
(2.070)*
-3.150
(0.990)*
-0.940
(0.921)
-0.222
(1.015)
-1.021
(0.789)
-1.021
(0.789)
-1.021
(0.790)
0.181
(1.416)
1.089
(0.549)**
0.801
(0.961)
0.001
(0.853)
-2.207
(0.587)*
1.453
(0.664)**
0.118
(0.596)
0.340
(0.589)
1.544
(0.806)***
-2.128
(0.695)*
2.213
(0.630)*
0.621
(0.828)
0.182
(1.416)
1.089
(0.549)**
0.800
(0.961)
0.001
(0.853)
-2.207
(0.587)*
1.452
(0.663)**
0.118
(0.596)
0.340
(0.588)
1.543
(0.806)***
-2.128
(0.695)*
2.213
(0.630)*
0.621
(0.828)
0.182
(1.418)
1.094
(0.550)**
0.801
(0.963)
0.001
(0.855)
-2.208
(0.588)*
1.453
(0.665)**
0.120
(0.598)
0.342
(0.590)
1.551
(0.807)***
-2.127
(0.696)*
2.217
(0.631)*
0.618
(0.829)
110.105
(4.717)*
-4.260
(1.994)**
-0.687
(1.881)
-0.844
(2.124)
-3.108
(1.629)***
1.155
(3.254)
1.986
(1.114)***
1.737
(1.997)
-0.219
(1.708)
-4.114
(1.189)*
4.309
(1.378)*
1.126
(1.211)
1.080
(1.214)
1.222
(1.624)
-3.817
(1.446)*
2.768
(1.322)**
4.015
(1.762)**
109.706
(4.583)*
-4.176
(1.931)**
-0.706
(1.824)
-0.824
(2.060)
-3.050
(1.58) ***
1.185
(3.160)
1.883
(1.09)***
1.661
(1.937)
-0.249
(1.657)
4.035
(1.152)*
4.171
(1.337)*
1.065
(1.172)
1.037
(1.177)
1.153
(1.571)
-3.710
(1.401)*
2.689
(1.281)**
3.961
(1.710)**
109.683
(4.576)*
-4.173
(1.928)**
-0.707
(1.821)
-0.823
(2.056)
-3.047
(1.58) ***
1.186
(3.155)
1.878
(1.08)***
1.658
(1.933)
-0.250
(1.657)
-4.030
(1.150)*
4.164
(1.335)*
1.062
(1.170)
1.036
(1.175)
1.152
(1.568)
-3.704
(1.399)*
2.685
(1.279)**
3.956
(1.707)**
91.146
(2.418)*
-2.048
(1.107)***
0.151
(1.030)
-0.001
(1.155)
-0.889
(0.884)
1.453
(1.811)
1.308
(0.615)**
1.314
(1.114)
0.837
(0.967)
-2.055
(0.666)*
1.701
(0.736)**
-0.048
(0.678)
0.188
(0.666)
0.068
(0.873)
-2.216
(0.805)*
1.124
(0.713)
1.175
(0.942)
91.147
(2.417)*
-2.048
(1.107)***
0.151
(1.030)
-0.001
(1.155)
-0.889
(0.884)
1.453
(1.810)
1.308
(0.615)**
1.313
(1.113)
0.836
(0.966)
-2.055
(0.666)*
1.701
(0.736)**
-0.049
(0.678)
0.188
(0.666)
0.067
(0.873)
-2.216
(0.804)*
1.124
(0.713)
1.176
(0.942)
91.135
(2.421)*
-2.052
(1.109)***
0.151
(1.031)
-0.001
(1.157)
-0.890
(0.885)
1.455
(1.813)
1.315
(0.616)**
1.320
(1.115)
0.842
(0.968)
-2.053
(0.667)*
1.703
(0.737)**
-0.046
(0.679)
0.191
(0.667)
0.074
(0.875)
-2.222
(0.806)*
1.1246
(0.714)
1.169
(0.943)
Sole Prop.
Partnerships
Other
Private
Corporation
Other Industry
Service
Agriculture
Construction
Small: <50
employees
Large: >500
employees
<5 years of age
>15 years of age
No Competitors
>3 Competitors
Foreign Owned
Gov. Owned
47
Export
Audits
Obstacle:
Financing
Obstacle:
Political
Instability
Obstacle:
Organized
Crime
Obstacle: Taxes
& Regs.
Obstacle:
Corruption
Obstacle:
Inflation
Obstacle:
Exchange Rate
Obstacle: Laws
& Regs.
In consistent
Country
Dummies
σ
(s.e.)
Mean ŷ *
(s.e.)
[min; max]
# >100
Pseudo LogLikelihood
Value
McKelvey &
Zavoina Pseudo
R2
0.630
(0.971)
5.556
(0.986)*
-
0.614
(0.937)
5.361
(0.952)*
-
0.614
(0.935)
5.351
(0.950)*
-
0.826
(0.544)
3.252
(0.560)*
-
0.826
(0.544)
3.252
(0.560)*
-
0.829
(0.545)
3.257
(0.561)*
-
0.687
(1.125)
5.581
(1.155)*
-2.277
(1.058)**
-0.031
(1.152)
0.691
(1.091)
5.394
(1.119)*
-2.214
(1.025)**
-0.020
(1.116)
.691
(1.089)
5.384
(1.117)*
-2.210
(1.024)**
-0.0186
(1.114)
0.883
(0.615)
3.357
(0.642)*
-1.124
(0.582)**
-0.389
(0.636)
0.883
(0.616)
3.357
(0.642)*
-1.124
(0.582)**
-0.389
(0.636)
0.883
(0.617)
3.362
(0.643)*
-1.127
(0.583)**
-0.389
(0.637)
-
-
-
-
-
-
-
-
-
-
-
-
-1.787
(1.067)***
1.728
(1.03)***
-1.724
(1.03)***
-0.687
(0.601)
-0.687
(0.601)
-0.689
(0.602)
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-5.187
(1.188)*
-11.504
(1.065)*
0.388
(1.236)
-1.729
(1.136)
1.701
(0.976)***
-5.037
(1.151) *
-11.181
(1.029)*
0.381
(1.198)
-1.701
(1.101)
1.652
(0.95)***
-5.028
(1.149) *
-11.165
(1.028)*
0.381
(1.196)
-1.698
(1.099)
1.649
(0.95)***
-2.561
(0.650)*
-6.808
(0.606)*
0.617
(0.683)
-1.263
(0.627)**
0.768
(0.542)
-2.561
(0.650)*
-6.808
(0.605)*
0.617
(0.683)
-1.264
(0.627)**
0.769
(0.542)
-2.563
(0.651)*
-6.819
(0.607)*
0.616
(0.684)
-1.260
(0.628)**
0.769
(0.543)
-
-
-
-
-
-
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
30.487
(0.416)
91.153
(17.213)
[21; 142]
2016
29.591
(0.375)
91.100
(16.470)
[30; 141]
1960
29.545
(0.374)
91.098
(16.436)
[30; 141]
1960
18.749
(0.170)
82.058
(9.810)
[35;104]
47
18.749
(0.170)
82.060
(9.804)
[36;104]
47
18.778
(0.172)
82.043
(9.858)
[36;104]
47
29.763
(0.477)
92.584
(18.062)
[5; 144]
1761
28.989
(0.434)
92.520
(17.332)
[19; 142]
1724
28.948
(0..433)
92.520
(17.299)
[20; 142]
1721
18.083
(0..191)
82.915
(10.021)
[25;105]
66
18.080
(0..191)
82.916
(10.015)
[26;105]
66
18.102
(0..192)
82.899
(10.083)
[23;105]
66
-11829.6
-11866.4
-11869.3
-21158.4
-21158.4
-21156.2
-8508.7
-8532.4
-8534.29
-15648.8
-15649.0
-15647.5
0.098
0.138
0.138
0.215
0.215
0.161
0.108
0.149
0.149
0.235
0.235
0.180
48
LRT-OS
[d.o.f.; P-Value]
1648.5
[96; 0.00]
1637.2
[96;0.00]
1636.8
[96;0.00]
1674.9
[96;0.00]
1674.9
[96;0.00]
1677.6
[96;0.00]
1364.45
[102;0.0]
1356.75
[102;0.0]
1356.54
[102;0.0]
1367.43
[102;0.00]
1367.20
[102;0.00]
Pseudo
4.557
9.989
Functional Form
[.0328]
[0..002]
Test
[P-Value]
Homoskedastici
1520.841
1727.343
[0.000]
[0.000]
ty Test of
Unknown Form
[P-Value]
Normality Test
2448.197
1490.885
[P-Value]
[0.000]
[0.000]
Observations
7320
7320
7320
7320
7320
7320
5393
5393
5393
5393
5393
NOTES: § ***,** and * denote statistical significance at the 10%, 5% and 1% level respectively.
± d.o.f. denotes degrees of freedom.
‡ Omitted category is 1 “Public Corporations”, 2 “Manufacturing”, 3 “Medium”, 4 “Between 5 and 15”, 5 “Between 1 and 3”, 6 “United States”.
₣ Standard errors (s.e.) are noted in parenthesis.
1369.38
[102;0.00]
-
-
5393
49
Table 6: Estimation Results for Semi-log Model
Variable
Model 5a
1<lnYi*<100
Constant
Sole Prop.
Partnerships
Other
Private
Corporation
Other Industry
Service
Agriculture
Construction
Small: <50
employees
Large: >500
employees
<5 years of age
>15 years of age
No Competitors
>3 Competitors
Foreign Owned
Gov. Owned
Export
Audits
Obstacle:
Financing
Obstacle:
Political
Instability
Obstacle:
Organized
Crime
Obstacle: Taxes
& Regs.
Model 11a
1<lnYi*<100
1.898
(0.153)*
0.267
(0.073)*
0.106
(0.068)
0.040
(0.075)
0.128
(0.058)**
0.017
(0.105)
-0.055
(0.040)
-0.087
(0.071)
0.068
(0.063)
0.178
(0.043)*
-0.135
(0.049)*
Model 5b
1<lnYi*<100
Het. Corr.
1.273
(0.150)*
0.225
(0.066)*
0.065
(0.062)
0.070
(0.072)
0.094
(0.052)***
0.110
(0.088)
-0.041
(0.037)
-0.041
(0.065)
0.079
(0.058)
0.179
(0.040)*
-0.131
(0.0453)*
1.185
(0.182)*
0.211
(0.082)*
0.031
(0.077)
0.040
(0.086)
0.131
(0.066)**
-0.069
(0.136)
-0.070
(0.046)
-0.096
(0.083)
0.029
(0.072)
0.174
(0.050)*
-0.163
(0.055)*
Model 11b
1<lnYi*<100
Het. Corr.
0.934
(0.155)*
0.340
(0.070)*
0.060
(0.065)
0.118
(0.083)
0.132
(0.059)**
0.030
(0.125)
-0.086
(0.043)**
-0.087
(0.080)
0.041
(0.069)
0.201
(0.048)*
-0.193
(0.051)*
-0.037
(0.044)
-0.066
(.043)
-0.159
(0.059)*
0.143
(0.051)*
-0.203
(0.046)*
-0.121
(0.061)**
-0.037
(0.040)
-.238
(0.041)*
-
-0.054
(0.039)
-0.079
(0.040)**
-0.175
(0.056)*
0.116
(0.047)**
-0.173
(0.043)*
-0.117
(0.056)**
-0.048
(0.036)
-0.192
(0.038)*
-
-
-
-0.041
(0.050)
-0.063)
(0.050)
-0.064
(0.065)
0.148
(0.060)**
-0.124
(0.053)**
-0.150
(0.071)**
-0.049
(0.046)
-0.239
(0.048)*
0.096
(0.04)**
-0.012
(0.048)
-0.057
(0.027)**
-0.059
(0.045)
-0.119
(0.062)**
0.084
(0.053)
-0.011
(0.036)
-0.154
(0.065)**
-0.024
(0.042)
-0.150
(0.042)*
0.045
(0.035)
-0.0154
(0.036)
-
-
0.092
(0.045)**
(0.065)
(0.042)
-
-
0.212
(0.049)*
0.202
(0.046)*
50
Obstacle:
Corruption
Obstacle:
Inflation
Obstacle:
Exchange Rate
Obstacle: Laws
& Regs.
In consistent
Country
Dummies
σ
(s.e.)
Pseudo LogLikelihood
Value
McKelvey &
Zavoina Pseudo
R2
LRT-OS
[d.o.f.; P-Value]
-
-
-
-
-
-
-
-
Yes
0.527
(0.045)*
-0.013
(0.051)
0.068
(0.047)
-0.067
(0.04)***
0.534
(0.046)*
0.028
(0.044)
-0.057
(0.027)**
-0.046
(0.039)
Yes
Yes
Yes
1.379
(0.012)
-
1.347
(0.013)
-
-16242.508
-16010.435
-11788.379
-11614.318
0.200
1589.549
[96; 0.000]
0.228
4475.80
[96; 0.000]
1357.86
[102;0.000]
16652.51
[102; 0.000]
Pseudo
0.757
0.001
Functional Form
[0.384]
[0.979]
Test
[P-Value]
Homoskedastici
1062.329
1111.411
ty Test of
[0.000]
[0.000]
Unknown Form
[P-Value]
Normality Test
3409.997
1699.6542
[P-Value]
[0.000]
[0.000]
Observations
7320
7320
5393
5393
NOTES: § ***,** and * denote statistical significance at the 10%, 5% and 1% level respectively.
± d.o.f. denotes degrees of freedom.
‡ Omitted category is 1 “Public Corporations”, 2 “Manufacturing”, 3 “Medium”, 4 “Between 5 and 15”,
“Between 1 and 3”, 6 “United States”.
₣ Standard errors (s.e.) are noted in parenthesis.
5
51
18.27%
100%
10.74%
16.56%
6.659%
44.43%
33.65%
90-99%
Latin America
10.85%
Source: World Business Environment Survey
8.562%
6.435%
6.603%
11.13%
9.615%
9.615%
6.868%
Africa and Middle East
80-89%
18.83%
70-79%
OECD
15.09%
3.247%
2.078%
2.857%
5.584%
11.7%
9.481%
9.007%
5.312%
8.083%
8.16%
Transition Europe
57.92%
42.65%
7.967%
14.28%
11.09%
5.417%
60-69%
8.097%
9.198%
11.16%
14.53%
50-59%
8.031%
<50%
38.69%
32.7%
Former Soviet Union
5.503%
11.08%
22.25%
Asia
Figure 1: Percentage of Sales Reported to Tax Authorities
FIGURES
52
Figure 2: Histograms of Pseudo Residuals
0
.2
Density
.4
.6
Histogram of Pseudo Residuals for Model 5a
-2
-1
0
Pseudo Residuals
1
2
0
.2
Density
.4
.6
Histogram of Pseudo Residuals for Model 11a
-2
-1
0
Pseudo Residuals
1
2
53
REFERENCES
Allingham, M. and A. Sandmo. 1972. Income Tax Evasion: A Theoretical Analysis. Journal of
Public Economics 1:323-338.
Andreoni, James. 1992. IRS as Loan Shark: Tax Compliance with Borrowing Constraints.
Journal of Public Economics 49:35-46.
______. Brian Erard, and Jonathan Feinstein. 1998. Tax Compliance. Journal of Economic
Literature 36:818-860.
Apel, Mikael. 1994. An Expenditure-Based Estimate of Tax Evasion in Sweden. Working Paper,
Department of Economics, Uppsala University.
Batra, Geeta, Daniel Kaufmann and Andrew H.W. Stone. 2003. Investment Climate Around the
World: Voices of the Firms from the World Business Environment Survey. Washington,
D.C.: The World Bank.
Border, Kim C. and Joel Sobel. 1987. Samurai Accountant: A Theory of Audit and Plunder.
Review of Economic Studies 54:525-540.
Breham, J.W (1966) A Theory of Psychological Reactance. Academic Press: New York.
Chan, K. Hung and Phyllis Lai Lan Mo. 2000. Tax Holidays and Tax Non-compliance: An
Empirical Study of Corporate Tax Audits in China’s Developing Economy. The
Accounting Review 75:469-484.
Chen, Kong-Pin and C. Y. Cyrus Chu. 2005. Internal Control and External Manipulation: A
Model of Corporate Income Tax Evasion. RAND Journal of Economics 26:151-164.
Chesher, A. and M. Irish. 1987. Residual Analysis in the Grouped and Censored Normal
Regression Model. Journal of Econometrics 34:33–61.
Christian, Charles W. 1994. Voluntary Compliance with Individual Income Tax: Results from
the 1988 TCMP study. The IRS Research Bulletin Publication 1500 9:35-42.
Cowell, Frank. 2004. Sticks and Carrots. In The Crises in Tax Administration , eds. Henry Aaron
and Joel Slemrod, 230-258. Washington, D.C.: The Brookings Institution Press.
Cremer, Helmuth, Maurice Marchand and Pierre Pestieau. 1990. Evading, Auditing and Taxing:
the Equity-Compliance Tradeoff. Journal of Public Economics 43:67-92.
Crocker, Joel and Joel Slemrod. 2005. Corporate Tax Evasion with Agency Costs. Journal of
Public Economics 89: 1593-1610.
54
Deregulation Commission (1991) Opening of Markets and Competition, Report Presented to
German Federal Government, Bonn.
Erard, Brian. 1992. The Influence of Tax Auditing on Reporting Behaviour. In Why People Pay
Taxes, ed. Joel Slemrod. Ann Arbor: University of Michigan.
______ and Chih-Chin Ho. 2001. Searching for Ghosts: Who are the Nonfilers and How Much
Tax do they Owe? Journal of Public Economics 81:25-50.
Giles, David E.A. 2000. Modeling the Tax Compliance Profiles of New Zealand Firms. In
Taxation and the Limits of Government , eds. Gerald W. Scully and Patrick J. Caragata,
243-270. Boston: Kluwer Academic Publishers.
______ and Lindsay M. Tedds. 2002. Taxes and the Canadian Underground Economy.
Toronto: The Canadian Tax Foundation.
Hanlon, Michelle, Lillian Mills, and Joel Slemrod. 2005. An Empirical Examination of
Corporate Tax Non-compliance. Working Paper, University of Michigan.
Harvey, Andrew. 1976. Estimating Regression Models with Multiplicative Heterocedasticity.
Econometrica 44: 461-65.
Huber, P.J. 1967. The Behaviour of Maximum Likelihood Estimates under Non-standard
Conditions. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics
and Probability, 221-223. Berkeley: University of California Press.
Internal Revenue Service. 2004. Interactive Tax Gap Map. National Headquarters Office of
Research.
Johnson, Simon Daniel Kaufamnn and Pablo Zoido-Lobatόn (1998a) “Regulatory Discretion and
the Unofficial Economy,” 88 American Economic Revue 387-392.
______ (1998b) Corruption, Public Finances and the Unofficial Economy. MIT Press:
Cambridge.
Johnson, Simon, Daniel Kaufmann, John McMillan, and Christopher Woodruff. 2000. Why do
firms hide? Bribes and Unofficial Activity After Communism. Journal of Public
Economics 76:495-520.
Joulfaian, David. 2000. Corporate Tax Evasion and Managerial Preferences. Review of
Economics and Statistics 82:698-701.
______ and Mark Rider. 1998. Differential Taxation and Tax Evasion by Small Business.
National Tax Journal 51:676-87.
55
Kennedy, P. 1981. Estimation with Correctly Interpreted Dummy Variables in Semilogarithmic
Equations. American Economic Review 71: 801.
Lyssiotou, Panayiota, Panos Pashardes and Thanasis Stengos. 2004. Estimates of the Black
Economy Based on Consumer Demand Approaches. Economic Journal 114:622-640.
Machin, S., and M. Stewart. 1990. Unions and the Financial Performance of British Private
Sector Establishments. Journal of Applied Econometrics 5:327–350.
McElvey, W. and Zavoina, R. 1975. A Statistical Model for the Analysis of Ordinal Level
Dependent Variables. Journal of Mathematical Sociology 4:103-120.
Mirus, Rolf and Roger S. Smith. 1997. Self-Employment, Tax Evasion, and the Underground
Economy: Micro-Based Estimates for Canada. Working Paper, International Tax
Program, Harvard Law School.
Mookherjee, Dilip and Ivan P.L. P’ng. 1989. Optimal Auditing, Insurance and Redistribution.
Quarterly Journal of Economics 104:399-415.
Orme, C. 1990. The Small-Sample Performance of the Information Matrix Test. Journal of
Econometrics 46:309–331.
Papke, Leslie E. and Jeffrey M. Wooldridge. 2005. Panel Data Methods for Fractional
Response Variables with an Application to Test Pass Rates. Working Paper, Department
of Economics, Michigan State University.
Pissarides, Christopher A., and Guglielmo Weber. 1989. An Expenditure-based Estimate of
Britain’s Black Economy. Journal of Public Economics 39:17-32.
Ramsey, J.B. 1969. Tests for Specification Errors in Classical Linear Least-Squares Regression
Analysis. Journal of the Royal Statistical Society 31:350 – 71.
Rice, Eric M. 1992. The Corporate Tax Gap: Evidence on Tax Compliance by Small
Corporations. In Why People Pay Taxes, ed. Joel Slemrod, 125-166. Ann Arbor:
University of Michigan Press: Ann Arbor.
Sanchez, Isabel and Joel Sobel. 1993. Hierarchical Design and Enforcement of Income Tax
Policies. Journal of Public Economics 50:345-369.
Schuetze, Herb J. 2002. Profiles of Tax Non-compliance Among the Self-Employed in Canada,
1969-1992. Canadian Public Policy XXVIII: 219-237.
Schnieder, Friedrich and Dominik Enste (2000) “Increasing Shadow Economies All Over the
World – Fiction or Reality? A Survey of the Global Evidence of Their Size and Their
Impact From 1970 to 1995.” 38 Journal of Economic Literature 77-114.
56
Scotchmer, Suzanne. 1987. Audit Classes and Tax Enforcement Policy. American Economic
Review 77:229-233.
Slemrod, Joel. 2004. The Economics of Corporate Tax Selfishness. Working Paper 10858,
National Bureau of Economic Research.
______ and Shlomo Yitzhaki. 1996. The Cost of Taxation and the Marginal Efficiency Cost of
Funds. International Monetary Fund Staff Papers 43:25-34.
Smith, J. and T. Adams. 1987. The Measurement of Selected Income Flows in Informal Markets,
1981 and 1985-1986. Working Paper, University of Michigan Survey Research Centre.
Smith, Stephen with Christopher A. Pissarides and Guglielmo Weber. 1986. Evidence from
Survey Discrepancies. In Britain’s Shadow Economy, ed. Stephen Smith, 137-153.
Oxford: Clarendon Press.
Stewart, M. 1983. On Least Squares Estimation when the Dependent Variable is Grouped.
Review of Economic Studies 50:737-753.
Tannenbaum, Robert. 1993. Corporate Tax Disclosure: Good or Bad for the Commonwealth?
Working Paper, Massachusetts Special Commission on Business Tax Policy.
Tedds, Lindsay M. 2005. Nonparametric Expenditure-based Estimation of Income Underreporting and the Underground Economy. Working Paper, Department of Economics,
McMaster University.
Trandel, Greg and Arthur Snow. 1999. Progressive Income Taxation and the Underground
Economy. Economics Letters 62:217-222.
Veall, Michael R. and Kurt F. Zimmermann. 1996. Pseudo-R2 Measures for Some Common
Limited Dependent Variable Models. Journal of Economic Surveys 10:241-259.
Watson, Harry. 1985. Tax Evasion and Labor Markets. Journal of Public Economics 27:231246.
World Bank. 1997. World Development Report 1997: The State in a Changing World.
Washington, D.C.: World Bank Group.
Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Section and Panel Data.
Cambridge: The MIT Press.
57
APPENDIX
A.
SURVEY DESIGN1
The World Business Environment Survey (WBES) was conducted over a 20 month
period between the latter half of 1998 and mid-2000. The WBES consists of a core survey
instrument alongside a subsidiary module that addresses issues of importance to particular
regions. The core survey covers a broad range of measures for business environment and firm
attributes. The survey was completed by over 10,000 firms across eighty countries and the West
Bank and Gaza. Approximately, one third of these countries were low income, one half were
middle income and a little less than one fifth were high income. There were seven countries for
East Asia, twenty-three from Eastern Europe and Central Asia, twenty from Latin America and
the Caribbean, nineteen from the Middle East and North Africa, nine from the Organization for
Economic Co-operation and development (OECD), and three from South Asia. In Africa,
surveys were conducted primarily by mail while direct interviews were the predominant method
used for all other firms in the sample. Response rates were lowest in Africa but generally
satisfactory in other region.
The WBES used a form of stratified random sampling along with oversampling and
nonrandom methods to correct for biases in the representation across firms with respect to
industry characteristics that were not common. To ensure adequate representation of firms by
industry, size, ownership, export orientation, and location, the following sampling targets were
agreed on across all regions:
ƒ
The number of manufacturing versus service firms were allocated according to their
contribution to GDP, with a 15 percent minimum for each type of firm
1
The information presented here is based on that provided in Batra et al (2003).
58
ƒ
At least 15 percent of the sample was in the small category (fewer than 50
employees) and at least 15 percent was in the large category (more than 500
employees)
ƒ
At least 15 percent of the companies in the sample were firms with foreign content
ƒ
At least 15 percent of firms exported at least 20 percent of their output
ƒ
At least 15 percent of firms were located in small towns or countryside
In addition, the sample size for each country or territory was greater than or equal to 100.
Of the firms in the final sample, twenty percent of the firms were large (more than 500
employees) with the remaining 80 percent spit almost equally between small (fewer than 50
employees) and medium sized firms. Firms were also groups into five broad industry categories
as follow: 43% in services and commerce; 35% in manufacturing; 9% in construction; 7% in
agriculture and 4% in other. Firms that fell into the other category were largely from the SubSahara and North Africa. The majority of firms were domestically owned and privately owned.
One third of the firms were export oriented, with the highest proportion in the Middle East and
North Africa. Finally, the most predominant form of organization were privately held
corporations followed by sole proprietorships, partnerships, other, public corporations, and
finally cooperatives, respectively.
Unfortunately, the survey does not contain survey weights. The lack of survey weights
means that individual indicators should not be used for precise country rankings in any particular
dimension measured by the survey.
59