# Individual Income Tax

> Briefs, arguments, decisions, and more.

URL: https://www.frixlaw.com/law-library/documents/agency%3Airs%3A848e5e1d472acc56

## Record

- **Collection:** Agency decision
- **Document type:** Agency decision

## Text

The Determinants of
Individual Income Tax
Compliance

Estimating The Impacts of
Tax Policy, Enforcement, and IRS Responsiveness

Department of the Treasury
Internal Revenue Service
Publication 1916 (Rev. 11-96)
Catalog Number 22555A

Department
of the
Treasury
Internal
Revenue
Service

The Determinants of
Individual Income Tax
Compliance
Estimating the Impacts of
Tax Policy, Enforcement, and
IRS Responsiveness
November 1996

Alan H. Plumley
IRS Research Division

This is a copy of a Ph.D dissertation prepared for Harvard
University, slightly modified for distribution by the Internal
Revenue Service. The estimates presented herein are subject
to change. In fact, this revision is the second IRS printing,
and includes corrections of a number of typographical errors
contained in the August 1996 printing. Other refinements are
being made continually. Comments and questions are
welcome, and may be directed to Alan Plumley as follows:
Mail: IRS Research Division CP:R:R:AR:E
1111 Constitution Ave. NW, Washington, DC 20224
Phone/VoiceMail: (202) 874-0508
Fax:

(202) 874-0634

E-mail:AHPLUM50@M1.IRS.GOV

Suggested Citation
Internal Revenue Service
The Determinants of Individual Income Tax Compliance:
Estimating The Impacts of Tax Policy, Enforcement, and IRS
Responsiveness
Publication 1916 (Rev. 11-96)
Washington, DC: 1996

ABSTRACT
This paper presents an econometric analysis of the impact of a wide variety of
potential determinants of voluntary compliance with individual income tax filing and
reporting obligations. Based on perhaps the richest dataset yet compiled (by state
and year, from 1982 through 1991), including data on taxpayer behavior, IRS
actions, and other factors, the analysis finds significant compliance effects
attributable to many tax policy and tax administration parameters, including: audits;
the matching of third-party information documents; the issuance of targeted nonfiler
notices; criminal tax convictions; marginal tax rates; the burden associated with
completing the myriad tax forms and schedules; and the preparation of returns by
the IRS Taxpayer Service function.

The Determinants of Individual Income Tax Compliance:
Estimating The Impacts of Tax Policy, Enforcement, and IRS Responsiveness

Contents

1. Introduction . . . . . . . . . . .
1.1 Background . . . . . . . . .
1.2 Previous Research . . . . . . .
1.3 Advances Made By This Research .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

1
1
2
3

2. The Model . . . . . .
2.1 The Data . . . . .
2.2 Estimation Approach
2.3 Model Specification

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

6
7
20
24

3. Results . . . . . . . . . . . . . . . . . . . . .
3.1 Filing Compliance . . . . . . . . . . . . . . .
3.2 Reporting Compliance . . . . . . . . . . . . .
3.3 Alternative Definitions of Income and Offsets . . . .
3.4 Other Potential Determinants of Voluntary Compliance .
3.5 Implications for Resource Allocation . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

27
28
30
38
38
39

4. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . .

42

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . .

44

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Appendices
A. Data Sources and Derivations . . . . . . . . . . . . . . . . . .
B. Method for Using the Current Population Survey to Estimate the Number
of Returns Required to be Filed . . . . . . . . . . . . . . . .
C. Three Definitions of Income and Offsets . . . . . . . . . . . . . .
D. Personal Income Compared With Reportable Income . . . . . . . . .
E. Descriptive Statistics . . . . . . . . . . . . . . . . . . . . .
F. Estimation of the Burden Associated With Tax Forms . . . . . . . . .
G. Estimated Revenue Impact of the Decline in Audit Coverage . . . . . .
H. Interactions With the AuditRate Variable . . . . . . . . . . . . . .
I. Derivation of Total and Marginal Indirect Revenue Functions
for Five IRS Activities . . . . . . . . . . . . . . . . . . .

. .

46

.
.
.
.
.
.
.

.
.
.
.
.
.
.

63
67
79
80
82
85
86

. .

90

The Determinants of Individual Income Tax Compliance:
Estimating The Impacts of Tax Policy, Enforcement, and IRS Responsiveness

1. Introduction
For many years, the Internal Revenue Service (IRS) has presumed that its activities
promote better income tax compliance in the general population—both through deterrence and
taxpayer service—but it has never been able to quantify this impact, or even verify that it exists.
This paper does both, providing the first empirical basis for choosing the best combination of
major IRS activities to improve voluntary filing and reporting compliance among individuals—a
capability that is increasingly needed in this climate of declining budgets. It also estimates the
compliance impact of important tax policy parameters—most notably marginal tax rates.
The success in estimating these compliance effects is not the result of sophisticated, new
statistical techniques, but rather the creative application of straightforward techniques to the right
data, which took eight years to compile from IRS reports and databases, as well as numerous other
sources. The analysis uses panel data over a ten-year period (1982-91) aggregated to the state level
to estimate one filing compliance equation and three separate reporting compliance equations.
Among the findings: the deterrent effect of audits in the general population is about 11 times as
large as the adjustments proposed by the audits themselves, but nonfiler notices, information
document matching, and return preparation assistance are more cost-effective in boosting revenue.
1.1 Background
The federal income tax system operates on a self-assessment basis. That is, the
government expects taxpayers to determine their own tax obligations and to pay voluntarily
whatever is due—both regularly (through withholding from wages and through estimated tax
payments, if necessary) and at year end (by filing tax returns and paying any additional balances
due). By placing the onus on taxpayers, the government avoids the costly alternative of
determining each individual’s tax liability and doing whatever it must to collect it.
However, one cost of relying so heavily on the voluntary compliance of taxpayers is that
not all tax is voluntarily paid. The IRS estimates that the gross individual income tax gap (the
difference between what taxpayers should pay and what they actually do pay voluntarily and
timely) was about $94 billion for Tax Year 1992—about 20 percent of total individual income tax
receipts and over half the size of the budget deficit!1
Congress has taken several steps to strengthen voluntary compliance with the income tax
laws. These actions fall into two major categories: requirements (e.g., requiring withholding of
tax at the source of income and requiring third party information reporting) and deterrents (e.g.,
giving the IRS certain enforcement powers, and stipulating the penalties that those who are caught
through this enforcement must pay).
Probably the most widely known example of an IRS deterrent is the “audit” of an
individual’s tax return. In Fiscal Year 1992, for example, IRS completed examinations of just
over 1 million tax returns of individuals. Although these returns were for several tax years of
liability, this is approximately the same number of examinations conducted on Tax Year 1992
1

IRS (1990).

2

The Determinants of Individual Income Tax Compliance

returns, which numbered about 115 million—implying an audit coverage rate of roughly 0.9
percent. As a result of the examinations completed in FY 1992, IRS recommended that the
taxpayers in question pay additional tax and penalties totaling $6.0 billion.2 The additional
revenue that the government will collect from this “yield” (following all appeals, litigation, and
collection efforts) is the direct revenue effect of those 1 million examinations. However, it is quite
likely that the audits also had an indirect revenue effect—inducing some amount of voluntary
compliance in the population at large through the general deterrent effect of the examinations
(referred to at IRS as the “ripple effect” of the examinations), and (perhaps) by influencing the
voluntary compliance of the contacted taxpayers in subsequent years (referred to as the
“subsequent year effect”). Indeed, one of the purposes of IRS enforcement is establishing a
credible deterrent to noncompliance. It has generally been believed, for example, that many
taxpayers would perceive increased auditing by IRS as an increase in their chances of being
audited, and that they would improve their voluntary compliance as a result.
In addition to the various requirements and deterrents that Congress has implemented
specifically to improve voluntary compliance, other actions or laws may have influenced
compliance indirectly. Examples of this phenomenon may include a change in compliance
behavior resulting from a change in tax policy (e.g., the marginal tax rate structure), or from some
change in the public’s attitude toward the IRS (which may arise from changes in IRS’s
responsiveness to taxpayers’ needs).
The focus of this paper is the indirect behavioral response of taxpayers (as measured by
changes in their voluntary filing of required income tax returns, and their reporting of income and
offsets to income on those returns) to changes in IRS enforcement, IRS’s responsiveness, and
basic tax policies.3 Quantifying these responses could help to shape tax policy and tax
administration for the foreseeable future—especially given the need to reduce the budget deficits
and to make the best use of government resources.
1.2 Previous Research
The indirect revenue effect of audits is beginning to receive attention from researchers, but
little—if any—empirical research has been done to quantify the separate compliance effects of
enforcement, tax policy and IRS responsiveness. What makes such research challenging, of
course, is that the compliance impact of government actions is never observed in isolation; it can
only be estimated. Voluntary income tax compliance is probably determined by a wide variety of
factors that interact differently for each individual. Although many such factors have been
suggested, and several studies have focused on some of them, nothing has emerged to guide
policy-makers concerning the relative merits of alternative approaches to improving compliance.
This may be due in part to the fact that no comprehensive theory exists that explains the compliance
behavior of taxpayers. It may also be because very little data are available to test such theories.
One of the earliest attempts to model taxpayer compliance was done by Allingham and
Sandmo (1972), which applied the utility-maximization approach of Becker (1967) to tax
compliance. Allingham and Sandmo’s simple model predicts the intuitive result that taxpayers will
voluntarily report more income in response to either an increase in the probability of being
detected, or an increase in the penalty imposed on those who are caught. However, the model is
inconclusive in predicting the response to an increase in the tax rate; the net response is the sum of
two terms in their model—one negative (suggesting a decrease in income reported as the tax rate
increases), and another, which is most probably positive, assuming that taxpayers’ risk aversion
2
3

IRS (1992).

Strictly speaking, my analysis quantifies such behavioral responses, but it does not address why taxpayers behave
the way they do. Although traditional deterrence mechanisms may be responsible, this study cannot prove it.

The Determinants of Individual Income Tax Compliance

3

decreases with income. Allingham and Sandmo liken these two terms to a positive income effect
and a negative substitution effect. They reason that the substitution effect means that an increase in
the tax rate makes it more profitable to underreport income at the margin (i.e., the higher the tax
rate, the more money is retained when one underreports a dollar of income). The income effect,
however, is most likely positive because an increase in the tax rate reduces net income, and
assuming decreasing absolute risk aversion, the taxpayer is less willing to underreport income than
before. Which of these two effects is stronger depends in part on the taxpayer’s degree of risk
aversion.
This early model—and virtually everything that has followed—says nothing about IRS
responsiveness to taxpayers, or about nonfiling. Roth, Scholz, and Witte (1989) gives an
excellent overview of the theoretical and empirical work on tax compliance through the 1980’s, but
several empirical studies are worth mentioning here. At least three studies (Clotfelter (1983), Cox
(1984), and Dubin, Graetz and Wilde (1990)) have explored the possibility that the actual marginal
tax rate to which individuals respond is the sum of the federal and state marginal rates, and that
since the federal income tax law is applied uniformly across states, using the state marginal tax rate
alone has equivalent explanatory power. The primary reason for making this assumption arises
from econometric pragmatism: to gain cross-sectional variation in the marginal tax rate variable.
However, for panel studies such as this one (i.e., a time series of cross sections), the difficulties
inherent in creating a state marginal tax rate variable, and the opportunities associated with the
variation in federal tax rates over time, caused me to explore the role of federal tax rates alone.
Four econometric studies that attempted to estimate the indirect effect of audits on
compliance are worth noting. Erard (1992) focused solely on the “subsequent-year effect” of
audits on the reporting compliance of those who were audited, and reports inconclusive results.
Three other studies examined the general deterrent effect of audits. 4 Tauchen, Witte, and Beron
(1989) was a cross-sectional study based on the 1979 Taxpayer Compliance Measurement
Program (TCMP) micro database, and concluded that the impact of audits is weak at best. The
other two studies were aggregate analyses. Dubin, Graetz and Wilde (1990) used panel data
(aggregated at the state level, 1977-1986) to estimate both reported tax per return and returns filed
per capita, and finds a very large and significant deterrent effect of audits. Beron, Tauchen, and
Witte (1992) uses 1969 data aggregated at the 3-digit ZIP Code level to estimate average Adjusted
Gross Income and average tax reported, and finds a weak indirect effect of audits that is limited to
certain taxpayer groups.
1.3 Advances Made By This Research
This study improves upon all of the earlier work in this area. It is based on one of the most
comprehensive datasets ever compiled on the potential determinants of voluntary compliance, and
provides an empirical basis for choosing the best mix of strategies for improving and maintaining
voluntary compliance.
Before describing the details of my model (section 2) and the results (section 3), I provide
in this section an overview of the advances made by this research. These advances are almost
entirely in the realm of improved data and an improved econometric specification; no new
estimating procedures are developed. Since my approach is most similar to that of Dubin, Graetz
and Wilde (1990) (which I will hereafter refer to as DGW), I will describe the advances with
respect to that important work. There are several obvious similarities between this study and
4

Actually, these studies examined the combined indirect effect of audits on the compliance of those who were not audited
(the “ripple” effect) as well as of those who were audited (the “subsequent-year” effect). But since the methodologies could
not distinguish between these effects, and since the “ripple” effect is presumed to dominate, the studies can be thought to
examine general deterrence.

4

The Determinants of Individual Income Tax Compliance

DGW. Each attempts to estimate the impact of various factors (including audits and tax rates) on
the voluntary income tax filing and reporting compliance of individuals using a 10-year panel of
data aggregated to the state level. Each accounts for the endogeneity of audit rates. However,
there are many important differences between these studies. I describe these differences—and
explain why they represent improvements—below.
1.3.1 Dependent Variables
DGW estimates three equations: one for reported tax per return (a measure of reporting
compliance), one for returns filed per capita (reflecting filing compliance), and one for “assessed
liability” (reported tax plus additional tax and penalties proposed by audits) per return. These
dependent variables control for cross-sectional and time variations by dividing by the number of
returns filed or by population. My dependent variables use as denominators exogenous surrogates
for reporting and filing obligations, giving them a meaning closer to traditional measures of
voluntary compliance. My estimation of filing obligations by state from Census data, for example,
creates a very powerful measure of the filing rate—obviating the need for such DGW variables as
the number of households per capita, and the percent of households on welfare. Moreover, using
reported tax per return filed as a dependent variable makes it difficult to interpret the results since
the explanatory variables could conceivably influence both the numerator and the denominator.
For example, a positive coefficient could indicate that the explanatory variable increases reporting
compliance, but it is also possible that it increases average tax reported by decreasing filing
compliance among low-income taxpayers. The implications of those two possibilities are
dramatically different. In fact, one may initially view the latter possibility as the correct
interpretation of DGW’s positive coefficient on audit rate in the reporting equation and its negative
coefficient on audits in the filing equation—especially given DGW’s assessment that “one way to
escape audits has been simply not to file.” DGW reports, however, that the magnitudes of these
coefficients is such that the net effect of audits is to increase dollars reported—but extraneous side
calculations are necessary to conclude this, illustrating the cumbersome nature of the specification.
The numerators are also quite different in the two studies. DGW uses the tax reported on
returns as its measure of voluntary reporting compliance. I avoid tax as a measure of reporting
compliance because several of the potential determinants of voluntary compliance (e.g., marginal
tax rates, filing thresholds, marital status, and allowable child exemptions) also have a direct role in
the calculation of tax from gross income, making it difficult to separate their impact on compliance.
I use three more useful measures instead: total income reported, total offsets reported, and net
income reported (income minus offsets). These three equations also have the advantage of
providing insight into the major forms of noncompliance (underreporting income vs. overstating
offsets to income or to tax), and they allow consistency comparisons across equations (since
income minus offsets equals net income). I have also controlled for the extent to which the tax
rules have changed concerning the amount of income that must be reported and the amount that
may be claimed as offsets. I have done this largely by defining income and offsets in three
different ways, and by estimating separate equations for each definition.
My data also reflect two important qualitative improvements. First, DGW uses dollars of
tax reported on returns as tabulated from IRS’s Statistics of Income (SOI) samples. I use the same
source, but since these samples are not designed to be accurate at the state level, I have adjusted the
sample weights (by state and tax form type) so that they conform to actual return filings by state.
Second, I have restricted the number of returns filed and the dollars reported on those returns to
include only those returns that were required to be filed—excluding those with no tax liability, but
were filed to claim a refund of withheld tax or to claim the Earned Income Tax Credit.

The Determinants of Individual Income Tax Compliance

5

Finally, I do not estimate anything like “Assessed Liability per Return” (ALR) because it is
too misleading. Not only does it suffer from all of the disadvantages of reported tax as a
dependent variable, but the proposed audit adjustments component of ALR is an incorrect measure
of the direct revenue effect of audits. This is because the proposed audit adjustments
(“recommended additional tax and penalties”), which DGW compiled from IRS’s Commissioner’s
Annual Reports, are not the amounts assessed through enforcement; a large portion of these
“recommended” adjustments is never assessed, due to successful taxpayer appeals and litigation,
and the rate at which these recommendations are ultimately assessed has varied greatly over time
and across states. Moreover, even the proposed adjustments are endogenous with audit rates;
since IRS allocates its audit resources so as to audit only those returns it perceives to be most
noncompliant, the average audit “yield” declines with audit rate. Even though DGW recognizes
that ALR is endogenous with audit rates, and that as voluntary compliance increases (say, due to
increased audit rates) average audit yield is likely to decrease proportionately, the paper does not
recognize that audit yield decreases with audit rates by design. This complicates the choice of
instruments for audit rate, since an effective instrument now has to be unrelated to both taxpayer
compliance and overall resource levels. A final source of confusion introduced by the inclusion of
proposed audit adjustments with the tax voluntarily reported arises from the fact that they relate to
very different time frames, which could be important in a longitudinal study such as this.
Recommended audit adjustments are typically made more than a year after the return is filed; the
larger the adjustment, the greater the likelihood that the time lag is more than one year. Therefore,
the impact of variables whose effect on tax obligations varies over time may be difficult to estimate
reliably with this dependent variable.
1.3.2 Independent Variables
A more obvious improvement in this study is the inclusion of a much richer set of
explanatory variables—richer both in quantity and in quality. This allowed the specification of
each equation to be uniquely suited to the differences in the dependent variables; DGW employed
the same specification for each equation. I have grouped my variables into the following
categories: Tax Policy, Burden/Opportunity, IRS Enforcement, IRS Responsiveness, and
Demographics/Economics. I have constructed federal marginal tax rate variables in lieu of the
average state tax rate used in DGW, and have included a variety of other tax policy variables, as
well: the filing threshold, a state amnesty indicator, allowed child exemptions, and state and local
taxes that are deductible federally. I include the prevalence of sole proprietors rather than of
farmers, since they are a more prevalent (and, arguably, more important) indicator of the
opportunity to avoid (and, perhaps, to evade) taxes. Other Burden/Opportunity variables I have
included are the burden (in hours) needed to complete and file all required returns and schedules,
and the percentage of returns prepared by a paid practitioner.
The only IRS activity DGW includes is the audit rate. This study also includes the audit
rate, but it is the first to use the audit start rate instead of the audit closure rate. Since audits are
typically closed several years after the returns are filed, and closures in any given year relate to
many different prior tax years, the start rate better represents the percentage of returns filed in a
given year that are audited. I have also included variables for four additional enforcement
activities: the information return matching program, nonfiler notices, refund offsets, and criminal
tax convictions. This is also the first study of its kind to include variables related to IRS’s nonenforcement activities. Two included variables relate to IRS’s Taxpayer Service telephone
assistance and return preparation services. Variables considered, but not found to have a
significant impact on compliance include other Taxpayer Service activities (correspondence and
educational outreaches), and the speed with which refunds are processed and sent to taxpayers.

6

The Determinants of Individual Income Tax Compliance

1.3.3 Estimation Technique
DGW actually employs a more sophisticated econometric procedure, but I do not believe
that it is appropriate in this context. That study uses a “random effects” model to control for
unobservable state-specific factors that do not vary over time, whereas I employ a “fixed effects”
model to account for this common peculiarity of panel data. (I explain in section 2.2.1 why this is
a preferable approach in this context.)
1.3.4 Identification
One of the most critical features of any study of this kind is the choice of instrument(s) for
audit rates. DGW uses two: Budget per Return (BPR), and information documents (other than
W-2s) filed divided by the number of tax returns filed. Although these variables reflect a very
creative use of available data, each has serious drawbacks as an instrument for the audit rate. (The
reasons for this are given in section 2.2.2.) This paper introduces for the first time two
instruments that help to explain the audit rate, but which are unrelated to compliance: the
percentage of auditor time directly devoted to audits (Direct Examination Time, or DET), and the
average DET per audit. As productivity-related measures, these variables cause the audit rate to
increase or decrease quite independent of taxpayer compliance. Strong evidence that my reporting
equations are identified is given in section 3.2.1.
1.3.5 Functional Form
The DGW specification is strictly linear. In contrast, I use the logarithms of those
independent variables that are likely to have a non-linear effect on compliance. This is especially
important among the enforcement variables, which almost certainly achieve diminishing indirect
marginal returns to effort, much like their direct revenue effects.
2. The Model
It is tempting for economists to develop theoretical models of individual (micro) tax
compliance behavior. However, there are two significant reasons—one theoretical and one
practical—why these models may be inadequate. The theoretical reason is that much of that
behavior is governed by what is called in the literature “general deterrence.” 5 As Nagin (1978)
correctly observes, “general deterrence is inherently an aggregate phenomenon since it is reflected
in the behavior of the entire population.”6 That is why analyses of criminal sanctions have
generally been aggregate studies. Many studies of tax compliance have been aggregate, also, but
this has typically been due to a lack of access to micro-level compliance data, such as IRS develops
in its Taxpayer Compliance Measurement Program (TCMP). The aggregate studies, therefore,
have tended to start with micro-level economic models, and attempted to estimate these using the
lowest level of aggregation possible—such as at the 3-digit, or 5-digit ZIP Code level. (See, for
example, Beron, Tauchen, and Witte (1992) ). The lower the level of aggregation, however, the
less realistic is the model. That is because such models implicitly assume that the general deterrent
operates only within the strict confines of each unit of observation (e.g., a ZIP Code boundary),
5

We may think of general deterrence as including both negative influences, such as IRS enforcement actions, as well as
potentially positive influences, such as IRS responsiveness to taxpayers’ needs. Although the latter may not intuitively be
considered a deterrent, it undoubtedly influences the general population just like a deterrent. For example, many of those
who have good or bad experiences with IRS efforts to help them presumably share their experiences and perceptions with
their friends, who may change their own compliance accordingly. This is completely analogous to the way in which
perceptions about IRS enforcement are developed in the general population.
6

Nagin (1978), p. 99.

The Determinants of Individual Income Tax Compliance

7

and it seems obvious that people will develop their compliance perceptions and propensities based
on the information they get from a wide variety of sources from many locations. In fact, many
people today interact more with people from outside their ZIP Code (such as at work) than they do
with others in the immediate vicinity of their residence.
The practical reason why micro models may be inadequate is that it is virtually impossible
to quantify deterrence-type activities (like audit rates and Taxpayer Service phone calls) in any
meaningful way for each individual observation. Inevitably, these variables are aggregated in
some way, then imputed to individual observations; they are therefore subject to the same
limitations as the low-level aggregate studies, and are poor substitutes for the individual
perceptions called for in the theoretical models.
In order to avoid or minimize these problems, I have aggregated all data to the state level.
It would be nice to have been able to aggregate to the IRS district level,7 but it would have been
extremely difficult to have aggregated most of the non-IRS data in this way. Fortunately, the IRS
data could easily be combined to derive state-level aggregations. Such an aggregate analysis suits
the IRS from the perspective of its usefulness, as well; IRS is not as interested in individual
behavior as it is in aggregate behavior. Although the aggregate behavior is certainly the sum of the
behaviors of individuals, IRS is interested in the bottom line: how could we allocate resources
differently to improve voluntary compliance, and hence, net revenues? Aggregation bias cuts both
ways: it makes it hard to use aggregate results to estimate a micro model of individual behavior,
but by the same token, it is equally problematic to estimate a micro model using micro data, and
then to generalize the results to make aggregate calculations. It is best to make aggregate
calculations based on aggregate data. Although it may not have the sophistication of some micro
models, and it cannot model individual motivations, it seems appropriate in this context.
2.1 The Data
The data collected for this study form one of the most comprehensive datasets ever
compiled of potential determinants of voluntary filing and reporting compliance, and took over
eight years to assemble in usable form. Many of the IRS variables had never before been
assembled for any study, and were available only on paper or microfiche tables—often in a form
requiring some manipulation to derive the desired concept. Some IRS and external data were
available as representative samples of individuals; certain IRS variables were available at the district
level; and certain external data were available at the state level. All variables were aggregated to the
state level,8 and were compiled for a ten-year period: 1982-1991. The panel nature of the data
increased the number of observations, and also captured important variations in both compliance
and in its determinants over time. Appendix A contains a detailed summary of the sources and
derivations of the raw data used to create the variables included in this study.
2.1.1 Measures of Voluntary Compliance
The IRS recognizes three types of voluntary compliance: filing compliance (the timely
filing of any required return); reportingcompliance (the accurate reporting of income and of tax
liability); and payment compliance (the timely payment of all tax obligations). This study focuses
on both filing compliance and reporting compliance (both income reporting and offset reporting—
subtractions such as deductions, exemptions, adjustments, and credits). The most basic measure
7
8

Until just recently, IRS had 63 districts, each of which was a single state or a portion of a single state.

The District of Columbia (DC) is included in Maryland both because of its small size and because most IRS data are not
available for DC separately, since IRS’s Baltimore District includes DC with all of Maryland. Moreover, Alaska is excluded
from the data for reasons explained below.

8

The Determinants of Individual Income Tax Compliance

of voluntary compliance is what the taxpayers actually did: how many returns were filed, and how
much income and offsets they reported on those returns. However, while analyzing the possible
determinants of this taxpaying behavior, one clearly has to control for the corresponding “true”
obligations—returns required to be filed, income required to be reported, and offsets allowed to be
claimed. One could control for these true obligations either by including them among the
explanatory variables, or by dividing the basic voluntary measures by them to derive appropriate
compliance ratios. I have adopted the ratio alternative. Either way, these true obligations are never
observed; they can only be estimated.
It is tempting to use TCMP’s compliance data—thorough audits on a representative sample
of taxpayers—to estimate true obligations, but even these data suffer greatly from a general lack of
good information on the true filing and reporting obligations of the individuals in the samples.9
Moreover, TCMP data on individual income tax reporting compliance are currently available for
only four years (1979, 1982, 1985, and 1988), and only one survey on filing compliance is
available (1988), making it difficult to construct a useful panel of data from the TCMP. I therefore
controlled for true filing and reporting obligations using non-TCMP data.10
Filing Compliance
My measure of filing compliance [FilingRate] is the ratio of the number of required returns
actually filed to the total number required to be filed, expressed as a percentage. The number of
required returns actually filed was aggregated by state from IRS’s Statistics of Income (SOI)
samples of individual returns for each year.11 All three basic compliance measures (the number of
returns, and the amount of income and offsets reported on those returns) correspond to returns
required to be filed. This includes all returns having a positive tax liability or net losses. This
definition excludes returns filed “unnecessarily,” as well as those having no tax liability, but are
filed to claim a refund of any withholding, or solely to claim the refundable Earned Income Tax
Credit (EITC). I have excluded returns filed solely to claim the EITC since EITC noncompliance
tends to increase filing, whereas usual filing noncompliance decreases filing.
I have defined the denominator—the number of returns required to be filed—in the same
way, and have estimated it by state for each year from the micro data files of the Current
Population Survey (CPS) compiled by the Bureau of the Census.12 The CPS is the most
comprehensive annual U.S. census of individuals, and recognizes families and households, as
well. In order to structure the data to reflect potential tax returns instead of individuals, I first
combined the information about spouses into combined records, approximating a jointly-filed tax
9

Since IRS auditors are not omniscient, TCMP inevitably falls short of identifying all unreported income and the
corresponding tax liability. (Alexander and Feinstein (1987) attempted to estimate the undetected unreported income using
a sophisticated detection-controlled econometric technique, but while this would be useful for making aggregate
adjustments to estimates of noncompliance, using it to impute specific amounts of undetected unreported income to the
individuals in the sample is likely to make inferences about taxpayer behavior very sensitive to any errors introduced by
the imputation.) Estimating true filing obligations is even more troublesome at the micro level, since nonfilers, by
definition, tend not to be identified. (Erard and Ho (1995), for example, implemented a sophisticated analysis of the recent
nonfiler TCMP micro database to estimate both the number of nonfilers and the tax gap associated with them.)
10

It is somewhat ironic that I have chosen not to use TCMP data—even though I have access to them—while others have
used aggregate data only because they had no access to TCMP. This seems to be the right choice for this study, however.
11

State-by-state aggregations of filing data are available from IRS’s Individual Masterfile (IMF) of all returns, but
comprehensive reporting data are much too difficult to produce from the IMF—making it impossible to restrict our counts
to required returns. The SOI samples are the only alternative—both for filing data and for reported data. However, even
though the SOI samples of 100,000 or so individuals are relatively easy to access and manipulate, they are not designed to
be representative of state populations. Therefore, I adjusted the SOI sample weights for each year to make these files
conform to the actual number of returns filed in each state.
12

The “March Supplement” of the CPS is compiled each year, and typically has a sample size of 80,000 to 90,000
records. Each file reflects income received in the prior calendar year, so I compared the SOI estimates for a given tax year
with CPS estimates based on data compiled the following March.

The Determinants of Individual Income Tax Compliance

9

FilingRate (%) *

return.13 I then estimated which records corresponded to dependents in the tax sense, since
special filing rules apply to them.14 Next, I estimated which individuals might have been eligible
to claim the Head of Household filing status, making it a simple matter, then, to divide the
remaining records into the Single and Married-Filing-Jointly statuses. Finally, I calculated the
correct standard deduction and exemption values based on the estimated filing status and the tax
law for the year in question. Using the amounts of various types of income reported on the CPS, I
was then able to estimate whether the “return” had income over the filing threshold for that year.
Any record that did, or which showed negative income, I counted as a required return. Appendix
B contains detailed information about the logic of this estimation process, as well as a summary of
the CPS variables used and the relevant tax parameters for each year. Figure 1 illustrates the
resulting national trend in the FilingRate variable.

100
99
98
97
96
95
94
93
92
91
90

* FilingRate is the number of required
returns filed, as a percent of the total
number of returns required to be filed.

79

80

81

82

83

84

85

86

87

88

89

90

91

Tax Year

Figure 1. Filing Compliance Measure: The number of required individual income tax returns filed (per
SOI), as a percent of the total number of returns required (estimated from CPS), Tax Years 1979-1991

As Figure 1 illustrates, the national FilingRate has remained in the low 90’s, with a
general downward trend until 1988, and a significant improvement thereafter. (Average
FilingRates by state are tabulated in Appendix E.)
Reporting Compliance
Reporting noncompliance takes two forms: underreporting income and overstating
offsets to income (i.e., exemptions, adjustments, and deductions) or to tax (i.e., credits). Since
13

I had to assume that all couples filed jointly, since there is no conclusive information on the CPS to suggest otherwise.
This is a fairly small approximation, however, since very few couples file separately in reality.
14

Unfortunately, the CPS files compiled for me excluded records for children under the age of 15. This resulted in a slight
underestimation of filing obligations—and a corresponding overestimation of the FilingRate variable—for all states, but
this was especially fatal for Alaska. The state of Alaska pays from its oil holdings a “Permanent Fund Dividend” on the
order of $1,000 each year to every permanent state resident—including children—and this dividend is taxable federally.
Virtually every child in Alaska, therefore, has an obligation to file a federal income tax return—especially after the Tax
Reform Act of 1986, when the filing threshold for dependent children was drastically reduced in most cases. Because of
this, the number of Alaska returns filed doubled for tax year 1987. Since my CPS files could not reflect this, I chose to drop
Alaska from my analysis, leaving me with 49 states rather than 50.

10

The Determinants of Individual Income Tax Compliance

the various determinants of compliance presumably influence income reporting and offset reporting
differently, I have measured reporting compliance by means of three amounts that taxpayers
voluntarily report on filed returns: the total amount of income that they report, the amount of
offsets15 they report, and the corresponding net income—income minus offsets. It is tempting to
monitor the tax reported on the returns, but the impact of tax policy parameters (e.g., marginal tax
rates and filing thresholds) on the tax voluntarily reported is twofold—influencing the tax
calculation directly, and potentially influencing compliance propensities, as well. By focusing on
income and offsets separately—before the tax calculation—we can isolate the impact of tax policy
parameters on reporting compliance. I estimate the net income equation both because it is close to a
tax concept, and because it is a useful check on the results from the separate income and offsets
equations.
The amounts of income and offsets reported were aggregated by state from the SOI files for
each year. However, during the 1982-1991 time period, the rules that govern what income must
be reported and what amount of offsets may be claimed changed significantly for many items (see
Appendix C for complete details). Unless we can control for these rule changes by modifying the
data to reflect constant-law amounts, or by including adequate explanatory variables, we may
misinterpret the effects of tax policy variables (like marginal tax rates) that are highly correlated
with these other tax law changes.16 I handled this problem by defining the three reporting
compliance measures (income, offsets, and net income) in three separate ways: (A) excluding all
components whose reporting rules changed during the period (except for ones that could be
controlled for by creating constant-law data or by including appropriate explanatory variables; this
definition included on the order of 97 percent of total income, 30 to 60 percent of adjustments, 94
percent of itemized deductions, and 30 to 60 percent of credits); (B) making all of those
adjustments, but including income and offset components whose rules were changed only by the
Tax Reform Act of 1986 (TRA86); and (C) including all income and offset items regardless of rule
changes. Comparing the results using these three sets of definitions gives us insight into the
impact of changes in the rules governing what must (or can) be reported.
As with filing compliance, each of the reporting concepts was expressed as a ratio.
However, unlike for filing compliance, it is not possible to develop comprehensive exogenous
estimates of the amount of income or offsets that should be reported. Therefore, to control for
true reporting obligations, each of these reported amounts was divided by the amount of Personal
Income estimated for the National Accounts by the Bureau of Economic Analysis (BEA) for the
appropriate state and year. Although Personal Income does not correspond exactly to the amount
of “true” income that must be reported on federal income tax returns (it includes the income of all
those who do not need to file returns, for example; see Appendix D for a detailed comparison), and
it is certainly not the same as “true” offsets allowable, it is a very effective control for these
concepts. This is because it is probably the most comprehensive individual income variable
available annually at the state level, and because it is derived substantially independent of tax return
data. That is, it is reasonably exogenous to income tax compliance and income tax administration
decisions. Table 1 summarizes the data sources for each of the major components of Personal
Income. As the table shows, four major components are based (at least in part) on individual
income tax returns. However, BEA adjusts these amounts to account for underreporting using
data from the IRS Taxpayer Compliance Measurement Program, the Information Returns Program,
audits, and other data (see Parker (1984)).

15

All of these offsets are subtractions from income, except credits, which are subtractions from tax. In order to combine
these amounts into a single offsets concept, I converted the credit amounts on a given return to the equivalent income offset
amounts by dividing them by the marginal tax rate faced by the return.
16

I am indebted to Brian Erard for making this observation.

The Determinants of Individual Income Tax Compliance

11

Table 1. Data Sources for Major Components of Personal Income
Component of Personal Income
Wages and salaries
Other labor income
Non-farm proprietor income
Farm proprietor income
Rental income of persons
Royalty income of persons
Personal dividend income
Personal interest income
Transfer payments

Data Source
State Unemployment Insurance data (Form ES 202)*
Forms 5500 submitted by employers and plan managers
Forms 1040*
U.S. Department of Agriculture
Census of Housing, Bureau of the Census
Forms 1040*
Forms 1120 submitted by payers of dividends
Forms 1120 submitted by payers of interest*
Federal budget, Social Security Administration, etc.

Source: Thae Park, Bureau of Economic Analysis, Department of Commerce; see also BEA (1994).
* These data are adjusted by TCMP, IRP, audit , and other data; see Parker (1984).

The resulting ratios [IncomePct, OffsetsPct, and NetIncomePct] must not be interpreted
as compliance rates in the same sense that the FilingRate ratios can be, but in the context of the
regression analysis described below, dividing by Personal Income (and including additional
control variables among the explanatory variables) is an effective way to control for variations in
true obligations over time and across states.
The national trends in the income and offset reporting variables are illustrated in Figures 2
and 3, respectively. (Average values by state are tabulated in Appendix E.) One of the most
striking things about Figure 2 is to see that reported income as a percent of Personal Income
generally declined through 1986, and then rose sharply. A decomposition of total reported income
into the separate types of income reveals that most of the increase after 1986 (presumably in
response to TRA86), can be attributed to Schedule E income (consisting mainly of rents, royalties,
and income from partnerships, S-Corporations, estates and trusts), and to wage and salary income.
Of the Schedule E income types, S-Corporation income is not included in Personal income, and yet
increased after TRA86, so this may have contributed to the overall rise in IncomePct. However,
excluding all Schedule E income from IncomePct does not change the results appreciably;
IncomePct still rises after 1986, and the econometric parameter estimates are similar. The reason
for the rise in wage and salary reporting is less clear. The biggest difference in the three definitions
of income reported is that the unadjusted measure of total income reported declines after 1988.
This is due primarily to the decline in capital gains reported during that period. Given that capital
gains are not included in Personal Income, and that the timing of gains realizations is often
influenced by tax law changes, I excluded it from income definitions A and B.
The trend of offsets under definition A follows a pattern quite similar to that of income
reported. The main reason for the sharp rise after TRA86, however, is the large increases in
standard deductions and exemptions. Although these increases imply larger filing thresholds
(resulting in fewer required returns), their net effect is to increase offsets reported as a percent of
Personal Income. I control for the increase in standard deductions and exemptions by including
several explanatory variables. The large differences between the amount of offsets claimed under

The Determinants of Individual Income Tax Compliance

Income Reported as a Percent of Personal Income

12

75
74
73
72
71
70
69
68
67
66
65
79

80

81

82

83

84

85

86

87

88

89

90

91

Tax Year
A. Excluding capital gains and all income components with rule changes in 1983-91
B. Excluding capital gains and all income components with non-TRA86 rule changes
C. Total Income (unadjusted)

Figure 2. Income Reporting Compliance Measures: The amount of income reported on returns (per
SOI), as a percent of the amount of Personal Income (per BEA), for the three definitions of income,
Tax Years 1979-1991

definitions B and C compared with definition A reflect the large number of offset items excluded
from definition A because the rules for what can and cannot be claimed on these items changed
during the period—especially due to TRA86.
2.1.2 Potential Determinants of Voluntary Compliance
I have included over twenty variables likely to influence the four measures of voluntary
compliance described above—one filing equation and three reporting equations (income, offsets,
and net income). Apart from FilingRate, which I include in the reporting equations (since the
more returns that are filed, the more income and offsets will be reported overall), these explanatory
variables fall into five categories: Tax Policy, Burden/Opportunity, IRS Enforcement, IRS
Responsiveness, and Demographics/Economics. These variables are defined in Table 2, and are
described in greater detail below—including a discussion of my a priori judgment of the direction
(sign) of each one’s impact on compliance. Descriptive statistics (and units) for all the variables
are provided in Appendix E.

Offsets Reported as a Percent of Personal Income

The Determinants of Individual Income Tax Compliance

13

30
29
28
27
26
25
24
23
22
21
20
79

80

81

82

83

84

85

86

87

88

89

90

91

Tax Year
A. Excluding all offset components with rule changes in 1983-91 *
B. Excluding all offset components with non-TRA86 rule changes *
C. Total Offsets (unadjusted)
* Except standard deductions, exemptions, and state & local taxes

Figure 3. Offsets Reporting Compliance Measures: The amount of offsets reported on returns (per SOI),
as a percent of the amount of Personal Income (per BEA), for the three definitions of offsets, Tax Years
1979-1991

Tax Policy
I control for changes in the values of standard deductions and exemptions through the use
of two variables: FThresholdPct, which is the aggregate filing threshold value17 among all
required returns (estimated from the CPS), expressed as a percentage of Personal Income; and
ChildExemptsPct, which is the aggregate exemption value for all children among all required
returns (estimated from the CPS), also expressed as a percentage of Personal Income. Increased
filing thresholds will decrease the number of returns filed, but it is not clear whether increased
thresholds would affect the FilingRate one way or the other; it may depend on whether those who
no longer need to file have a higher filing rate than those who are still required to file. If filing
thresholds and dependent exemptions have any impact on how much income is reported
(controlling for the filing rate), that impact is likely to be positive, since the larger these offsets are,
the more income that can safely be reported without paying more tax. FThresholdPct and
ChildExemptsPct should both have a positive impact on OffsetsPct, since these are two of the
most important income offsets.
17

The filing threshold for a given return is the amount of total income below which the taxpayer generally does not need
to file a return. In most cases, this is defined as the sum of the standard deduction applicable to the taxpayer plus the value
of personal exemptions to which he is entitled. Personal exemptions include those for the taxpayer and spouse (if filing
jointly), and exclude exemptions claimed for dependents; claiming dependent exemptions requires filing a return.

14

The Determinants of Individual Income Tax Compliance

Table 2. Definitions of Explanatory Variables
Variable

Numerator

Tax Policy
FThresholdPct
Amnesty5
MargTaxRate@$15K
MargTaxRate@$57K
ChildExemptsPct
StateTaxPct
Burden / Opportunity
AvgBurden
SoleProps
SolePropTFS

Filing threshold on required returns (CPS)
Dummy indicating whether state has had
an amnesty in the last 5 years
Marginal tax rate at $15K taxable income
(weighted for married-single mix by state and year)
Marginal tax rate at $57K taxable income
(weighted for married-single mix by state and year)
Value of exemptions for children (CPS)
State income, property & sales tax revenues

Personal Income (BEA)

Total tax form burden on required returns (CPS)
Number of sole proprietors (CPS)
SoleProps x percentage of non-farm employment
in Trade, Finance & Service sectors
No. of returns prepared by paid practitioner (SOI)

Number of potential returns (CPS)
Number of potential returns (CPS)

PaidPrep
IRS Enforcement
AuditRate
Number of district audits started in fiscal year (AIMS)
IRP_DocRate
No. of IRP documents matched against returns
TDI_TotRate
Total number of TDI notices issued
RefOffRate
Number of refunds offset for outstanding debts
CID_ConvRate
Criminal convictions
IRS Responsiveness
TPS_CallsPC
Number of telephone calls handled by TPS
TPS_RetPrepPC
Number of returns prepared by TPS
Demographics / Economics
Singles
Number of singles among potential returns (CPS)
Under30
Number of potential returns under age 30 (CPS)
Over64
Number of potential returns over age 64 (CPS)
PCBirths
Number of births (HHS)
AvgPI
Personal Income (BEA)
AvgPIgrowth
Annual growth in AvgPI
ExclIncomePct
Income on potential returns that is not taxable
UnemplRate
Unemployment rate (among those 16 and older)
Abbreviations:

AIMS
BEA
CID
CPS
HHS
IRP
SOI
TDI
TPS

Denominator

Personal Income (BEA)
Personal Income (BEA)

Number of returns filed (SOI)
Returns filed in prior tax year (SOI)
Number of potential returns (CPS)
Number of potential returns (CPS)
Number of refunds
Population, in millions (Census)
Population, in thousands (Census)
Population, in thousands (Census)
Number of potential returns (CPS)
Number of potential returns (CPS)
Number of potential returns (CPS)
Population, in thousands (Census)
Number of potential returns (CPS)
Personal Income (BEA)

Audit Information Management System (IRS Examination function)
Bureau of Economic Analysis, national accounts (Commerce Department)
Criminal Investigation Division (IRS)
Current Population Survey (Census Bureau)
U.S. Department of Health & Human Services
Information Returns Program, document matching (IRS)
Statistics of Income (IRS)
Taxpayer Delinquency Investigation, nonfiler program (IRS Collection function)
Taxpayer Service function (IRS)

The Determinants of Individual Income Tax Compliance

15

60
55
Taxable Income of $57,000 (in 1982 $)
50

Marginal Tax Rate (%)

45
40
35
30

Taxable Income of $15,000 (in 1982 $)

25
20
15
10
5
0
79

80

81

82

83

84

85

86

87

88

89

90

91

Tax Year
Single

Married

Weighted Average

Figure 4. Marginal Tax Rates: The national trend for two taxable income levels, Singles and Marrieds Filing
Jointly, and the weighted average of the two, Tax Years 1979-1991

Marginal tax rates are difficult to reflect in an aggregate analysis—especially since the tax
rate schedules apply to everyone equally in a given year. Also, since the average marginal tax rate
for each state population does not by itself reflect the progressivity of the tax rate schedules, or the
possibility that marginal rates have different impacts on the reporting compliance of taxpayers at
different income levels, and since any average marginal rate calculated from filed returns is
endogenous, I developed an alternative way to estimate the impact of marginal rates. This involved
determining the marginal tax rate for each year from the tax rate schedules (included in the Form
1040 tax package) at two separate levels of taxable income: $15,000 and $57,000, expressed in
constant 1982 dollars.18 These marginal rates saw significant changes over time (see Figure 4),
and varied across states each year due to the widely variable mix of single and married taxpayers
from state to state.19 I constructed a composite marginal tax rate variable for each income level by
weighting the single and married rates by the relative mix of singles and marrieds among potential
18

$15,000 was chosen as a fairly modest income, and $57,000 was chosen because it is the highest level of taxable
income (in 1982 dollars) for which the Single and Married-Joint marginal tax rates were different for each year. If the
marginal tax rate were the same for these two rate schedules in any year, then we would not observe any variation in the
variable across states for that year.
19

The ratio of marrieds to singles in a given state typically ranged from 0.6 to 1.2.

16

The Determinants of Individual Income Tax Compliance

returns (which I estimated from the CPS). (I also included the percent of singles in the state
population among the demographic explanatory variables to ensure that these marginal tax rate
variables are not simply surrogates for the married-single mix.) As Figure 4 shows, the marginal
tax rate at $15,000 of taxable income remained relatively stable, whereas the rate at $57,000 of
taxable income dropped significantly over the 1982-1991 period of my analysis. It is not
completely clear a priori whether the effect of marginal tax rates is positive or negative; it most
likely depends on how risk-averse the typical taxpayer is relative to the magnitude of other
important tax parameters. Since low-income taxpayers are most likely more risk averse than higher
income taxpayers, however, marginal rates are likely to have more of a positive effect on those
with low income.
Two additional tax policy variables relate to state taxes. StateTaxPct is the amount of state
income, property, and sales tax revenues that were deductible federally, expressed as a percentage
of Personal Income. We should expect this to have a positive impact on OffsetsPct. Amnesty5
is a dummy variable that indicates whether the state has had a tax amnesty20 within the five years
up to and including the year in question. Since state and federal tax administration is linked in
many important ways, it is conceivable that the federal government realizes some of the short-term
gains that the states have enjoyed following their amnesties. However, it is also possible that
taxpayers reduce their compliance following an amnesty—either because they expect another one
will follow, or because they feel they deserve to receive some of the government’s leniency from
which their less compliant neighbors have benefitted. Therefore, it is not clear a priori whether
state amnesties have a positive or a negative impact on federal compliance.
Burden / Opportunity
It seems reasonable to presume that the more complex the tax system becomes, and the
harder it becomes to comply with one’s tax obligations, the more likely people will become
noncompliant—either unintentionally due to confusion, or willfully out of frustration. On the other
hand, the fewer opportunities there are to be noncompliant (e.g., with the introduction of
requirements that payers of certain types of income report this information both to the recipient and
to the IRS), the less noncompliance we should expect. These two factors are actually much the
same: the complexity of the tax system (and therefore the burden associated with complying with
it) arises to a large extent from the various opportunities left open for noncompliance (e.g.,
underreporting business income), and the many mechanisms in the tax system to minimize those
opportunities (e.g., detailed forms and schedules, complicated rules, and lengthy instructions).
Moreover, what is complexity to one (e.g., itemized deductions) may be opportunity to another.
Recognizing that business income among individuals presents both tax-paying complexities
and opportunities, I have included two variables: SoleProps, the percent of potential returns
having non-farm sole proprietor income (per the CPS); and SolePropTFS, an interaction term
between SoleProps and the percent of nonfarm employment in the Trade, Finance, and Services
sectors. Sole proprietors generally keep a fairly visible public profile, and thus may find it difficult
to hide from the IRS’s notice. However, dealing largely in cash and “moonlighting” may provide
many entrepreneurs—especially in the Trade, Finance, and Services sectors—the opportunities to
evade taxes by not filing or by not reporting all of their income. It is not clear whether SoleProps
should have a positive or negative impact on compliance, but we should expect SolePropTFS to
have a less positive, or even negative, impact compared with SoleProps.
20

By 1991, 33 states had conducted some form of amnesty. Virtually all of these waived some or all penalties associated
with nonfiling if a delinquent return were filed within a specific time period. Many amnesties also focused on accounts
receivable.

The Determinants of Individual Income Tax Compliance

17

The extent to which taxpayers pay others to prepare their returns has been related to
compliance before, and it seems appropriate to control for it. It is not obvious, however, whether
paid tax practitioners improve reporting compliance (say, by asking their clients if they had any
extra income from easy-to-forget sources, or by avoiding misunderstanding as to what offsets are
available to their clients, and how to calculate them), or whether they reduce the amount of net
income reported (say, by pointing out legal tax avoidance strategies open to their clients).
A variable that has not been related to compliance before is the burden associated with
getting, learning how to use, completing, and filing the various required tax forms and schedules.
This “burden” was estimated by an IRS study conducted in response to the Paperwork Reduction
Act. Estimates of the burden (measured in hours) associated with each form and schedule have
been published in the Form 1040 tax package in the most recent years, and I have applied the same
methodology to develop estimates for the earlier years. (These estimates are tabulated in
Appendix F.) I then multiplied these burden estimates for each form and schedule by the number
of corresponding forms and schedules required to be filed, which I estimated from the CPS (see
Appendix F for details on the logic employed). I then derived AvgBurden by dividing this total
burden estimate for the state and year by the corresponding number of potential returns indicated
by the CPS. (I use potential returns instead of filed or required returns because an increase in
the filing threshold could easily increase the average burden among the remaining required returns,
but should be considered a burden reducer.) We should expect this burden to decrease filing
compliance, as well as the reporting of offsets, since the hassle associated with maintaining the
correct records and completing the paperwork can only diminish taxpayers’ willingness to file
returns and to claim offsets to which they might be entitled. It is not clear a priori what impact
burden might have on income reporting compliance, though the “hassle factor” could have a
negative effect.
IRS Enforcement
The primary enforcement variable thought to influence voluntary compliance is the audit
coverage rate—the percentage of returns audited. Presumably, if taxpayers respond to the
deterrent effect of audits, they (subjectively, at least) try to estimate their chances of being audited.
This probability depends both on what they report on their returns (making the audit rate
endogenous with compliance) and on the prevailing level of audit resources in their area.
Traditionally (in IRS reports, and therefore in academic research), the latter concept has been
expressed as an average audit coverage rate, which is usually defined as the number of audits
closed in a given fiscal year divided by the corresponding number of returns filed in the prior
calendar year. Since the length of audits varies widely, and is often longer than one year, the
traditional coverage rate concept does not accurately reflect the average percentage of returns filed
in a given year that are eventually audited. That is better captured by the percentage of audits
started in a given year, which has never been used in an analysis of taxpayer compliance. It is
possible, however, that audits send different signals to the general population when they are started
compared with when they are closed. For example, the message that gets “rippled” to friends and
neighbors when an audit begins presumably focuses on the fact of the audit, and may shape their
perception of their own likelihood of getting audited. In contrast, the message communicated
when an audit ends probably has more to do with the quality of the audit, and may shape others’
perceptions more of the consequences of the audit (good or bad) than of its likelihood. Including
both audit measures—which are obviously highly correlated—however, introduces the problem
of multicollinearity into the analysis, so I have included just the audit start rate alone, since it
displayed the greater predictive power.21
21

My definition includes only the person-to-person audits conducted by the district offices; it excludes the simple
correspondence audits conducted by the service centers.

18

The Determinants of Individual Income Tax Compliance

Many have observed that over the last two decades voluntary compliance seems to have
fallen concurrent with a decline in the audit coverage rate. However, as shown in Figure 5, much
of the decline in conventional, labor-intensive audits has been accompanied by a very significant
rise in IRS’s ability to detect noncompliance through the use of automated matching of third-party
information documents with tax returns in its Information Returns Program (IRP). We ought to
control for this shifting of enforcement resources, since taxpayers may not perceive much of a
difference between getting caught by a person and getting caught by a computer. For this reason,
some recent analysis22 has included the number of computer-generated notices to taxpayers (called
CP-2000’s) arising from such mismatches. I do not think that that is the correct variable,
however. The deterrent effect of the IRP document matching program is achieved when taxpayers
believe that virtually every mismatch will be detected and pursued. It is not the number of
mismatches found that reflects the level of enforcement, but rather the number of documents
actuallymatched. In fact, in recent years, the reporting of wage, salary, interest and dividend
income has steadily improved while the number of mismatches has generally declined.
Compliance has improved in these areas because taxpayers have increasingly understood that
virtually all mismatches will be detected (a “coverage rate” approaching 100 percent); the number
of mismatches has fallen as a result. Therefore, I have included the average number of IRP
documents processed per potential return [IRP_DocRate] as an explanatory variable. (As with
AvgBurden, this is divided by the number of potential returns; since IRP documents are
submitted even for those who have no filing obligation, we should not compare the number of IRP
documents with only the required returns.) Since one of the uses of these documents is to identify
nonfilers, the more IRP documents that are processed, the greater the likelihood that a potential
nonfiler will choose to file. Although the overall impact of IRP has been to improve the reporting
of income, as well, it is not clear what impact new types of IRP documents have had in the recent
past. Their positive deterrent may be mitigated somewhat by a “What the IRS doesn’t know won’t
hurt them” type of mentality. That is, taxpayers could improve their reporting of IRP-covered
income types, but reason that income not reported to IRS is easy to conceal. The more that
taxpayers are aware of the limits of IRS knowledge, the more opportunity they have to underreport
their income.
I am not aware of any other study that has attempted to measure the impact of three other
IRS enforcement programs. The first, the Taxpayer Delinquency Investigations (TDI) program, is
specifically targeted toward nonfilers. Based on either the presence of sufficient IRP-detected
income without a return having been filed (known as the “IRP-Nonfiler” program), or on the
absence of a return from someone who had filed the previous year (known as the “Stopfiler”
program), IRS issues up to four TDI notices to potential nonfilers. If the notices do not yield the
required returns, a more-intensive investigation may be conducted by the IRS Collection function.
TDI_TotRate includes all such notices issued by state and year, and we should expect that this has
a positive effect on the FilingRate.
Most of the effort of IRS’s Criminal Investigation Division (CID) is intended to improve
voluntary compliance by catching and prosecuting tax fraud cases. This can improve compliance
in two ways: either as a deterrent among those tempted to defraud the government, or as an
encouragement to the general population (to the extent that they don’t want to see criminals go
scott-free). It is possible, however, that the latter effect primarily works in reverse; that is,
criminal convictions might not improve compliance, but the lack of them might erode compliance.
Since one of the primary mechanisms for influencing the general population by CID activities is the
publicity surrounding the cases—especially if convictions result—I have included
CID_ConvRate, which I have defined as the number of criminal convictions obtained per million
people in the population.
22

For example, an early draft of Dubin, Graetz and Wilde (1990).

The Determinants of Individual Income Tax Compliance

19

8
7
IRP_DocRate
(information documents
processed per potential return)

6
5
4
3

AuditRate
(district audits started per
100 returns filed in prior year)

2
1
0
80

81

82

83

84

85

86

87

88

89

90

91

Year

Figure 5. IRP and Audit Coverage Trends: The number of IRP documents processed per potential return,
and the audit start rate, 1980-1991

The third IRS enforcement activity unique to this study involves not paying some or all of a
refund to a taxpayer in order to satisfy delinquent child support payments, or some other debt. I
have included RefOffRate, which is the percentage of refunds offset in this way. It seems likely
that if a taxpayer has such a debt, he is likely to find ways to avoid being offset in the future (such
as by adjusting his withholding, not filing altogether, or both) if he has been offset once, or if he
hears of others getting smaller refunds than they had claimed.
IRS Responsiveness
One of the most significant contributions of this study is a better understanding of the
extent to which IRS responsiveness to taxpayer needs influences voluntary compliance. To my
knowledge, this has not been studied before. Unfortunately, reliable and consistent measures of
such responsiveness have generally not been maintained by IRS for very long. In most cases,
only one or two years worth of data are available. However, I have compiled data for two
important types of Taxpayer Service (TPS) activities: the number of telephone calls handled per
thousand people in the population [TPS_CallsPC], and the number of returns TPS helps to
prepare, also per thousand people in the population [TPS_RetPrepPC]. We should expect
TPS_RetPrepPC to contribute positively to both filing and reporting compliance. The impact of
the telephone calls is somewhat ambiguous, however. Since these calls are almost always initiated
by taxpayers, the ones who choose to call are are probably not representative of the overall
population. Taxpayers call for two major reasons: seeking information about the administrative
progress or status of their account, or seeking clarity on some substantive tax law issue that they
face. Generally, a pleasant experience with the IRS (e.g., getting the correct answers in a
reasonable amount of time) ought to contribute to higher voluntary compliance, but since the

20

The Determinants of Individual Income Tax Compliance

quantity of phone calls is not always related to their quality, the likely impact of the calls is not
straightforward. There are two additional factors that add uncertainty: (1) the people who ask the
IRS substantive tax law questions are likely to call because they are generally compliant, and will
tend to seek the correct answer rather than not call and take a risk, and the answers to their
questions might either be in their favor or not; and (2) the number of calls answered does not
always relate to the number attempted, but not answered, or to the number of times a taxpayer had
to try before being served. Including both of these TPS variables, however, could shed important
light on the impact that these activities have.
Demographics and Economics
It is probable that a variety of demographic and economic factors help to shape an
individual’s tax compliance behavior. Other factors help to control for fluctuations in the
dependent variables (specifically Personal Income in the denominator of the reporting compliance
measures) that are not related to compliance.
The demographic variables include the prevalence of singles as percent of potential returns
[Singles], the percentage of potential returns in young and old age categories [Under30 and
Over64], and the number of births per thousand in the population [PCBirths]. Marital status has
important tax implications, and is likely to be negatively correlated with compliance. The two age
categories included tend to have less income, but it is not obvious whether they contribute to higher
filing compliance. The per capita birthrate may be related to people’s present satisfaction level and
to their optimism about the future. If so, it may be related to better compliance.
The economic variables include: the level and rate of growth of Personal Income [AvgPI
and AvgPIgrowth, respectively]; a measure of the income included in Personal Income, but
excluded from taxable income [ExclIncomePct]; and the unemployment rate among those age 16
and older [UnemplRate]. AvgPI and AvgPIgrowth control for income differences across states,
which must be done through the inclusion of independent variables in the FilingRate equation.
(Recall that the three reporting compliance measures are divided by Personal Income, so including
Personal Income on the right-hand side of these equations would introduce non-linearity with
respect to income, which may or may not be realistic.) ExclIncomePct controls for fluctuations in
Personal Income (and therefore in the income reporting compliance measures) related to sources of
income not reported on income tax returns—such as the income of those not required to file,
veterans’ benefits, and child support payments—and should therefore have a negative sign.
Finally, a high unemployment rate is likely to cause taxpayers to become less compliant; to the
extent that they have less disposable income, they are more likely to cut corners on their taxes—or
maybe not even file at all.
2.2 Estimation Approach
Three important complexities had to be addressed in order to estimate successfully the
impact of these various potential determinants on voluntary compliance. The first arose from the
panel structure of the data (a time series of cross sections). The second dealt with the endogeneity
of audit rates. The third involved the problem of accounting for changes in the rules as to what
income should be reported and what offsets could be claimed.

The Determinants of Individual Income Tax Compliance

21

2.2.1 Panel Data
In the standard regression model, the unexplained error term is assumed to vary randomly
across all observations. With a cross-sectional and time-series panel structure, however, certain
left out (perhaps unmeasurable) variables may vary cross-sectionally (i.e., across states in this
study), but not over time (such as cultural factors), while other variables may vary over time, but
not across states (such as features of the tax law). The effects of these left-out variables are
captured in the error term, which is no longer random across all observations. The accepted
econometric approach in such cases is to assume that the error term is made up of components:
one varying only across individuals (i.e., states), reflecting the “individual effect;” sometimes
another component that varies over time, but not across individuals, reflecting a “time effect;” and
then the usual error term, varying randomly across all observations. Pioneering work applying
this technique to economic problems include Kuh (1959) in estimating investment equations,
Mundlak (1961) and Hoch (1962) estimating production functions, and Balestra and Nerlove
(1966) on demand functions.
The existence of such individual and time effects can be determined from a simple analysis
of variance examination of the residuals obtained from the standard regression that assumes no
error components. Such a test on my data reveals a very significant presence of both individual
(i.e., state) and time effects.
There are two principal ways to model these error components. One is to view them as
“fixed effects” for each individual (state) and time period (year). Under this assumption, the
effects are not only unique to each state or time period, but they are constant, or “fixed.” If a
separate regression of the same equation were estimated for each state on its own, all of the
estimated parameters would be identical for each state, except for the constant term; the “individual
effect” would appear in combination with the overall constant term, making the constant vary
across states. The other possibility is that the individual effect is random rather than constant.
Under the “random effects” assumption, all components of the error term (the individual effect, the
time effect, and the remaining disturbance) are distributed with a mean of zero and a constant
variance. However, the variance of the individual effect is unique to each individual (state), and
the variance of the time effect is unique to each time period (year), while the variance of the
remaining disturbance is common across all individuals and time periods.
Fixed Effects vs. Random Effects
Which is the best way to model the error term? Two considerations have led me to choose
the fixed effects model. First, the random effects approach is best suited when the panel consists
of individuals tracked over time who are representative of a much larger population. But if the
individuals are the population of interest, then the individual effects would most appropriately be
fixed.23 In this problem, the “individuals” are states, and are the population of interest; they are
not a sample drawn from a larger population, so the fixed effects model would seem the most
appropriate. Second, the random effects model may lead to biased parameter estimates if the
individual effects include unobserved factors that are correlated with the included explanatory
variables.24 The usual example of such an unobserved factor is an individual’s inherent “ability”
(or intelligence) which is undoubtedly correlated with education (say, in an equation to explain
income differences). In this study, the unobservable state effects undoubtedly include factors
having to do with people’s attitudes toward and perceptions about the federal government
generally, and tax compliance specifically. If so, then these effects are almost by definition
23

See Hsiao (1986), pp. 42-43, and Baltagi (1995), p. 13.

24

See Mundlak (1978), and Hsiao (1986), pp. 43-46.

22

The Determinants of Individual Income Tax Compliance

correlated with the included explanatory variables, since these variables are included for the very
reason that they are thought to influence taxpayer attitudes and perceptions. For example, if IRS is
not responsive to taxpayer needs (e.g., by answering telephone enquiries correctly and efficiently),
then those taxpayers—and perhaps others influenced by them—may develop unobserved attitudes
toward taxpaying (built on frustration, or perceived unfairness, for example) that cause them to be
less compliant next time around. So, by both considerations, the fixed effects model would seem
to be the most appropriate specification.
Estimation Procedure
The most straightforward way to estimate a fixed effects model is to include an overall
constant term as well as dummy variables for all but one of the states and for all but one of the
years. This captures the fixed effects explicitly, leaving only one error component: the standard
random disturbance. I employed this procedure, which is sometimes called a Least SquaresDummy Variable (LSDV) approach (or 2SLSDV for Two-Stage Least Squares), since it gives
insight into the magnitude and distribution of the effects.
2.2.2 Endogeneity of Audit Rates
By carefully constructing my dependent and explanatory variables (for example, using the
CPS to estimate filing obligations and certain economic and demographic characteristics of the
potential filing population, and using marginal tax rates taken directly from the tax rate schedules),
I have avoided a number of potential endogeneity problems. However, the problem of
endogenous audit rates remains, and is one of the most significant estimating challenges in any
comprehensive study of IRS’s impact on voluntary compliance. The problem arises because not
only do audit rates presumably influence taxpayers’ perceptions of their chances of getting
audited—and, therefore, their compliance decisions—but IRS allocates its audit resources based on
its perceptions of taxpayers’ noncompliance. That is, audit rates and voluntary compliance are
jointly (simultaneously) determined. Since these audit resources are allocated on a district-bydistrict basis (where most districts encompass an entire state), this endogeneity can be expected to
manifest itself in state-level data. All of the most recent attempts to estimate the determinants of
voluntary compliance have included audit rates, and have recognized this endogeneity. However, I
am not confident that these earlier studies have successfully identified the compliance equations in
which audit rates appear. Dubin, Graetz, and Wilde (1990) employs an instrumental variables
approach in which IRS budget per return filed and the the number of information returns per tax
return filed are used as instruments for audit rate. Since district budgets vary in large part due to
IRS perceptions of the variations in taxpayer compliance, since information documents have their
own impact on voluntary compliance (particularly on filing compliance), since the number of
returns filed (the denominator for each instrument) is itself endogenous, and since data on
information documents were not compiled by state, these would seem to be poor instruments for
the audit rate. Beron, Tauchen, and Witte (1992) uses a similar approach, including as an
instrument the number of returns filed per IRS employee in a given district—in essence, the
inverse of the budget per return instrument described above. This is intended to reflect the fact that
district audit staffing levels are often dictated by constraints unrelated to compliance, but this
variable includes all types of district staffing, and it is unclear whether the constraints on optimal
resource allocation are significant enough to say that this variable is an effective instrument for
audit rates. Since the paper does not include, for the sake of comparison, results using the
endogenous audit rate in the reporting equations, we can only speculate whether the predicted audit
rates based on this instrument corrected the bias of the parameter estimate, or exacerbated it.

The Determinants of Individual Income Tax Compliance

23

Direction of Bias
To understand the presence and direction of this bias, consider the following generic
income reporting and audit rate equations:
IncomePct = α0 + α1AuditRate + ε
AuditRate = β0 – β1IncomePct + ν
We would expect the impact of AuditRate on IncomePct (i.e., α1) to be positive, and the
impact of IncomePct on AuditRate (β1) to be negative. If the endogeneity is not controlled for,
what is the relationship between the desired parameter estimate (α1) and the error term? A positive
ε implies a larger value of IncomePct. The second equation, then, because the coefficient on
IncomePct is negative, suggests that the larger value of IncomePct is associated with a lower
value of AuditRate. So, a positive ε is associated with a smaller AuditRate. This means that the
coefficient on AuditRate (α1) would be biased downward if we estimated the equation using the
endogenous AuditRate directly, with no correction for the endogeneity. This negative bias can be
seen in state level data; many states with high audit coverage also have low compliance. This does
not mean that audits reduce compliance; rather, it simply reflects that IRS intentionally allocates its
audit resources where they are needed most. Using similar logic, one would expect a positive
bias of the AuditRate parameter in the OffsetsPct equation—on the assumption that OffsetsPct
has a positive impact on AuditRate. I examine these biases, and the extent to which they are
corrected, in the discussion of results, section 3.2.1.
Instruments for AuditRate
The ideal instrument for AuditRate would be something that helps to predict AuditRate,
but is not related to compliance. I have chosen two audit productivity-related measures to fulfill
this purpose. Presumably, the more productive the auditors are in a given year, the more audits
they can start in that year. And, to a lesser extent, perhaps, the more time applied per audit last
year, the greater the likelihood that those audits will be completed (and new audits started) this
year. Audit productivity is routinely measured in several ways, but the best definition for this
purpose is probably the percent of all examiner time available that is applied to the direct
examination of returns. Direct Examination Time (DET) is generally on the order of 50 percent of
total time, and varies widely across districts and by year. Non-direct activities include vacations,
training, travel, and an assortment of administrative duties. I have used as instruments both the
DET percent and the one-year lag of the average DET per audit.
2.2.3 Alternate Definitions of Income and Offsets
As discussed earlier, the amount of income that taxpayers report is often influenced by
changing rules that dictate what must be reported. For example, prior to the Tax Reform Act of
1986 (TRA86), taxpayers could exclude from total income the first $100 of dividends received
($200, on married-joint returns). After 1986, there was no such dividend exclusion. Unless this
is controlled for, we might interpret the increase in reported income to be an improvement in
voluntary compliance instead of a straightforward response to the new rule. This would be
especially detrimental if the elimination of the dividend exclusion were correlated with other policy
parameters included as explanatory variables (like marginal tax rates). Many more such rule
changes have occurred over the 1982-1991 timeframe of this study—especially in the rules
governing offsets.

24

The Determinants of Individual Income Tax Compliance

No single approach to this problem is ideal, so I have employed a combination of
techniques. Where possible, I have included explanatory variables that control for certain rule
changes—such as FThresholdPct, ChildExemptsPct, and StateTaxPct. In the case of dividends,
I was able to construct a constant-law dividend variable by applying the dividend exclusion to the
post TRA86 micro data before aggregating to the state level. But more often than not, neither of
these two approaches could control for specific rule changes. Therefore, I have constructed the
reporting compliance variables according to three different definitions to test the sensitivity of my
results to these rule changes. The basic approach (Definition A) was to exclude from total income
or total offsets any component whose rules changed during the period, and neither of the other
forms of controlling for the changes was available. The other extreme (Definition C) included all
income and offsets components regardless of rule changes. A hybrid of these two (Definition B)
excluded only those components of total income or total offsets whose rules changed in years other
than 1986. Definition A has the advantage of being purged of such rule changes, allowing a
straightforward interpretation of the parameter estimates. However, it has the drawback that it is
incomplete; it says nothing about the influence of the explanatory variables on the excluded
components of income and offsets. By comparing the results from all three alternatives, however,
it should be possible to conclude something more definitive about the sensitivity of the results to
unaccounted-for tax rule changes.
2.3 Model Specification
The model consists of four compliance equations (FilingRate, IncomePct, OffsetsPct,
and NetIncomePct), and one first-stage AuditRate equation. They are all estimated using singleequation procedures (LSDV and 2SLSDV) to avoid the likelihood of introducing omitted variable
bias across equations. The equations are estimated from the panel of 49 states (i.e., excluding
Alaska, as discussed earlier) over ten years (1982-1991), giving 490 observations for each
variable.
2.3.1 Functional Form
Many of the explanatory variables can be expected to have a non-linear effect on
compliance, reflecting diminishing returns to IRS effort, for example. When this is plausible, I
have expressed the independent variables as logarithms.25 For such variables that frequently take
on values between zero and one, or values near one, I have used the logarithm of one plus the
variable. Otherwise, all variables are modeled linearly.
2.3.2 First-Stage AuditRate Equation
It is both difficult and unnecessary to estimate a structural equation for AuditRate. We
need only estimate a first-stage AuditRate equation using appropriate instruments, and then use the
predicted AuditRate in the structural compliance equations. My specification for the AuditRate
equation, therefore, was as follows:
AuditRate =

δ0 + Σ σ 1iStatei +Σ τ1tYeart+ δ1DET% + δ2Ln(AvgDET-1 +1) + ε1

As with all of the specifications that follow, there are 48 State dummies (i = 1 to 48) and 9
Year dummies (t = 1 to 9), and all of the state and time subscripts on the other variables are
omitted for simplicity.
25

I have used natural logarithms. However, since the instruction for natural logs in my econometric software is simply
“log,” I inadvertently thought I had been using base 10 logs in prior versions of this report. Although this did not affect
the econometric results, it does change any calculations based on those results. This version of the report includes the
corrected nomenclature and calculations. For example, see Figure 6, Table 5, and Appendices G, H and I.

The Determinants of Individual Income Tax Compliance

25

2.3.3 FilingRate Equation
Many of the determinants of reporting compliance are not relevant to filing compliance, and
vice-versa. I used the following specification for the FilingRate equation:
FilingRate =

φ0 + Σ σ 2iStatei +Σ τ2tYeart+ φ1FThresholdPct + φ2Amnesty5
+ φ3Ln(AvgBurden) + φ4SoleProps + φ5SolePropTFS
+ φ6IRP_DocRate + φ7Ln(TDI+1) + φ8Ln(RefOffRate+1)
+ φ9TPS_RetPrepPC
+ φ10Singles + φ11Under30 + φ12Over64
+ φ13AvgPI + φ14AvgPIgrowth + φ15UnemplRate + ε2

The explanatory variables unique to the FilingRate equation are: Amnesty5, since the
common characteristic of all the various state amnesties was that they targeted nonfilers;
Ln(TDI+1), since the TDI program is only targeted to nonfilers; Ln(RefOffRate+1), since the
incidence of refund offsets is likely to induce some nonfiling, but is likely to manifest itself in
reporting compliance (if at all) through smaller amounts of tax withheld, which is not reflected in
any of my reporting compliance variables; and AvgPI and AvgPIgrowth, in order to control for
income variations across states and over time (this is handled in the reporting compliance equations
by dividing the dependent variables by Personal Income). In addition, IRP_DocRate is included
as a linear variable here, whereas it enters the income reporting equations in logarithm form. This
is because the filing decision is an either-or choice; more IRP documents would induce more
people to file—not greater filing compliance among those who already file.
The other variables of primary interest in this equation are: FThresholdPct, which should
indicate whether raising the filing threshold increases or decreases filing compliance;
Ln(AvgBurden), which is likely to diminish filing compliance; and TPS_RetPrepPC, which we
would expect to improve filing compliance.
2.3.4 IncomePct Equation
The income and offset reporting equations include FilingRate as an explanatory variable,
but this is not endogenous. This is because, although we would expect that greater filing
compliance (i.e., more people filing) would increase the aggregate amounts of income and of
offsets reported, the amounts reported on filed returns do not affect whether or not people file
returns. The specification of the IncomePct equation, then, is as follows:
IncomePct =

α0 + Σ σ 3iStatei +Σ τ3tYeart+ α1FilingRate
+ α2FThresholdPct + α3MargTaxRate@$15K + α4MargTaxRate@$57K
+ α5ChildExemptsPct
+ α6Ln(AvgBurden) + α7SoleProps + α8SolePropTFS + α9PaidPrep
+ α10Ln(pAuditRate+1) + α11Ln(IRP+1) + α12Ln(CID+1)
+ α13TPS_CallsPC + α14TPS_RetPrepPC
+ α15Singles + α16Under30 + α17Over64
+ α18PCBirths + α19ExclIncomePct + α20UnemplRate + ε3

26

The Determinants of Individual Income Tax Compliance

Like the other two reporting compliance equations, this equation includes several
variables—in addition to the FilingRate variable itself—that are not in the FilingRate equation.
Additional enforcement variables include Ln(pAuditRate+1) (the predicted results from the firststage AuditRate equation), and Ln(CID+1); additional tax policy variables include
MargTaxRate@$15K, MargTaxRate@$57K, and ChildExemptsPct (which controls for the
changing value of dependent exemptions); PaidPrep is an additional Burden/Opportunity variable,
which is relevant only to filers; and TPS_CallsPC is an additional IRS Responsiveness variable,
which seems to be mostly relevant to filers. The reporting compliance equations also include
PCBirths, which does not appear to affect FilingRate significantly. Finally, as pointed out
earlier, Ln(IRP+1) is used for the income reporting equations instead of IRP_DocRate in linear
form.
The IncomePct equation has two variables not included in the OffsetsPct equation:
Ln(IRP+1), since virtually all information documents report income rather than offsets; and
ExclIncomePct, which accounts for types of income included in Personal Income that do not need
to be reported on tax returns—either because the income falls below the filing threshold (and no
return is required), or the type of income is not taxable.
The explanatory variables of primary interest in the IncomePct equation are:
Ln(pAuditRate+1), FThresholdPct, the two MargTaxRate variables, Ln(AvgBurden),
Ln(IRP+1), Ln(CID+1), and the two TPS variables. These represent the most important tax
policy and tax administration variables that could conceivably be manipulated so as to foster better
voluntary compliance.
2.3.5 OffsetsPct Equation
This equation is much like the income reporting equation, but with some important
differences. The specification is as follows:
OffsetsPct =

β0 + Σ σ 4iStatei +Σ τ4tYeart+ β1FilingRate
+ β2FThresholdPct + β3MargTaxRate@$15K + β4MargTaxRate@$57K
+ β5ChildExemptsPct + β6StateTaxPct
+ β7Ln(AvgBurden) + β8SoleProps + β9SolePropTFS + β10PaidPrep
+ β11Ln(pAuditRate+1) + β12Ln(CID+1)
+ β13TPS_CallsPC + β14TPS_RetPrepPC
+ β15Singles + β16Under30 + β17Over64
+ β18PCBirths + β19UnemplRate + ε4

As discussed above, the OffsetsPct equation does not include two of the variables included
in the IncomePct equation: Ln(IRP+1), and ExclIncomePct. However, it includes one other
variable not in the IncomePct equation: StateTaxPct, which controls for variations in the amount
of state and local taxes that are deductible federally (and therefore contribute to offsets).
The same tax policy and tax administration variables are of primary interest in the
OffsetsPct equation (to the extent that they are included). However, some of these variables are
likely to affect offsets and income in opposite ways, while others have the same kind of impact.

The Determinants of Individual Income Tax Compliance

27

2.3.6 NetIncomePct Equation
The final reporting equation represents income minus offsets, and so includes all of the
explanatory variables included in either of the two separate reporting equations, as follows:
NetIncomePct = γ 0 + Σ σ 5iStatei +Σ τ5tYeart+ γ 1FilingRate
+ γ 2FThresholdPct + γ 3MargTaxRate@$15K + γ 4MargTaxRate@$57K
+ γ 5ChildExemptsPct + γ 6StateTaxPct
+ γ 7Ln(AvgBurden) + γ 8SoleProps + γ 9SolePropTFS + γ 10PaidPrep
+ γ 11Ln(pAuditRate+1) + γ 12Ln(IRP+1) + γ 13Ln(CID+1)
+ γ 14TPS_CallsPC + γ 15TPS_RetPrepPC
+ γ 16Singles + γ 17Under30 + γ 18Over64
+ γ 19PCBirths + γ 20ExclIncomePct + γ 21UnemplRate + ε5
This equation represents the “bottom line” of reporting compliance; it differs from tax
reporting compliance only because the amount of tax reported reflects the application of the tax rate
schedule to net income. (Recall that I have defined offsets—and therefore net income—to include
the income-offset value of credits, which are tax offsets. Therefore, this net income is not
synonymous with taxable income, as reported on tax returns.)
Even though the NetIncomePct equation alone cannot provide the insight into the method
of noncompliance (i.e., underreporting income or overstating offsets) that the other two equations
provide, it does serve two useful purposes. First, when we are interested in tax compliance,
NetIncomePct is the most direct way to estimate it. Second, it provides a useful check on the
other two reporting equations; we should expect that when a variable is included in all three
equations, the coefficient in the NetIncomePct equation should be roughly equal to the coefficient
in the IncomePct equation minus the coefficient in the OffsetsPct equation. This should
especially be true when the coefficients are all significant, and can help to evaluate the results when
one of the coefficients is not significant.
3. Results
The principal results for all four compliance equations are summarized in Table 3. These
results pertain to the most restricted definition of the dependent variables IncomePct, OffsetsPct,
and NetIncomePct (Definition A), which excludes all income and offset components for which the
rules changed during the 1982-1991 period as to what should be reported—unless the rule change
could be reflected in the data (as with the dividend exclusion), or the change could be controlled for
with explanatory variables (as with the value of standard deductions and exemptions, and the
deductibility of state and local income taxes). These results are discussed below, first with respect
to filing compliance, and then with respect to reporting compliance. In the case of reporting
compliance, all three equations are discussed concurrently. For convenience, the explanatory
variables in Table 3 are grouped in the five major categories (Tax Policy, Burden/Opportunity,
Enforcement, IRS Responsiveness, and Demographics/Economics) used in the discussion.
Following this discussion of the principal results is a section that compares the results for
all three definitions of the reporting compliance variables (Definitions A, B, and C), focusing on
the sensitivity of the results to rule changes in the 1982-1991 time period. The fourth section
describes a number of additional variables that were tested as potential determinants of voluntary

28

The Determinants of Individual Income Tax Compliance

compliance, but were found not to have any significant impact. Finally, section 3.5 discusses the
estimated relative merits of expanding the five IRS activities found to have a positive impact on
voluntary compliance.
3.1 Filing Compliance
3.1.1 Impact of Tax Policy Parameters
Of the two policy variables included, only the filing threshold (the sum of one’s standard
deduction and personal exemptions) significantly influences the FilingRate. The impact is
strongly negative, which probably reflects two phenomena: first, as the filing threshold increases,
some people who are still required to file stop doing so (perhaps out of confusion); and second,
low-income people (who are most affected by a change in the filing threshold) may exhibit higher
filing compliance than those with higher incomes—raising the overall FilingRate while they are
required to file, but causing a drop in that rate when they no longer need to file. The latter
possibility is consistent with the strongly negative impact of AvgPI on FilingRate. State tax
amnesties apparently have no significant impact on federal filing compliance, but the results
suggest that they may have a weak positive influence.
3.1.2 Impact of Burden / Opportunity
All of the included variables significantly affect filing compliance—at least at the 10 percent
level of significance. The burden (in hours) associated with the various tax forms and schedules
seems to reduce the FilingRate, as one might expect. This seems to confirm IRS concerns that as
the forms get more numerous and complex, requiring more time to maintain records and to
complete the paperwork, more people decide to forget about filing altogether. The coefficients on
the two sole proprietor variables suggest that sole proprietors, generally, improve the
FilingRate—except for those within the Trade, Finance, and Service sectors, who have a strong
negative impact on filing compliance.26 This may reflect the fact that businesses typically have
multiple “paper trails,” making it hard for them to hide from the IRS, but that these three special
sectors tend to be associated with the “underground economy,” including cash-based businesses
and “moonlighting.”
3.1.3 Impact of Enforcement Activities
Both the matching of third-party information documents (IRPDocRate) and the issuance of
TDI nonfiler notices (Ln(TDI+1)) provide a fairly strong deterrent against nonfiling, as we would
expect. The more information that is provided to the IRS by the payers of income, the more people
will file required returns, since it is harder for them to hide. Furthermore, when IRS uses this
information (and prior filing patterns) to issue TDI notices to presumed nonfilers, the general
population seems to respond with a higher FilingRate than it would otherwise. The two programs
apparently complement each other nicely to promote the filing of required returns.
Refund offsets seem to have a negative impact on filing compliance, but it is too weak to be
considered significant. This suggests that some taxpayers may stop filing in order to avoid paying
the debts being addressed by these offsets, but the most likely response—if any—seems to be to
adjust their withholding to minimize their refunds (a phenomenon that I cannot verify with a model
focused solely on filing and reporting compliance).
26

Note that SolePropTFS is actually an interaction term, the product of SoleProps (the number of proprietors as a
percentage of all potential returns) times TFSEmplPct (Trade, Finance, and Service employment as a percent of total
nonfarm employment). It seems reasonable to view the interaction term, though, as representing the relative concentration
of proprietors in these three sectors.

The Determinants of Individual Income Tax Compliance

29

Table 3. Determinants of Voluntary Filing and Reporting Compliance, Definition A *
Equation
Explanatory Variables

FilingRate

IncomePct

OffsetsPct

NetIncomePct

0.345586

0.137683

0.207853

1.182627

0.857427

0.339758

MargTaxRate@$15K

1.221297

-0.663976

1.921272

MargTaxRate@$57K

-1.978458

0.530545

-2.442911

ChildExemptsPct

1.475395

0.457696

1.000080

0.145114

-0.101522

FilingRate

(7.64)

FThresholdPct

-3.569438

Amnesty5

0.207335

(-8.31)

(4.04)

(8.72)
(8.33)

(5.57)
(1.40)

(0.67)

(1.17)

(-1.00)
(1.86)

StateTaxPct

(-1.80)
(0.76)
(1.63)
(1.99)

(2.22)

(-1.50)
(1.52)

(-0.59)

Ln(AvgBurden)

-11.929189

3.383676

-3.550471

4.888900

SoleProps

1.953925

1.428688

0.274123

1.169128

SolePropTFS

-3.414896

-2.925128

-0.527449

-2.399908

PaidPrep

-0.166282

-0.014858

-0.153009

Ln(pAuditRate+1)

16.158539

3.313904

13.892113

IRP_DocRate

(-1.78)
(2.44)

(-2.30)

(0.55)
(1.84)

(-2.02)
(-4.81)
(3.37)

(-1.77)
(0.98)

(-1.01)
(-1.23)
(2.00)

(0.96)
(1.77)

(-1.95)
(-5.36)
(3.46)

1.565057
(2.71)

-1.121633

Ln(IRP+1)

0.675205

(-0.23)

Ln(TDI+1)

3.850765

Ln(RefOffRate+1)

-0.873704

(0.16)

(1.81)

(-1.13)

Ln(CID+1)

0.932191

0.314909

0.593380

TPS_CallsPC

-0.003994

-0.000742

-0.003378

(3.08)

(-1.34)

(2.96)

(-0.71)

(2.37)

(-1.36)

TPS_RetPrepPC

0.146118

0.130914

-0.007756

0.136453

Singles

-0.551763

0.266954

0.072872

0.190927

Under30

0.186049

-0.098600

-0.008269

-0.091372

Over64

0.242260

-0.075873

-0.042744

-0.022223

0.991262

0.253864

0.734990

(2.18)

(-5.77)
(1.95)
(2.40)

PCBirths

(2.07)
(1.46)

(-1.10)
(-0.76)
(4.55)

AvgPI

-0.920930

AvgPIgrowth

0.192294

(-0.35)
(1.14)

(-0.26)
(-1.23)
(3.58)

(2.61)
(1.26)

(-1.23)
(-0.27)
(4.02)

(-4.32)
(3.68)

ExclIncomePct

-0.642278

-0.917979

(-1.60)

(-2.77)

UnemplRate

-0.200099

-0.370384

0.078118

-0.428625

Adj. R-Squared

0.627007

0.757573

0.919223

0.801260

(-1.50)

(-2.97)

(1.84)

* 2SLSDV estimates (just LSDV for the FilingRate equation) from state-level panel data for 1982-1991;
t-statistics in parentheses; variables in bold are the primary tax policy and tax administration parameters of interest.

(-4.14)

30

The Determinants of Individual Income Tax Compliance

3.1.4 Impact of IRS Responsiveness
The effort of IRS’s Taxpayer Service (TPS) function to help taxpayers prepare their returns
(TPS_RetPrepPC) seems to have a significant, positive impact on the overall FilingRate.
Specifically, we can interpret the estimated coefficient to mean that for every additional seven
returns prepared by TPS per thousand of population,27 the FilingRate will increase by one
percentage point. This suggests that some of the induced returns may be ones prepared by TPS,
but also that some of the new filers are influenced indirectly by the TPS outreach.
3.1.5 Impact of Demographic and Economic Variables
Five of the six demographic and economic control variables included have significant and
intuitive effects on the FilingRate. A greater concentration of singles within a state is strongly
associated with lower FilingRates. This may reflect less of a requirement to file among singles, a
misunderstanding as to when dependents need to file their own returns (especially now with
relatively low filing thresholds for dependents), or—like auto insurance companies have learned—
that singles are typically less careful than their married counterparts. However, higher
concentrations of the population in the under-30 and over-64 age categories each seem to improve
filing compliance. This may reflect fewer opportunities among the young and the old to avoid IRS
notice.
As noted earlier, income seems to be strongly and negatively associated with filing
compliance. Specifically, every additional thousand (1992) dollars of average Personal Income
(AvgPI) is associated with a drop in the FilingRate of almost one percentage point. This
suggests that those who are able to hide from IRS entirely can conceal significant amounts of
income—and as long as it works, they’ll find ways to make more such income—while those who
contribute little to aggregate Personal Income either have no requirement to file, or derive most of
their income from wages and interest, which are hard to hide from IRS.
Of the two variables reflecting the state of the economy, only the rate of real income growth
(AvgPIgrowth) has a strongly positive impact on FilingRate, which we would expect. As real
income increases in the general population, fewer people are tempted to cut corners by not filing a
tax return. By the same token, as the unemployment rate increases, filing compliance seems to
decline. While this result is intuitive, however, it does not appear to be significant.
3.2 Reporting Compliance
Before discussing the impact that specific variables have on voluntary reporting
compliance, it is necessary to ensure that these equations are appropriately identified, given the
endogeneity of audit rates.
3.2.1 Identification
As mentioned earlier, the first-stage AuditRate equation is a function of two audit variables
related to productivity: DET% and AvgDET, where DET (Direct Examination Time) is the time
that auditors apply directly to the examination of returns, as opposed to the time they spend on
leave, in training, or performing various administrative duties. The specification for this equation
is as follows:
27

Note that the coefficient (0.146) is approximately one-seventh.

The Determinants of Individual Income Tax Compliance

AuditRate =

31

δ0 + Σ σ 1iStatei +Σ τ1tYeart+ δ1DET% + δ2Ln(AvgDET-1 +1) + ε1

In words, the percentage of returns subject to a started audit this year is a function of the
percentage of their time that auditors apply directly to such audits (DET_Pct) and the logarithm of
one plus the average amount of time spent directly on audits last year. The more time that auditors
are able to apply to examinations this year ought to increase the number of examinations
conducted, suggesting that δ1 ought to be positive. Similarly, if more time was applied to the
average audit last year, more audits should be closer to being completed—and new audits started—
this year, suggesting that δ2 also ought to be positive. The actual results for this regression (with
t-statistics in parentheses) are as follows:
δ1 = 0.007570,
(1.84)

δ2 = 0.057984,

Adj. R2 = 0.766978

(0.58)

Except for the fact that the impact of the prior year’s AvgDET is not statistically significant, these
results seem encouraging.
There are two indications that the reporting equations estimated using AuditRate predicted
from this first-stage regression (pAuditRate) are identified. First, when we compare the
coefficient on the predicted (and presumed exogenous) Ln(pAuditRate+1) from these reporting
equations with what is obtained using same specifications, but using the actual (endogenous)
AuditRate (i.e., Ln(AuditRate+1)), we see that the two-stage approach corrects the downward
bias anticipated in the coefficients in both the IncomePct equation and the NetIncomePct equation
(see section 2.2.2 above).
These comparisons are given in Table 4 for each of the reporting
Table 4. Comparison of AuditRate Coefficients in the Three Reporting Compliance
Equations, Estimated Using the Endogenous and Exogenous AuditRate Variables
Definition A

Definition B

Definition C

-0.188470

-0.234193

0.747762

(-0.17)

(-0.21)

(0.54)

16.158539

15.823423

17.100229

(3.37)

(3.30)

(2.79)

-0.014032

0.400095

0.468282

(-0.04)

(0.91)

(0.93)

3.313904

0.838515

-0.525384

(2.00)

(0.44)

(-0.24)

-0.164486

-0.600534

0.505934

(-0.18)

(-0.69)

(0.38)

13.892113

15.751865

17.449324

(3.46)

(4.11)

(2.97)

IncomePct Equation
Ln(AuditRate+1) [endogenous]
Ln(pAuditRate+1) [exogenous]

OffsetsPct Equation
Ln(AuditRate+1) [endogenous]
Ln(pAuditRate+1) [exogenous]

NetIncomePct Equation
Ln(AuditRate+1) [endogenous]
Ln(pAuditRate+1) [exogenous]

t-statistics in parentheses. All equations specified as in Table 3. For a complete tabulation of these results, see Appendix C.

32

The Determinants of Individual Income Tax Compliance

compliance equations for all three definitions of the dependent variables. For both the IncomePct
and NetIncomePct equations, assuming that the AuditRate is exogenous yields parameter
estimates that are very small and very insignificant. Accounting for the endogeneity of AuditRate,
however, yields parameter estimates that are larger (as anticipated) and highly significant.
Notice, however, that the same pattern is true with the OffsetsPct equation—even though
we anticipated that the coefficient on AuditRate would be biased upward, and that the corrected
coefficient would be negative. Understanding this counter-intuitive result first requires making
two observations: first, the estimated impact of Ln(pAuditRate+1) on OffsetsPct is much less
than its impact on IncomePct; and second, the estimated impact of Ln(pAuditRate+1) on
NetIncomePct (shown as about 13.89 in Table 4) is very significant, and is the logical
combination of the separately-estimated coefficients in the IncomePct and OffsetsPct equations
(shown as 16.16 and 3.31, respectively in Table 4—their difference being 12.85). All the results
are significant and internally consistent, but why the unanticipated sign on the AuditRate
parameter in the OffsetsPct equation? The logical explanation seems to lie in the fact that the
claiming of offsets is not a simple matter. For example, some offsets, such as medical expenses
and miscellaneous deductions, are specifically limited by the amount of Adjusted Gross Income
reported on the return; as more income is reported, we should expect more to be claimed for these
types of offsets. Likewise, most credits are not refundable; they are limited by the amount of tax
due. If more income (and, therefore, more tax) is reported, we should expect more to be claimed
as credits. It may also be true that if taxpayers feel compelled to report more income (e.g., in
response to an increased AuditRate), they may seek to find additional offsets to reduce the bite
somewhat. This may be especially true if (as one might expect) taxpayers perceive that they may
become audit targets if the offsets they claim seem out of line to the IRS with respect to their
income; so, if they do not report all of their income, they may consciously avoid claiming all of
their potential offsets. The fact that the impact of AuditRate on OffsetsPct is much smaller than
its impact on IncomePct may also mean that audits have some negative impact on the amount of
offsets claimed, after all, but that this effect is more than compensated for by the phenomena just
described. In any event, if the net impact of OffsetsPct on the AuditRate is actually negative (as
is the impact of IncomePct), then the coefficient on AuditRate in the OffsetsPct equation is
biased downward, rather than upward, and the 2SLS estimates correct this bias. This reasoning
suggests that OffsetsPct is dependent, in part, on IncomePct—a possibility not controlled for in
this specification. Not controlling for this does not affect the results of either the IncomePct
equation or the NetIncomePct equation (which represents the “bottom line” results); in fact,
estimating the three separate reporting equations allows us to identify such dependencies. As long
as we are able to interpret the results for the OffsetsPct equation with this dependency in mind, we
can make the correct inferences.
The second fact consistent with the proposition that the reporting equations are identified is
the result of specification tests for endogeneity based on Hausman (1978). This procedure tests
for statistically significant differences between a regression that assumes endogeneity and an
otherwise identical one that does not. The results of this test are given for each pair of such
specifications in the tables in Appendix C. In every important case,28 the test statistic is strongly
significant, which is consistent with notion that AuditRate is endogenous, and that using
pAuditRate corrects for that endogeneity. Moreover, experimentation with alternate specifications
for the first-stage AuditRate equation yielded no alternatives that produced Hausman test statistics
as strongly significant as these. This, together with the observed bias correction, suggests that the
reporting equations are identified.
28

The only exceptions to this are the OffsetsPct equation under Definitions B and C.
drawbacks of these two definitions.

This is consistent with the

The Determinants of Individual Income Tax Compliance

33

3.2.2 Impact of Tax Policy Parameters
Now let us turn to discussing the the impact of the various determinants of reporting
compliance, as summarized in Table 3. As expected, FilingRate positively and significantly
influences the reporting of both income and offsets; as more people file returns, they obviously
report some of each.
More interesting are the effects of the five tax policy parameters included in the three
reporting equations. First, having controlled for the FilingRate, an increase in the filing threshold
seems to increase the reporting of both income and offsets significantly. However, since
FThresholdPct has a strongly negative impact on FilingRate, which is included in the reporting
equations, the net impact of FThresholdPct must take this into account. For example, a one
percentage point increase in FThresholdPct decreases the FilingRate by roughly 3.57 percentage
points, which, in turn, decreases IncomePct by about 1.23 percentage points (-3.57 times 0.35,
the coefficient on FilingRate in the IncomePct equation). So, the increase in FThresholdPct has
virtually no net impact on IncomePct; the apparent increase of 1.18 is just offset by the reduction
of 1.23 due to a decline in the FilingRate. Similar arithmetic reveals that the net impact of a one
percentage point increase in FThresholdPct is an increase in OffsetsPct of 0.37 and a decrease in
NetIncomePct of 0.40. (The fact that these two numbers are opposite and roughly equal is
consistent with the net impact on IncomePct being zero. That is: 0.00 – 0.37 ≈ -0.40). So, an
increase in the filing threshold—controlling for its impact on FilingRate—increases the amount of
offsets claimed (as we would expect), but has no effect on income reported.
The influence of marginal tax rates on the voluntary reporting of income and offsets is even
more interesting. An increase in the marginal tax rate for those with $15,000 of taxable income
increases IncomePct by 1.22 and decreases OffsetsPct by 0.66, resulting in an increase in
NetIncomePct of 1.92, which is strongly significant. In other words, low-income taxpayers seem
to respond to an increase in marginal rates by becoming more compliant. This is consistent with
the Allingham and Sandmo (1972) prediction when the positive “income effect”

[Text truncated at 120,000 characters. The full text is on the page linked above.]

---

Source: Frix Law Library, https://www.frixlaw.com/law-library/documents/agency%3Airs%3A848e5e1d472acc56. Public record. Not legal advice.