Content uploaded by James Alm
Author content
All content in this area was uploaded by James Alm on Nov 13, 2017
Content may be subject to copyright.
Journal of Public Economics 48 (1992) 21-38. North-Holland
Why do people pay taxes?
James Aim, Gary H. McClelland and William D. Schulze*
Received February 1990. revised version received January 1991
Why do people pay taxes when they have an opportunity. even an incentive. IO evade? The
experimental results in this paper suggest that tax compliance occurs because some individuals
overweight the low probability of audit. although such overweighting is not universal. The
results also indicate that compliance does not occur simply because individuals believe that
evasion is wrong. since subject behavior is unchanged by the use of either neutral or loaded
terms. Finally. there is evidence that individuals pay taxes because they value the public goods
that their taxes linance. In short. individuals exhibit much diversity in their behavior.
I. Introduction
In rcccnt years economists have dcvotcd incrcasinp attention to the study
of individual tax evasion. Despite these efforts. our understanding of the
reasons behind individual tax compliance behavior remains limited. In fitct,
the puzzle of tax compliance is that most pcoplc continue to pay their taxes.
This papa uses cxpcrimcntid methods to examine the role that various
factors play in the compliance dccision.
One explanation for compliance strcsscs that the threat of dctcction and
punishment is responsible for complinnce. This theory stems from the
economics-of-crime itpproach, based upon traditional expcctcd utility theory
and first applied to titx evasion by Allingham and Sandmo (1972).’ Here a
rational individual is viewed as weighing the expected utility of the benefits
from successful tax evasion with the uncertain prospect of detection and
punishment. and an individual pays taxes because he or she is afraid of
getting caught. Although it is clear that detection and punishment affect
compliance to ii dcgrec, * it is equally clear that these factors cannot explain
all. or even most, tax compliance behavior. The percentage of individual
Correspondmce ro: J. Aim. University of Colorado. Boulder, CO 80309-0256. USA.
‘We wish Lo thank Charlie PIott, Mark Isaac. and Jorge Martinez for helpful comments on
an earlier version of this paper. Mark Evrrs. Steve Elliot, and Julie Irwin provided excellent
research assistance.
‘The basic model of Allingham and Sandmo (1972) has been extended in a variety of
dimensions. See. for example. Yitzhaki (1974). Pencavel (19791. Sandmo (1981). Cowell (1985).
and Alm (19X8). Cowell (1990) provides a comprehensive survey of the literalurr.
‘For empirical evidence on the deterrent &ccl of government audit and penalty policies. see
Dubin and Wilde (19XX).
0057-2727.92,SOS.oO 3’) l992-Elsrvicr Science Publishers B.V. All rights reserved
22 1. A/m er al.. Why do people pay mres.?
income tax returns that are subject to a thorough tax audit is quite small in
the United States, less than I percent in recent years. In addition. the penalty
on fraudulent evasion in the United States is only 75 percent of unpaid taxes,
and the penalties on non-fraudulent evasion are even less. A purely economic
analysis of the evasion gamble implies that most individuals would evade if
they are ‘rational’, because it is unlikely that cheaters will be caught and
penalized.’ Yet compliance with the individual income tax remains relati-
vely high; that is. individuals pay far more in taxes than suggested by the
standard expected utility theory of compliance. It seems implausible that the
low penalties and the low probability of detection that prevail in the United
States. indeed in most countries, can by themselves act as an effective
deterrent to evasion, unless individuals’ aversions to risk far exceed conven-
tional assumptions. In fact, the Internal Revenue Service (1978) has found
that there are numerous factors other than detection and punishment that
affect the decision to pay taxes.
It therefore appears that there is some discrepancy between the way in
which people actually decide to pay their taxes and the models based on
expcctcd utility theory that have been used by economists to explain this
behavior. Such anomalous behavior has frcqucntly been found in other arcas
of choice under uncertainty, particularly in those arcas that involve low
probability-high loss cvcnts, such as natural or man-made disasters, or in
those arcas whcrc the decisions of individuals arc intcrdcpcndcnt and
repcatcd. such as the voluntary provision of public goods.
Scvcral explanations for this behavior have been suggcstcd in rcccnt years.
One argument stems from thcorctical work of Machina (1983) and Kahncman
%rartr and Wilde (19X5) and Skinner and Slemrod (19X5) make a simihlr nryumcnt. To
illustrutc more prcciscly. consider the stand;ud cvas~on modrl of Allinyh;un and Sandmo (1972).
Hcrc an individual receives a fixed cndowmant of income. I, and must choose how much of this
income to declare IO the lax authorities and how much to under-report. The individual must
pay taxes 31 thr rate I on every dolhlr D of income that is declared. while no 13xcs are paid on
evaded income. tiowcvcr. the individual may be audited with probability p; if audited. then all
unreported income is discovrrcd. and the individual must pay :I penalty a~ the rate 1’ on each
dollar that he or she was supposed to pay in taxrs but did not pay, The individual’s incomr. I,.
if caught under-reporting equ~ds I,=f -rD-l[r(l-01). while income. I,. if not caught is
I, = I -rl). Expec~cd utility theory then assumes that the individual will choose D in order to
maximize expected utility EU =pU(/,) +( I +p)U(/,). where U(,) is the utility function. Now
suppose that this optimization is solved for specilic. realistic values of the various parameters
and ti,r the spccitic utility runction. I,! -‘!‘I I -e). where the subscript i rcfcrs to the sta~c of the
world (i =C.N) and c is 3 measure of the individual’s constant relative risk aversion. For
ex;lmple, il r = 0.4. /’ = 2, p = 0.0 2. and e = I. then the individual will optimally declare no income.
Very large valucx for rrlativrs risk aversion are rcquirrd IO generate compliance consistent with
aclual U.S. experience. When e= 3. declared income is only I4 percent of true income: when
e= 5. it is still only 44 percent; when c= IO. it is only 71 percent. Risk aversion must ctceed 30
for compliance to exceed 90 percent. However. existing ticld evidence suggests that c rangrs
between one and two [Cohn et al. (1975). Friend and Blume (1975). Morin and Suarcz (19X3)].
Consequently. cxpeclcd utility theory predicts that the optimal strategy for most individuals is IO
report littlr or no income. Tar Ices than is actu;dly observed.
J. Alm et al.. Why do peopie pu.v ta.re.~.7 23
and Tversky (1979). Using different approaches. they argue that individuals
either can show great sensitivity to or can overweight low probabilities.
Suppose. for example. that the true probability of an event is 0.02. In making
their decisions, however, many individuals will systematically behave as if
they think the likelihood of the event exceeds 0.02, when their behavior is
viewed from an expected utility perspective. Overweighting of low probabili-
ties may therefore provide an additional explanation for compliance. If
taxpayers give more weight to the probability of an audit than they ought to
relative to an expected utility model, then compliance will be greater than
the level suggested by expected utility theory.’
Another factor arises from theoretical and experimental work on public
good provision. Samuelson (19%) argued that the private provision of public
goods will be inefficiently low because each individual will have an incentive
to ‘free ride’ on the private purchases of others. The Samuelson result has
received some support from the experimental work of Kim and Walker
(1984). Isaac, Walker and Thomas (1984). and Isaac, McCue and Plott
(1985). However, other work argues that voluntary provision may not always
play as a prisoners’ dilemma game; that is. an individual’s optimal decision
may depend upon the actions that hc or she cxpccts others to follow, both
now and in the future.’ Under some circumstances. full voluntary contribu-
tions (or cooperation) may be the dominant individual strategy. This work
thcrcforc suggests that individuals pay taxes voluntarily because they value
the goods provided by govcrnmcnt and they rccognizc that their payment
may bc necessary to get others to contributc.h Howcvcr. the role of public
good provision in tax compliance has largely been ncglcctcd.’
The purpose of this paper is to use cxpcrimcntal methods to examine the
‘Studies of flood and carrhqudc insurance [Kunreuthcr cl aI (lY7X)j and of the vduc’ of
avoiding cxp~urc IO hazardous suhst;mccs [f~urncss CI al. ( IYX.3). Smith and C~csvousgcs (1956)J
alI suggest extraordinary lcvcls of risk aversion. Laboratory experiments by McClelland. Schulrr
and Coursry ( IYX6) also confirm that individds oftsn overweight low probabilities when they
have the opportunily IO purchase insurance IO protect against an uncertain loss of income. Marc
generally. there is growing evidrnce from numerous areas tha[ indicarcs that individuals do not
always behave in a manner consistent with cxpectrd utility theory. SW Machina (19X7) for a
&tailed discussion of this li(erraturr.
‘For rxamplc, the theoretical work of Taylor (lY76). Axelrod (IYXJJ, Palfrey and Rocrnthnl
(lW4). and Bagnoli and Lipman (1986) shows thar voluntary provision can lead to Ihe efiicirnl
level of public good provision, despite the usual prrsumplion that individuals will attempt to
frcr ride. The rxperimen[al litrralurr has mixed resulks. but there are numerous instances in
which these theorrlicnl results hove hccn verified [Brubaker (19X2). Rapoporr (19X5). rrrd
Uagnoli and McKee (IYYI)].
‘There are. of course. other factors that may explain compliance. For example. psychologias
and Other social scirntids argue that so&d norms and perceptions of fairness nlkcr compliancr;
thar is, individuds pay taxes because they furl thal it is a social obligarion IO do so. although
they comply less if they perceive th;ll they are tread less fairly than others.
‘Ser. however. Becker. Buchncr and Sleeking (19X7). who examine cxperimendly the role of
transfer paymrnls on compliance. and Cowcll and Gordon (19xX). who analyze theoretically the
dTcc[ of governmenl rxpcnditurcs on lax compliance and show Ihal a higher I;IX rale can Id IO
less compliance.
2‘4 J. Alm et of.. Why do people pay taxes.?
roles that overweighting of low probabilities and recognition of government
services play in the individual’s tax compliance decision. Experimental
methods have been applied to some issues in tax compliance by Friedland,
Maital and Rutenberg (1978). Spicer and Becker (1980), Friedland (1982).
Spicer and Hero (1985). and Becker. Buchner and Sleeking (1987). among
others. In the experimental design used here, subjects are faced with a typical
tax compliance decision: they receive income. they must decide how much
income to report as taxable income, knowing that there is some probability
that they will be caught and penalized if they do not report all of their
income. and they receive a return for their taxes (or a public good) that
depends upon the level of group tax payments. The parameters relevant to
their decision - the tax rate, the probability of detection, and the like - are
based upon values that individuals actually face. In particular, several levels
for the probability of detection are used in separate treatments, with the
values set at low levels, and several levels for the public good are also used
in separate treatments.
The experimental results provide strong evidence that some individuals
overweight low probability events; that is, when the probability is non-zero
but is low enough that evasion is the optimal strategy, the level of
compliance far cxcccds the level predicted by expected utility theory.
However, behavior is somewhat more complicated. For example, there is
cvidcncc that individuals do not always exhibit ovcrwcighting or extreme risk
aversion. since thcrc is still some compliuncc when the probability of
detection is zero and thcrc is still some evasion when individuals face a high
enough probability to make full compliance the optimal strategy. The results
also indicate that compliance behavior does not stem from a bclicf by
subjects that evasion is wrong. Some cxpcrimcnts arc run twice. once with
neutral terminology that makes no mention of taxes or evasion and instcnd
treats the cxpcrimcnt simply as a risky decision, and once with instructions
that clearly place the experiment in the context of tax evasion. Both
treatments yield identical results. Finally, the results suggest that some
individuals pay taxes bccausc they recognize that payment is necessary to
receive government goods. An increase in the payoff that individuals rcccive
from a given tax payment increases compliance, and compliance is positive
and stable over time cvcn when the probability of detection is zero.
Section 2 discusses the design of the expcrimcnts that examine the roles of
overweighting and public goods in the compliance decision, and section 3
reports the results from the experiments. Summary and conclusions arc in
section 4.
2. Experimental design
The subjects used in the experiments arc volunteers drawn from undcr-
J. Alm et al.. Why do people pay taxes.7 3
graduate classes at the University of Colorado at Boulder. and are allowed
to participate only once in the experiment.’ At the beginning of a round.
each of the eight subjects is given one of eight incomes between SO.25 and
52.00 in SO.25 increments, randomly chosen by computer. The subject must
decide how much income to report, and must pay taxes on all reported
income at the rate of 40 percent. The subject pays no taxes on unreported
income: however, the subject is told that there is some probability that his or
her underreporting will be detected by an audit, at which point he or she
must pay a penalty equal to 15 times all unpaid taxes.’ An audit is
determined by the draw of a chip from a bag that contains a total of 100 red
and white chips. If a red chip is drawn, an audit of all subjects occurs; if a
white chip is drawn, no audit occurs. Three different levels of red and white
chips are used in the experiments.
After taxes arc paid and penalties if any are assessed, the total taxes paid
by all subjects arc summed to give the ‘group tax fund’, increased by some
multiple (or the ‘group surplus multiplier’) to reflect the consumers’ surplus
that individuals dcrivc from povcrnmcnt provision of a public good. and then
divided equally among the subjects. Three different group surplus multipliers
on the group tax fund arc used. The net balance for each subject is
calculated (original income less taxes less pcnaltics plus share of the
multiplied group tax fund). A new round then begins, with the subject’s
balance carried over from the previous round. Subjects arc not allowed to
communicate with one another during the cxpcrimcnt. At the completion of
the cxpcrimcnt, the subject keeps all the money that hc or she has
accumulated; each is guaranteed a minimum of $5.00 for participating, and
earnings arc typically botwccn $ I5.00 and $25.00. The cxpcrimcnts arc
conducted in the Laboratory for Economics and Psychology (LEAP) at the
University of Colorado. AII cntrics arc made and rccordcd on computer
terminals. and all calculations arc pcrformcd by the LEAP MicroVAX
computer. The cxpcrimcnt typically lasts less than one hour.
Thcrc arc six basic trcatmcnts, with ;I different set of subjects each time.
Three sets of eight subjects face each of three different probabilities of audit
during the experiment: 0 percent, 2 pcrccnt. and IO percent. The order in
“The instructions given IO subjects arc awil;~hlc upon rcquert.
‘A penalty multiphrr of IS times unpaid taxes may seem quite large. since actwl pcnaltics for
incoms 1x1 fraud arc currently 75 percent of unpaid I~XL’S plus the unpaid taxes. However, it is
important IO rccognirc that the di>covcry of fraud in one yrar Icclds IO Internal Rcvenuc Service
(IRS) invcstigafion of potential fraud in previous years. The lncomc Tax Code (Section 6501 (C))
specilics that taxes 2nd penalties may he asscssrd al any time for fraud; that is, the IRS csn
cxtcnd its investigation any numhcr of years in the ptls~ when it disrovcrs fraud. If. for example.
the IRS cxtendcd its investigation for six years into the pss~. then the cfkctivr penalty multiplier
is 10.5 (or 6x 1.75). When intcrcst penalties and. more significantly. legal costs arc also
conridcrrd. a pentllty multiplwr of I5 does MI scum unreasonable. It also capture the type of
catastrophtc loss that dctectwn of evasion okn brings.
‘6 3. Aim er al.. Why do people pay taxrs?
Table I
Experimental design.’
Rounds
Session l-15 16-30 3135
lb p=o p = 0.02 p=O.lO
?b
5 p = 0.01 p=O.lO p=o > m=2
p=O.lO p=o p = 0.02
‘4 m=2 m=O m=6
5 m=O m=6 m=? p = 0.02
6 m=6 m=? m=O >
- ‘In all sessions. the tax rate is 0.40. the penally
rate is IS. and the individual’s share of the group
fund is 1 R.
%zssions 1. ? and 3 are run twice. once with
neutral inslructions and once with loaded
mstructions.
which the subjects fact those probabilities is varied across sessions. In the
first session the order is 0, 7, and IO pcrccnt; in the second session the order
is 2. IO, and 0 pcrccnt; the order is IO, 0. and 2 pcrccnt in the third session.
All subjects fact each probability for I5 rounds; they arc not told the number
of rounds, although they arc told that the number of rounds is prcdctcr-
mined. The group surplus multiplier equals ?. in all of thcsc trcatmcnts. Three
sets of tight subjects also fact each of thrco diffcrcnt lcvcls of the group
surplus multiplier. The lcvcls arc 0, 3, and 6. As with the probability, the
order in which the subjects fact thcsc multipliers is varied across ssssions,
and all subjects fact each multiplier for 15 rounds. The probability of
dctcction equals 2 pcrccnt in itll of thcsc latter cxpcriments. The cxpcrimcntal
design is summarized in table I.
This cxpcrimcntal design is diffcrcnt in several rcspccts from previous
cxpcrimcnts in tax compliance. “I Previous work has not always used values
for the various policy parameters that approximate actual real-world values.
In addition, some acccptcd proccduros of the cxpcrimcntal paradigm have
not always been followed. Many of the results have not been verified by
rcpcatcd rxpcrimcnts. Of perhaps more importance, previous work may have
inadequately induced subject preferences because the instructions given to
the subjects have placed the experiment squarely in the context of tax
evasion; that is. the terminology used in all previous work may have
provided an undcsirablc context to the experiments, and so the decision of
the subjects may have reflected the values that they associate with such terms
as ‘tax compliance’ or ‘tax evasion’, rather than the rewards or pcnaltics that
‘%e. for example. Fricdland. Mailal and Rulenherg (lY7X). Spicer and Becker (1980.
l.‘rlcdland (lYlc2). Spwx and ticro (lYN5). and Becker. Buchner and Sleeking (19X7).
J. Alm er al.. l4’h.v do people pay tares? 27
they faced in the experiment per se. To explore the role of such terminology
in the compliance decision, two treatments are examined, one that uses
neutral terminology and one that uses tax. or ‘loaded’. terminology.”
Finally, most previous experimental work on tax evasion has not considered
the role that government provision of public goods plays in tax compliance;
that is, individuals may voluntarily pay taxes to provide for public goods,
even if there is no penalty on the failure to pay. because they recognize that
they will receive something for their tax payments.” The experiments here
recognize that tax compliance may be affected by this exchange relationship
between the taxpayer and the government.
Note that the optimal single-period strategy for each subject can be easily
determined when the individual’s goal is to maximize expected value, as
implied by expected utility theory in the special case of a risk-neutral
individual who takes the actions of others as given (Cournot-Nash beha-
vior). The expected value, Er! from the choice of how much income to report
is
EV=f-tD+,ns(G+tD)-pJ‘[t(f-D)], (1)
whcrc I is the individual’s tixcd income, D is dcclarcd income, G is taxes paid
by all other group members. t is the tax rate on declared income, /’ is the
tine rate on unpaid taxes, p is the probability of detection, 111 is the group
surplus multiplier. and s is the individual’s share of the group tax fund.
Maximization of cq. (I) by the choice of dcclarcd income f1 indicates that the
individual will optimally report all income if
pj’ + IfIS > I, (2
while the individual will report zero income if the inequality is reversed.
When the group surplus multiplier equals 2 (and the share of the public
good is 0.125), inequality (2) suggests that the dominant strategy for risk-
neutral individuals is to report zero income for the experiments in which the
probability of detection is 0 and 2 percent, and to report all income when the
probability is IO percent. More generally, the critical (or ‘cutoff’) level of the
probability is 0.05 (=[ I-2 x O.lZS]/lS). However, if overweighting of low
probabilities occurs, or if individuals bccomc extraordinarily risk averse at
low probabilities, then there will be some compliance at 9 percent prob-
“The issue of loaded versus neu~rA instructions is discussed in more d&l b&w.
“An exception is Rcckrr. Ruchner and !&eking (19X7). who examine the role of tmnsfer
payments on compliancr.
Treatment
Table 2
Average group compliance rate.
Probabihty of audit (m = 2)
p=o p = 0.02 p=O.lO
XIesn, sessions I. 2. 3
Neutral instructions 0.200 0.503 0.675
Loaded instructlons 0.189 0.522 0.672
Group surplus multiplier
Ip = 0.02)
m=O m=Z m=6
Mean. sessions 4. 5. 6 0.435 0.537 0.592
ability. even though it is less than the critical probability. Similarly. when the
probability of detection is 2 percent (and s = 0. IX), risk-neutral individuals
who behave according to the expected utility model should report zero
income for the treatments in which VI equals 0 and 3, and should report all
income when IPI equals 6.” However, if subjects recognize the exchange
relationship, then there will bc some compliance at the lower multiplier
values.
3. Expcrirnental results
3.1. I’rohhility fj’mtfif
The results for the variations in the probability of audit arc summarized in
table 3. which gives the avcragc of the group compliance rates at each of the
probnbilitics for the three groups. The average group compliance rate is
calculated by dividing total reported income of all group members by total
group income.
Consider each of the three probabilities. According to expected utility
theory, the single-period dominant strategy for a risk-neutral individual is to
report zero income when the probability of detection is less than 5 percent.
However, there is clearly substantial compliance at 2 percent probability.
Compliance at p=O.O2 is on average 50.3 percent; the amount of compliance
by group varies little, from 45.3 percent for group 3 to 48.1 percent for group
I to 57.5 percent for group 2. Although expected utility theory is unable to
explain this result, it is consistent with the overweighting of low probabilities
that is suggested by Kahneman and Tversky (1979) or the extreme aversion
to risk at low probabilities that is suggested by ivlachina (1983). As discussed
below, compliance may also be motivated by the presence of the public good.
“The CUIOIT vaiuc for m is S.h( = [ I -0.02 x ISJO. 125).
J. Ah er al.. Why do people pay taxes? 29
In particular, coupling an enforcement mechanism with voluntary provision
may help overcome the free-rider problem.
Even at p=O there is substantially more compliance than is predicted by
expected utility theory. The average group compliance is 20.0 percent.
although there is some variation across the three groups (5.3 to 35.8 percent).
Expected utility theory predicts that there will be zero reporting of income
when the probability of detection is zero. This prediction is not observed,
due to the presence of the public good. It is also clear that this result cannot
be explained by overweighting or risk aversion, since the probability of
detection is zero.
Note, however, that compliance is actually less than that predicted by
inequality (2) when p=O.l. When the probability is 0.1. the expected return
to evasion becomes negative. and inequality (2) predicts that subjects will
fully comply. However, the avcragc group compliance is only 67.5 percent.
As with p=O. this behavior cannot be explained by overweighting or risk
aversion. Instead, subjects now appear to be risk-seeking, perhaps because of
the guaranteed payment of $5.00. Also, some subjects may dismiss the
likelihood of an audit bccausc of the relatively low level of the probability,
even though it exceeds the critical level, and some subjects may try to free
ride due to the prcscncc of the public good.
The rate of compliance rises in a non-linear way as the probability of
detection increases; that is. tax rcvenucs increase with grcatcr enforcement
efforts, but this payoff dcclincs as the probability incrcascs. At p=O,
compliance is 20.0 pcrccnt; the compliance rate incrcascs significantly to 50.2
pcrccnt at p = 0.02, and then rises but only to 67.5 percent at p=O.l. Thcsc
differcnccs arc highly statistically significant.14
‘*Statistical ;mdysis of the diffcrencc in mc;ms rcquircs that the observations be normal and
independent. Scvcral alkrnativc approxhcs arc followed to gener:lle average complisncc rates
that satisfy thcsc requircmcnts. In the lirst approxh. Ihc round is trc:lted as the unit of
observation. The average compliance raft for a given round is cdcula~cd by averaging for each
round the compliance rats across all right subjects in all three replications of a given
probability; this procedure generates 45 IOI~I obscrvdtons. or one for each of the fifteen rounds
fur each ol the three probabilities. The sample I~SI st;ltistic for the difTercnce between the average
compliance rates JI p=O and p=O.O2 is 16.56. and is 8.60 for rhc di!Tcrcncc bctwecn the average
compliance rdcs ;II p =0.02 and p = 0. I; the critical value of the r-statistic is 2.47 for a one-&led
ICSI 31 the 0.01 significance level with 28 degrees of freedom. In the second and more
conservative approach, the group is treated as the unit of analysis. Here the average compliance
rate IS calcula~cd by averaging across all right subjects und all liken rounds in a given group
for each probabdity level; this approach generates 9 total observations, or one fur each of the
three groups at each of the three problrbiliries. Bccausc there may be non-independence induced
by having the same subjects a~ each of the three probabilities. difkrencc scores arc then
calcula~cd, so that there is only one dirercncr for each group. For example, the sample test
slarisuc for the dillcrcncc between the average compliance rates 31 p =O and p =0.02 is computed
by first calculaling for each group the difTercncc between rhc compliance rates DI p=O and
p =O.OZ. and then using these three differences IO form the sample test statistic; the resulting I~SI
statistic is 5.69. The ICSI statistic for p=O.O2 versus p=O.l equals 4.1 I. Both difkrenccs arc
highly significant.
30 J. Ah er al.. Why do people pay tars”
Fig. I gives a more dctailcd picture of the avcragc amount of compliance
by round for the three probnbilitics. When the probability of dctcction is 0,
thcrc is substantial compliance and the amount of compliance dots not
decay as the cxpcrimcnt proceeds. I5 This trcntment corresponds to the .
voluntary provision of public good cxpcrimcnts - there is no penalty on non-
payment of taxcs - so that the results hcrc may bc compared with those
from the public good literature. In particular, the results contrast with those
of Isaac, McCuc and Plott (19X5). in which they find funding lcvcls for the
public good that arc near zero and that decay with rcpcatcd rounds. Here
there is substantial compliance that dots not on avcragc fall in later rounds.
Fig. I also shows compliance by round when the probability of dctcction
is 0.02. Although the dominant stratcpy from cxpcctcd utility theory is to
evade. the average group compliance rate is 50.3 percent, and this does not
vary much by round. On the other hand, compliance is lower than that
predicted by cxpcctcd utility theory when p=O.l.
Fig. L clearly shows that the level of compliance increases with the
probability of dctcction. It also demonstrates that the predictions of a risk-
neutral version of cxpectcd utility theory arc not vorificd: thcrc is grcatcr
compliance at low probabilities (p=O and p=O.O2) and less compliance at
Hidden within fig. 2 is some individual variation in compliance across
“Thcrc is. howcvcr. romc variation xross groups. In groups I and 3. compliance tluc~ua~cs
widely. Appurcntly some suhjccts cqxrimcnkd with tar payments in the hop Ihal others would
follow: when others dd not follow. maws Ml IO zero.
1. Aim et al.. Why do people pay taxes? 31
60
50
40
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.6 0.9 1
Fraction of Income Declared
Fig. 2
high probabilities (p=O.I) than is predicted by cxpcctcd utility theory. The
way in which individuals actually weight the probability of an uncertain
event is apparently not the way that is assumed under expected utility
t hcory.
It is also of some intcrcst to cxamine the individual data. Fig. 3 presents
the frcqucncy distribution of the individual compliance rates for the three
probability trcatmcnts.“’ Thcrc is much cvidcncc of all-or-none behavior, as
would result from any linear (or risk-neutral) payoff function for the
individual. Overall, 67 percent of the individual decisions arc all-or-none.
Howcvcr, this varies with the probability of detection. When p=O. 7X pcrccnt
arc all-or-none; for p=O.O2 this drops to 64 pcrccnt. and it drops still further
to 61 pcrccnt for p=O.l.
Recall, however, that (risk-neutral) cxpcctcd utility theory predicts that the
cut-off lcvcl for the probability of detection is 0.05: when ~~0.05, compliance
should be 0, while compliance should be complete when p>O.O5. These
predictions are complctcly rejcctcd by the results. It therefore appears that
individuals are largely risk neutral and rationally select and follow a cutoff
rule. However, many do not use expcctcd utility theory in the selection of
that cutoff value.
rounds. A subject apparently may try different strategies in the compliance
game in nn attempt. for example, to induce others in the group to comply.
Recall that the subjects know the size of the group fund at the end of each
lbFor each Irwtment lhcrr arc 360 observations. or three replications timcs right subjects
times IS rounds. The individual compliance rate for any round equals declared income divided
by true income.
32 J. Ah et al., Why do people pay tares.”
“OI
0.9 J
06-
0.1 - (4 mr6 1
0.0 1
0 2 4 6 6 10 12 14 5
Round
Fig. 3
round, so they can dctcrmins whcthcr other group mcmbcrs respond to their
decision to report more income (and they ciln infer whcthcr others arc
complying or cheating). Thcrc is in fact a positive and significant correlation
bctwccn each individual’s compliance rate in ;I given round and the amount
of the public good rcccivcd in the previous round.” Eventually. howcvcr,
nearly all subjects scttlc into ;I stable strategy. and their compliance dots not
fluctuate. Subjects also gcncrally bchavc consistently across the three diffcr-
cnt levels of probability. Thcrc arc virtually no instances in which ;I subject
complies at p = 0 or p = 0.03 and cvadcs at p = 0. I, or vice versa.
The results for the group surplus multiplier arc summarized in table 2 and
in figs. 3 and 4. Recall that under expected utility theory the sin&period
dominant strategy for ;L risk-neutral individual is to cvadc fully when the
multiplier equals 0 and 2 and to comply fully when IPI = 6. These predictions
“This rcsuh c‘onfr;rsfs somewhat with that of Spicer and Hero (I9XSj. who found in their
experiments Ihat an individual’s compliance did not depend upon his or her pcrcep(ion of the
comphancc behavior of others. However. in their experiments subjects dtd not know how other
mcmhcrs of rhcir group actually hchaved. since (heir perceptions were had upon drcuptivr
informntion provided by the experimenters ahoul average group compliance rates In cxpcrimcnrs
that were not m &I run. Also. suhjccls in their cxpcrimcnts were not ahlc IO AXI the behavior
of other group members by their own compliance decisions, since there was no group tax fund
and (hew was no interaction among the group mcmhers.
J. Aim et (11.. Why do people pay taxes? 33
0 0.I 0.2 0.3 0.4 0.5 0.6 0.7 0.6 0.9 1
Fraction of Income Declared
Fig. 4
arc not supported by the results. Compliance signiticantly cxcceds zero when
the multiplier is 0 and 2, .
,md fitlls below full compliance when the multiplier
is 6.
The trcatmont in which m=O corresponds to most carlicr compliance
cxpcrimcnts, since subjects rcccivc nothing for their tax payments. Thcrc is
substantial compliance in this trc:ttmcnt, with the avcragc group compliance
rate cqual to 43.5 psrccnt. Compliance hcrc is motivated solely by fear of
dctcction ;md punishment. If individuals ovcrwcight the low probability of
dctcction (p=O.O2). then compliance will cxcccd the zero level predicted by
cxpcctcd utility theory. The trcatmcnt in which m=2 gcncratcs an cvcn
higher avcragc group compliance rate (53.7 pcrccnt), since the payoff to taxes
is grcatcr. Again. the combination of enforccmcnt, overweighting of low
probabilities, and 3 voluntary provision mechanism helps to overcome frcc-
riding. Note that compliance of 53.7 percent is not statistically different from
complinncc (50.3 percent’ under the earlier probability treatment in which the
various paramctcrs ;Lrc identical. An increase in the multiplier to 6 incrcascs
compliance further. but only to 59.2 percent. There is not full compliance ;Lt
rr1=6, possibly due to risk-seeking. dismissal of the (low) probability of audit,
or free-riding.“’
‘“As with the probaility sessions. IWO alternative approaches are used to analyze the average
compliance ralrs (see footnote 14). When the round is the unit of andysis. so that there are 45
IOI:I~ observations. rhere arc signilicanl dillcrencrs between the average compliance rates ~1 m =O
and m = 2 (I-slalistic = 5.74). and al m = 2 dnd m = 6 (f-ctaGstic = 2.79). When the group is kated
These results demonstrate that compliance increases with the group
surplus multiplier. and suggest that government can increase compliance by
providing goods that their citizens prefer more. by providing these goods in a
more efficient manner, or by more effectively emphasizing that taxes are
necessary for receipt of government services. However, although compliance
increases with m, the increase is non-linear. As with the probability of
detection, there appear to be limits on how much governments can affect
compliance by increasing the individual payoff to tax payments.
Compliance by round for the three levels of the multiplier is shown in fig.
2. There is substantial variation in the average group compliance rate,
especially for m = 0 and VI = 6. However. this variability does not exhibit any
trend: that is, there is no systematic tendency for compliance to rise or to fall
as the experiment proceeds. More importantly. fig. 3 demonstrates again that
with few exceptions compliance is greater when the group surplus multiplier
is greater. Individuals pay more in taxes when they receive more for their tax
payments.
The frcquoncy distribution of individual compliance rates is shown in fig.
4. As with fig. 2. the compliance behavior tends to be all-or-none. with 60
pcrccnt of all decisions in these categories. although this tendency is
somcwhat wcakcr than in the probability trcatmcnts. Fig. 4 also shows that a
risk-neutral version of expected utility theory is not supported by the results.
Again, most individuals follow a cutofT rule in their compliance behavior, but
their behavior suggests that they do not use cxpcctcd utility theory in the
dctcrmination of the cutoff value. Individual subjects also bchavc consistently
across the throc group surplus multipliers, increasing (or not dccrcasing) their
cvmpliancc when the multiplier incroascs.
As discussed carlicr. all previous cxpcrimcntal work on tax evasion has
used instructions that make it clear to subjects that they are participating in
a tax compliance game. Unfortunately, use of such loaded instructions may
lead to unpredictable results. since subjects may respond to the values they
associate with the loaded terms rather than to the incentives in the
experiment itsclc that is, loaded instructions may lead to context effects that
can inllucncc the results in unpredictable ways [Kahneman and Tvcrsky
(1979). Machina (1987)]. Concern with framing cfTects has led most expcri-
mcntalists to USC it neutral terminology, one that masks the context of the
experiment and that gives the experimenter more control over the laboratory
setting. However. thcrc are few studies in experimental economics that have
J. Aim et ul.. W’hy do people pup faxes.? 35
directly examined whether results are affected by the use of loaded or neutral
instructions: there are no evasion experiments that have examined this issue.
This section looks at this issue by comparing two treatments (see table I). In
the additional treatment described here. the instructions include loaded terms
like taxes. audit, reported income. and penalty; in the base treatment. a
neutral terminology was used (payment. check, disclosed money. shortfall).
In experimental instructions, context effects might occur because the use of
loaded words and the inclusion of irrelevant material may lead subjects to
invoke different ‘mental scripts’. which enable the subject to till in missing
information in the instructions but which also may unpredictably influence
subject choices. Of course. the more explicit and complete the instructions.
even in the presence of loaded terms, the less will subjects have to rely on
scripts to fill in missing information. In this case it seems likely that there
should be little difference between loaded and neutral instructions.
This reasoning suggests that if subjects are given complete and precise
information on the situation they face in the laboratory - as they are in the
experiments here - then scripts should not bc needed to help subjects till in
missing context. Since the context is already complete, the USC of potentially
loaded words such as tax. audit, or penalty should have no impact on the
behavior of subjects relative to a treatment that does not WC such words.
Note, howcvcr. that if true uncertainty is present in an experiment (say,
prohahilities arc unknown), loaded words may call up scripts that suggcsl
the probability Icvcl and so affect choices.
The results arc summarized in table 2, which gives the average group
compliance rate across trcatmcnts. for each of the three probabilities.“’
Trcatmcnts using loaded and neutral instructions give results that arc
virtually identical. Dcbricfing of subjccls in the neutral terminology treatment
also suggcstcd. somewhat surprisingly, that Ihey did not realize that they had
participated in a ‘tax’ cxpcrimcnt. Thus, although the conlcxt apparently
dill&s for subjects in the two trcatmcnts, subjects arc not forced to rely upon
their own scripts (i.e. real-world knowlcdgc of the tax system) to complete
the context because in both trcalmcnts they are given identical and full
information on the parameters relevant to their choices, including probabili-
tics of audit. Consequently, their behavior is unaffcctod by the nature of the
instructions. This result implies that previous compliance experiments may
not have been substantially affected by their use of loaded terms. at Ieast if
they provided complctc information on the situation facing subjects.
4. Conclusions
Why do people pay taxes when they have an opportunity, even an
incentive. to evade? The experimental results in this paper suggest that tax
compliance occurs because some individuals are oversensitive to or over-
weight the low probability of audit that they in fact face. When combined
with the high penalty on detected evasion, individuals do not behave as if
their preferences are linear in probabilities. Rather, they often pay more in
taxes than a simple application of expected utility theory would suggest.
However. there is also some evidence that compliance is not always due to
overweighting or to extreme risk aversion, since there is some compliance
when there is no chance of detection and there is some evasion when the
expected value of the evasion gamble is negative. Furthermore, compliance
does not occur from a belief by subjects that evasion is wrong, since their
behavior is unchanged by the use of either loaded or neutral instructions.
Finally, the results suggest that compliance occurs because some individuals
value the public goods that their tax payments finance. An increase in the
amount that individuals receive from a given tax payment increases their
compliance rate, and individuals pay something in taxes to receive govern-
mcnt services even when there is no chance of detection and punishment.
In short. individuals exhibit a remarkable diversity in behavior. They
somctimcs appear to ovcrwcight low probabilities. they sometimes appear to
bc risk-seeking, they arc on occasion coopcrativc, and at other times they are
free-riders. Whcthcr any theory of compliance can bc dcvclopcd to explain
such behavior is unknown at present. However, it is apparent that some new
theory is nccdod. one can then dual with those factors that arc either ignored
or arc dealt with unsatisfactorily by cxpcctcd utility lhcory.
Hefwenccs
AIlingham. \llch;wl G. ;~nd Agnar Sandmo. lY72. Income lax ewsion: A thcorelicd analysis.
Journal of Public liconomlcs I. 323-338.
Aim. James. IYM. Compliance CO\I~ nnd rhr IU ;Ivod;mcc I;IK cvasiun decision. Public I.‘inance
Quarlcrly 16. no. I, 33 6’).
Awlrod. R.. IYXJ, The e\olulion of cooperation (Basic Books. New York).
B;lynoli. .LlJrk ;mJ B.L. Lipman. IYXh. Can private provision of public goods bc rllicicnt?.
llniverslty of ,Mlchlgan Working paper.
i%;lgnoli, %f;Irk and Xlichacl McKee. lY91. Vdunrury contrihulion games: ElTicicnr private
proviwn of puhl~c goods. IGxmomic Inquiry 29. no. 2. 351 -366.
Rcckcr. Winlfrd. tIcIn/-Jurgcn Buchncr and Simon Sleeking. 1997, The impact of public
transfer sxpcndilurc\ on tax cvaswn: An rxperimenral approach, Journal of Public &on-
omits 34. 243-251.
Bruhakcr. Ed. IYXZ. Sixty-cighr percent free rrvelntion and thirty-two percent free ride’!
Demxd disclosurw under varying conditions on exclusion. in: Vernon L. Smith, cd.,
Research in Experimcnlal ilconomics. vol. Z (JAI Press. Greenwich. CT).
Rurncss. CI S.. R.C. Cummings. A.F. \lchr and b1.S. Walhcrl. 19X3. Valuing policies which
reduce cn~lronmcnfal risk, Natural Rrsourccs Journal 23. 676-682.
J. Afm et al.. Why do people pay taxes? 37
Cohn. Richard A.. Wilbur G. Lewellen, Ronald C. Lease and Gary G. Schlarbaum. 1975.
Individual investor risk aversion and investment portfolio composition, Journal of Finance
30. 60.~620.
Cowell. Frank A., 1985. Tax evasion with labour income. Journal of Public Economics 26.
19-34.
Cowell. Frank A.. 1990. Cheating the government: The economics of evasion (MIT Press.
Cambridge. MA).
Cowell. Frank A. and James P.F. Gordon. 1988, Unwillingness to pay: Tax evasion and public
good provision. Journal of Public Economics 36. 30%3?1.
Dubin. Jeffrey A. and Louis L. Wilde. 1988. An empirical analysis of federal income tan auditing
and compliance. National Tan Journal 41. no. I. 61-74.
Friedland. Nehemiah. 1982. A note on tar evasion as a function of the quality of information
abtiut the credibdity of threatened fmes: Some preliminary research. Journal of Applied
Social Psychology 12. 54-59.
Fnedland. Nehemiah. Shlomo Xlaital and Aryeh Rutenberg. 197X. A simulation study of income
tax evasion. Journal of Public Econom& IO. 107-l 16:
Friend. lrwm and %larshall E. Blume. 1975. The demand for risky assets. American Economic
Review 65. 90&92.
Graetz. Michael J. and Louis L. Wilde. 198.5. The economics of tax compliance: Fact and
fantasy. National Tax Journal 3X. no. 3. 355-363.
Internal Revenue Service. United States Department of the Treasury, 197X. A dictionary of
compliance factors (Government Printing Oflice. Washington. D.C.).
Isaac. R. Mark. Kenneth F. McCue and Charles R. Plott. 19X5. Public goods provision in an
experimental environment. Journal of Public Economics 26. 51-74.
Isaac. R. Mark. J.&l. Walker and S. Thomas. 19X4. Divergent evidence on free-riding: An
experimental examination of possible explanations. public Choice 43. I IZ-149.
Kahncman, Daniel and Amos Tversky. lY79. Prospect theory: An analysis of decision under
risk, Econometrica 47. 263-291.
Kim, 0. and M. Walker. IYHJ. The free rider problem: Experimental evidcncc. Public Choice 43.
3-24.
Kunrcuther. H.. R. Ginxiherg. L. IMiller. P. Sapi. P. Slavic, B. Borkan and N. Katz. 107X.
Disaster insurances protc&on: public pohcy lessons (John Wiley and Sons. Inc.. New York.
NY).
Machina. Mark J.. 19X.3. Gcneralixd expected utihty analysis and the nature of oh*erved
violations of the mdcpcndcnce axiom. in: Bcrnt Stigum and I‘rcd Wcnstop. cds.. I:ound:~trons
of utility and risk theory with applications (I). Rcidel. Dordrccht. Holland).
Machina. Mark 1.. IYX7. Choice under uncertainty: Problems solved and unsolved, Journal of
Economic f’crspcclives I, no. I. I21 -1%.
McClelland. Gary II.. William D. Schulle and Don L. Coursey. 19X6. Valuing risk: A
ctrmparison of expected utility with models from cognitive psychology. unpublished manus-
cript. University 0r Colorxlo. Boulder. CO.
Morin. Kogrr A. and A. I:ernandcr SuareL 1983. Risk aversion revisited, Journal of Finance 3X.
1201-1316.
Palfrey. T. and t1. Rosenthal. 19X-t. Participation and the provision of discrele public goods.
Journal of Public Economics 24, I71 -193.
Pencavel. John H.. IY79. A note on income tax evasion. labor supply, and nonlinear tax
schedules. Journal of Public Economic 12. 115-124.
Rapoport. Amnon. 1985. Provision of public goods and the MCS experimental paradigm.
American Political Science Review 7Y, I-IX-ISS.
Samuelson. Paul, 1954. The pure theory of public expenditure. Review of Economics and
Statistics 36, 3X7-390.
Sandmo. Agnar. 19X1. Income tax evasion. labor supply. and the equity-efficiency tradeoff,
Journal of Public Economics 16. 265-2xX.
Skinner, Jonathan and Joel Slemrod. IYYS. An economic perspective on tax evasion, National
Tax Journal 3X. no. 3. 345-353.
Smith, V. Kerry and William H. Desvouspes. IYX6. Asymmetries in the valuation of risk and the
siting of hazardous waste faciltties, American Economic Review Papers and Proceedings 76,
29 I -2Y.t.
38 J. Aim et al.. U’hy do people puy roes?
Spicer. Michael W. and Lee Becker. 19X0. Fiscal inequity and tax evasion: An experimental
approach. Nntlonul Trir Journal 33. no. 2. 171-175.
Spicer, Mwhuel W. and Rodney E. Hero. IYL(S. Tax evwon and heuristics. Journal of Public
Economics 26. 263-267.
Taylor. 51.. 1976. Anarchy and Cooperation (John Wdey and Sons. Fl;ew York. NY).
Yitzhahi. Sholomo. 1974. A note on income tax evasion: A theoretical analysis. Journal of Public
Econamlcs 3 ‘01 -‘O’.
._ __