|Gemmell, Norman, Ratto, Marisa|
When taxpayers are audited, two effects on their declared tax liabilities might be expected. First, where non-compliance is detected, there is a direct effect in the form of additional tax yield, including any penalties and interest, for the year of the inquiry. Second, there are possible indirect effects in the form of changes to future compliance behavior. These behavioral responses take two forms: (1) changed behavior by audited taxpayers themselves (sometimes labeled a "corrective" or "preventive" effect); and (2) spillovers to non-audited taxpayers, sometimes labeled a "deterrent" effect. This paper focuses on the former "preventive effect" on audited taxpayers and investigates whether
Based on unique set of
We also find that behavioral responses (both positive and negative) by small and medium businesses (who generally have greater opportunities to evade) were greater than those for personal taxpayers. Medium businesses also showed a greater (positive) response to an audit as the amount of uncovered tax evasion increased.
The result of negative preventive effects associated with taxpayers identified by random audits as "compliant" might suggest that it would be preferable to substitute more risk-based audits for the random audit program. The data available to assess the possible impact of replacing the
This paper investigates whether
Despite an extensive literature on tax evasion in general, the literature on how taxpayers' compliance behavior responds to auditing is more limited. In part this reflects researchers' lack of access to confidential taxpayer unit record data. However, Slemrod, Blumenthal, and Christian's (2001) investigation of individual taxpayers' responses to the threat of an audit stimulated a number of subsequent empirical and experimental studies that seek to identify how taxpayers' compliance behavior responds to various compliance regimes and social settings. Alm and McKee (2004, 2006), for example, use experimental methods to assess the relative merits of random versus risk-based audits, and to examine how individuals' compliance decisions respond to announcements regarding audited selection and to the "productivity" of an audit. (2) Nevertheless, as Alm, Jackson, and McKee (2009, p. 392-393) note, it remains the case that, "the ways by which taxpayers become aware about audit intensity and then respond to their assessments of enforcement efforts are 'open questions' (Plumley, 1996)." Alm, Jackson, and McKee (2009) address some of these "audit response" questions using experimental methods. Our objective here is to address similar questions using actual, rather than experimental, taxpayer data. This also allows us to address responses to the audit process that have only recently begun to be considered in the experimental literature, such as the impact of differences in the "quality" of audit outcomes on taxpayer responses.
Focusing on the preventive effects described above, we argue that the predicted responses by taxpayers faced with random auditing are likely to depend both on their expected, endogenous probability of audit, and on the expected or perceived "quality" of those audits (e.g., how much evasion they detect). Neither of these features has figured prominently in previous analyses. Based on unique set of
The remainder of this paper is organized as follows. Section II briefly reviews the recent theoretical and empirical literature on responses to taxpayer auditing. Section III considers how the experience of an audit is expected to influence taxpayers' future compliance by changing their perceptions of being audited and their expectations regarding the amount of discovered tax evasion. This motivates our empirical analysis in Section IV, which describes the data used and our empirical methodology. The empirical results are discussed in Section V. Section VI considers the relevance of these results for the choice between random and targeted audits, and some conclusions are drawn in Section VII.
In brief, our results demonstrate that the verdict of a previous audit–in particular whether a taxpayer is identified as "compliant" or "noncompliant" and the amount of undeclared tax yield identified–affects whether subsequent compliance increases or decreases and the extent of any increased compliance. Distinguishing empirically between so-called "compliant" and "noncompliant" audited taxpayers turns out to be crucial for correctly identifying their responses. In addition, responses seem to be larger for audited business taxpayers–who generally have greater opportunities to evade–than for individual taxpayers.
II. EXISTING LITERATURE ON BEHAVIORAL RESPONSES TO TAX AUDITS
In the standard portfolio model of tax evasion, an individual chooses the optimal amount of income to evade in order to maximize expected utility. The tax rate and the penalty rate are exogenously determined by the tax authority, and taken as given by the individual. The probability of detection is typically modelled as a fixed subjective probability, as perceived by the taxpayer. (3) It is assumed that the audit is 100 percent successful so that, if audited, an evading individual is caught and fined for the full amount of evasion.
As is well known, in this model an increase in the audit probability unambiguously decreases the amount of evaded income as it increases the expected cost of being caught. The assumption of a fixed probability of detection is, however, rather strong. In the original model, Allingham and Sandmo (1972) also consider the case where the probability of detection varies with declared income and confirm the result that an increase in the probability of detection decreases tax evasion.
However, as Alm, Jackson, and McKee (2009) acknowledge, in reality individuals are likely to form expectations based on a wide range of available information. In addition, audits are not always 100 percent successful in the sense that they may detect less than the full amount of evasion. This can be expected to affect audited taxpayers' perceptions of the "quality"–as well as the likelihood–of future audits.
Recent theoretical contributions have modelled taxpayers' perceptions of being audited as different from the true probability of audit. Kim (2005) considers a model in which taxpayers have a private signal about the true probability of being audited. The author analyzes the role of audit misperceptions on compliance behavior, and shows that an increase in the noise of the private signal about the audit has a positive effect on compliance behavior only when the true probability of detection falls below the cut-off equilibrium value that defines evaders and non-evaders. Hence if the true probability of audit is very low, compliance will be higher the greater the imprecision of the taxpayers' private signal. In Kim's model, the probability of an audit corresponds to the probability of detection with the audit 100 percent successful, as in the standard model. Snow and Warren (2007) extend the standard portfolio model of tax evasion to a two-period framework, where the taxpayer has some initial expectations about being audited and updates these expectations on the basis of past audit experience, following a Bayesian updating rule. In their model, being audited in the past increases the perceived probability of being audited in the future; however, when an audit occurs evasion is fully detected and the updating rule is unaffected.
Turning to the empirical literature, the available evidence on the effects of taxpayer audits shows contrasting results on the magnitude of their tax compliance impact. We can distinguish two main approaches: studies that consider the effects of experiencing an audit and studies that consider the effects of an increased probability, or threat, of an audit. Among the former, early work by Long and Schwartz (1987) and Erard (1992) using Tax Compliance Measurement Program (TCMP) data, found little effect of audits on future behavior. On the other hand, Dubin, Graetz, and Wilde (1990), Tauchen, Witte, and Beron (1993), Plumley (1996), and Dubin (2007), using time series cross-sectional information from individual tax returns aggregated at state level, found that audits had a large positive compliance effect with the indirect effects of an audit outweighing the direct effect. (4)
These recent findings based on individual data find some support from experimental evidence. For example, some laboratory studies in the 1980s showed that subjects being audited in earlier rounds of experiments increased their compliance in later rounds (
The Slemrod, Blumenthal, and Christian (200 l) analysis of the effects of a threat of an audit is one of only a few studies that examine actual taxpayer responses. With the cooperation of the
These various results challenge the simple view that a tax audit, or increased audit threat, necessarily raises a taxpayer's compliance in future. Individuals form perceptions not only about the probability of being audited but also, conditional on being audited, about the probability of being caught evading and the extent of detected evasion. If an audit does not fully identify evaded tax, then perceptions of being caught in future could be lowered, potentially reducing compliance. Slemrod, Blumenthal, and Christian (2001) show that audit threat responses differ across types of taxpayers and can affect taxpayers' perceptions about being audited in different ways, depending on their income and opportunities to evade. In our analysis below, where we examine observed responses to an actual audit, we similarly argue that distinguishing between alternative taxpayer subgroups is important, although we consider the case in which these subgroups differ in the official information revealed to them from previous audits.
This is consistent with Alm, Jackson, and McKee (2009), who use experimental approaches to examine how "official" information from the tax authority and "unofficial" information from other taxpayers affects compliance choices. In particular, they show that responses to post-audit information are conditioned by how well-informed taxpayers are regarding audit rates prior to the audit–a result that is broadly consistent with the evidence we present below.
III. HOWTHE AUDIT EXPERIENCE INFLUENCES FUTURE COMPLIANCE: OUR APPROACH
In this paper we evaluate the impact of an audit on future compliance, by treating taxpayers' behavior as a function of their perceptions of being audited and the amount of tax evasion they expect to be uncovered. Since taxpayers are not normally informed of the audit rule adopted by the tax authority, their ex-ante perceived probability of audit may not correspond to the actual probability of being audited. We also allow for the case where the tax authority may not (fully) discover the true amount of evasion, which is the private information of the taxpayer. Crucially, taxpayers' expectations of being audited and of the amount of discovered evasion are unlikely to be fixed, but rather will change in response to the audit experience, as we clarify below. Our approach, then, differs both from the standard portfolio model, where the subjective probability of audit is fixed and the audit is 100 percent successful, and from the more recent literature, which does not allow partial audit success. As we show, our prediction is that being audited in the past does not necessarily increase future compliance.
In our approach the perceived probability of being identified as noncompliant is the product of the perceived probability of being audited and the perceived conditional probability that evasion is detected by the auditor, once an audit takes place. We assume that individual taxpayers adapt their beliefs about these probabilities and detection rates in the light of experience. In the present context, given data constraints, and in line with the findings of Lochner (2007), we assume that the only information used to update perceptions arises from the individual's own audit experience. (5)
We begin by clarifying the process by which taxpayers' perceptions are updated after receiving an audit. The result of an audit will affect both the expected probability of a future audit and the expected probability that all future evasion is detected. In particular, the expected probability of being audited in the future increases if tax evasion has been discovered (even if only partially) by a previous audit. This is because those taxpayers are more likely to be added to a risk-based audit pool. (6) Alternatively, if no evasion is discovered, two cases arise: either there was indeed no evasion or the audit completely failed to uncover it. In either case, if the individual thinks that future random audits are conducted "with replacement" (that is, where the probability of being randomly selected in current or future periods is independent of past selection for audit), then there will be no change in the expectation of being audited in future.
However, if the individual thinks that future audits are conducted "without replacement" (such that being randomly selected for audit last period reduces the probability, possibly to zero, of being randomly selected subsequently), or if the audit completely failed to discover tax evasion, then the expected probability of being audited in the future will decrease. That is, zero detected evasion at time t-1 reduces the probability of being selected "randomly," or via a risk-based audit program, at time t.
For perceptions of audit success, when an audit discovers less than an individual expects (based on their private information), the perceived probability of success of the audit in the next period will decrease. The converse holds when an audit discovers more than an individual expects, for example where a taxpayer was confident, prior to audit, that a particular form of evaded tax had a low probability of discovery. Alternatively, the taxpayer may believe, prior to an audit, that once selected for audit x percent of evasion will be discovered. If this percentage is confirmed by the audit, the perceived probability of audit success remains unchanged.
Table 1 summarizes the possible effects of random audits. The first column represents the known outcome of a random audit at time t, which classifies taxpayers into two categories: compliant (zero evasion) and non-compliant (positive evasion). Consider the first row of the Table 1. If the taxpayer was truly compliant or was noncompliant but believes that the next period's audit sample is selected "with replacement," expectations about the probability of being audited in the next period will remain constant. If, instead, the taxpayer was actually noncompliant but escaped detection and believes random audits are with non-replacement, expectations of an audit at t + I will decrease. The expected detection rate (column 3) remains constant if the individual was truly compliant, but decreases if the individual was actually noncompliant.
The second row of the table shows taxpayers classified by audit at t as noncompliant. These might expect to be specifically targeted in the future, such that expectations of a t + 1 audit increase. However, if only random audits are conducted (with replacement), an unchanged probability of audit in the next period might be expected. The expected rate of detection rises or falls depending on whether the proportion of evaded income discovered by the audit is greater or less than the individual expected prior to audit at t.
The overall effect on future tax evasion can be seen in column 4, and depends on the combination of the effects on the expected audit probability and on the expected detection rate. For those previously classified as "compliant," future compliance is expected to fall or remain constant. For those classified as "noncompliant," compliance is expected to rise unless the audit identified less tax evasion than the taxpayer expected, and this is sufficient to counteract any increase in the probability of a future audit attributable to the use of risk-based audits.
In our empirical analysis we analyze the effects of random audits on different groups of taxpayers, using individual level data from
IV. THE UKTAXPAYER DATA AND EMPIRICAL METHODOLOGIES
In seeking to identify the preventive effects discussed in the previous section, our analysis needs to distinguish between those taxpayers labelled as "compliant" and "noncompliant" by the audit process and identify how this information changes different taxpayers' expectations of future detection (via both the likelihood and detection rate of future audits). We have data on the direct yield (additional tax raised plus penalties and interest) from random audit inquiries in the year 2000, separated into compliant (zero direct yield) and noncompliant (positive direct yield) on completion of the audit process. Clearly this only captures "true" (non-) compliance to the extent that the audit succeeds. We should note that this distinction between compliant and noncompliant as determined by the audit process may not be very accurate. For example, taxpayers with positive but low yield values might be considered "compliant" because the low yield reflects genuine compliance mistakes rather than intended evasion. With this view, in Section V.D, we consider whether the behavioral response to a random audit of taxpayers with larger yield differs from those with a lower yield. Inevitably, no similar information is available for the non-audited.
We also cannot observe directly how taxpayers' subjective probabilities of audit change in subsequent years. However, from our previous analysis of the expected audit probabilities and detection rates for those labelled as "compliant" and "noncompliant" by the 2000 audit, we can predict the direction of change in indirect yield as described above. In addition, since some taxpayers have to submit their subsequent years' tax returns after knowing their 2000 return has been randomly selected but before knowing the audit outcome, we may expect to observe different responses related to these differences in information available to the taxpayer.
A. Sample Selection
For the present analysis, we identified all those taxpayers within each segment who had been audited in 2000 via the random selection process. The preventive effects of an audit may take some time to be assimilated, and may depend on when the audit was completed. We therefore chose initially to compare declared tax in the three years after 2000 with the three years before. To minimize the likelihood that previous or future audits influence observed reactions to the 2000 audit, we excluded all taxpayers who were audited in these six years (1997-1999, 2001-2003). (9) We then randomly selected a sample of "control" taxpayers who were not audited in 2000 but otherwise matched the selection criteria of the audited taxpayers. (10) This yields a balanced panel where, for each household, we observe all six years of data (11) for declared tax, and total income as well as its components (e.g., self-employment income, employment income). (12) Two further distinctions within each taxpayer segment are relevant for our subsequent investigations: whether audited taxpayers were (1) identified as "compliant" or "noncompliant;" and (2) whether their tax returns for years after 2000 were submitted before or after the outcome of the 2000 audit was known. (13)
For (1), the "compliant" are defined as those audited taxpayers for whom the 2000 audit led to no additional direct yield; the "noncompliant" are those for whom positive additional yield was deemed liable (additional PIT or NICs, plus any interest/penalties) as a result of the audit. For (2), our data identify whether audited taxpayers submitted their 2001, 2002, or 2003 returns before or after the end of the 2000 inquiry. Table 2 shows the numbers of taxpayers in each sample (segment), and distinguishes the "compliant" from the "noncompliant."
Each segment consists of around 2,500-3,000 individuals, of which around 20-25 percent were audited. Within the audited sample, the ratios of compliant to noncompliant are quite different. For the S- and P-segments, there are relatively large numbers of compliant taxpayers (with only 136 (= 18 percent) noncompliant P-segment taxpayers out of a total of 776 audited, for example). The audited M-segment, however, reveals 55 percent (320/575) noncompliant.
Table 3 reports the year in which the 2000 audits were completed. This table shows that only 42 percent of inquiries had been completed by the time audited M-taxpayers submitted their 2001 tax returns. By 2003, completed inquiries had risen to 90 percent, with the remainder completed within the next two years. This pattern is noticeably different from the P-taxpayers, where almost 80 percent submitted their 2001 returns after knowing the outcome of the audit. The S-taxpayers reveal an intermediate case with 65 percent of audits complete in 2001.
The three segments also reveal differences in the time profile of their direct yield. Table 3 shows (an index of) the average amount of additional tax obtained from the audits that were completed in each year, by segment. To respect the need for confidentiality of sensitive information, we do not report the amount of direct yield obtained per noncompliant taxpayer but instead report an annual index for each segment ,where 2001 = 100. The table shows that for small and medium businesses, the longer running audits on average tended to be associated with higher direct yield (e.g., more than four times the average yield per (noncompliant) audit in 2004-2005 compared to 2001); however, the reverse is true for the P-segment. This suggests that it takes longer to identify the extent of tax evasion in businesses compared to individuals, with the "easy" (low evasion) cases generally completed first in the case of businesses. (14)
Finally, if both audited and control firms are selected truly randomly, any observed differences between these two groups prior to audit should be due to random factors only. For example, on average, income should be the same in both groups. In the next subsection, we report tests for possible non-random differences between groups. This issue is also relevant to our use of DID methods in Section IV to separate out audit impacts. Though our combined audited group was selected randomly, the compliant and noncompliant subgroups were not, with the split being determined by the various aspects of the audit process. These subgroups could therefore differ for systematic reasons from the control groups, even before the 2000 audits.
B. Empirical Methodologies
This subsection describes the application of DID methods to audited and control group data to measure the impact of auditing. It then considers how the approach is affected by separating the audited group into compliant and noncompliant subgroups. First, we test whether the declared tax (including NICs) of audited and control groups can be considered to differ for random reasons only, based on a t-test and [chi square]-test for group mean and median differences, respectively, and the Wilcoxon rank-sum test. For all three segments these suggest that, using a l0 percent confidence interval, the inter-group differences prior to 2000 are not statistically significant confirming that, to a reasonable level of confidence, the two group means can be considered the same prior to 2000. (15)
Test results for the M-segment (though not for the others) provide some evidence of significantly higher tax paid by the audited after 2000–even though our hypotheses propose that the compliant and noncompliant audited subgroups could deviate in either direction from the controls as a result of auditing. This evidence is suggestive of a possible preventive effect of auditing. However, the key tests are whether, ceteris paribus, the "after versus before" difference for the audit group exceeds that for the control group and whether the impact of auditing differs as hypothesized between the two compliant and noncompliant subgroups.
To examine this we use a DID regression, applied to the two three-year periods before and after 2000 (1997-1999 and 2001-2003). This regression takes the form
(1) [y.sub.it] =[[beta].sub.0] + [[beta].sub.1] [D.sub.]T + [[beta].sub.2] [D.sub.G] + [alpha][D.sub.G] [D.sub.T] + [[epsilon].sub.it],
where i = taxpayer; t = time; y = declared tax (including NICs); [D.sub.T] = dummy for time (= 0 before 2000; = 1 after); DG = dummy for an individual being in a group (1 = audited, 0 = control); and [[epsilon].sub.it] = random error term.
Interpretation of the coefficients is as follows: [[beta].sub.0] = expected average outcome for the controls before 2000; [[beta].sub.1] = difference in expected average outcome before/after 2000; [[beta].sub.2] = difference in expected average outcome between audited and controls before the audit; and [alpha] = difference in the change in the outcome for the audited group compared to the control group–the DID parameter measuring the effect of the intervention. That is, [??] = ([bar.y.sub.1,1] – [bar.y.sub.1,0]) – ([bar.y.sub.0,1] – [bar.y.sub.0,0], where [bar.y] is average declared tax, and where the first and second subscripts refer to [D.sub.T] and [D.sub.G] respectively.
To capture the impact on tax declared for each post-audit year separately, (1) can readily be rewritten. We use OLS regressions for this first part of the analysis and cluster at the taxpayer level. Ideally we would control for personal characteristics. As already mentioned, our dataset includes age, gender, and trade description; however, given that there is a substantial number of missing values for those characteristics, to avoid losing large numbers of observations we do not control for age, gender, or occupation. (16) We also decided not to include the components of total income separately in the regressions since these come from the same declaration. Income component data show that the shares of the main components of total income vary across the three sample segments and over the six years. These shares are also potentially affected by auditing and possibly in different ways for the "compliant" and "noncompliant." (17)
To examine whether the impact of random audits is different for the compliant and noncompliant audited groups we use a regression that is similar to (1) but allows for two audited groups. In this case the specification becomes
(2) [y.sub.it] = [[beta].sub.0] + [[beta].sub.1] + [D.sub.T] + [[alpha].sub.1] [D.sub.C][D.sub.T] + [[alpha].sub.2] [D.sub.NC] [D.sub.T] + [[gamma].sub.i] + [[epsilon].sub.it],
where now there are three groups (controls, compliant audited (C), and noncompliant audited (NC), and [D.sub.C] ([D.sub.NC]) is a dummy for being in the compliant audited (noncompliant audited) group. In this case [[alpha].sub.1] ([[alpha].sub.2]) measures the difference in the change in the mean declared tax between compliant audited (noncompliant audited) and all non-audited controls.
In the case of(l), where both groups (audited and controls) are selected randomly, we do not need to allow separately for differences in unobservable characteristics across groups, e.g., via individual fixed effects (FE) captured by [[gamma].sub.i] in (2). Random selection should ensure there are no systematic group differences. In (2), however, we include [[gamma].sub.i] since the two audited subgroups are not randomly selected, but rather are selected on the basis of known taxpayer characteristics (their compliance) identified by the audit. Within the control group we cannot distinguish the compliant from the noncompliant. However, we can allow the impact on each audited subgroup (compliant versus noncompliant) to differ compared to the all non-audited group. We therefore adopt the (individual) (18) FE estimator to allow for the possibility of any systematic differences between the three groups prior to audit. This specification allows the compliant audited and noncompliant audited to differ on observable and unobservable characteristics that are constant over time. It may be, for example, that the differences between a truly compliant and a noncompliant taxpayer are related to taxpayer attitudes or the use of a tax advisor. Effectively our FE estimator treats any such differences that affect compliance levels as constant over the time period we consider. This assumption would be violated if trends in the control and audited groups changed in different ways after 2000 for reasons unconnected with auditing. (19)
We also argued in Section III that the closure of the inquiry would give taxpayers information that might affect their expectations of future audit or the extent of future noncompliance detection. We therefore also examine whether there is evidence of more robust preventive effects of auditing when each taxpayer's "post-audit" declared tax is measured only from the tax returns submitted after inquiry closure. (20)
Finally, we consider whether the amount of discovered tax evasion affects the behavioral change. In order to test whether taxpayers charged with larger direct yield (arising from their discovered noncompliance) have a stronger response to the audit, we include an interaction term in (2) between the dummy identifying the noncompliant after the audit ([D.sub.NC] [D.sub.T]) and the amount of their direct yield.
V. DIFFERENCE-IN-DIFFERENCES RESULTS
This section first considers whether the randomly audited taxpayers as a whole differ from the random control group. We then examine how the outcomes differ when the audited are split into compliant and noncompliant, and when preventive effects are estimated using only "after audit closure" data. Next, we calculate the implied revenue gains/losses associated with the preventive effects identified. Finally, we examine how the amount of discovered tax evasion affects the behavioral responses of audited noncompliant taxpayers.
A. Audited versus Controls
Our hypotheses above suggested that, within the audited group, differently signed responses to audit might be expected depending on a number of characteristics, such that the net response across all audited taxpayers is ambiguous. We therefore began by using (1) to test whether the audited and control groups display a significant DID parameter, comparing the two periods 1997-1999 and 2001-2003 (in the top part of Table 4). For each of the three segments this produced values of [alpha] in (1) that are positive for the M-segment and negative for the S- and P-segments, though all three are insignificantly different from zero. (21) That is, changes in the audited group (as a whole) after audit were not statistically larger or smaller than those observed for the control group. This result is perhaps not surprising, given the different subgroup responses expected by the audited. However, together with the evidence on different subgroups that we present below, it serves to highlight the importance of testing audited subgroups separately to avoid conflating, and hence potentially failing to identify, these different responses.
B. Separating the Compliant and Noncompliant Audited
Results when the audited are split into "compliant" and "noncompliant" subgroups are shown in Table 4. The first set of DID parameters are obtained when comparing average declared tax in 2001-2003 with 1997-1999; the second set compare declared tax in the first year after inquiry closure with the average over 1997-1999. (22) Because many inquiries last for several years we have insufficient data to consider more than one year after closure.
In principle, the parameters from regressions of the form in (1) and (2) represent average additional declared tax in pounds per year. However, HMRC is understandably sensitive to public release of information on the amount of direct tax yield from individual audits. Hence, to preserve the confidentiality of yield data, those shown in Table 4 have been rescaled by a common, undisclosed factor. The t-ratios shown in the table are those obtained for the original parameters–hence standard errors cannot be inferred from Table 4. The relative size of the parameters in Table 4 therefore indicates the relative magnitudes of indirect effects by the various taxpayer categories; t-ratios indicate the statistical significance of the original parameters from zero.
The results in Table 4 are revealing. In all cases positive parameters are obtained for the noncompliant group and negative parameters are obtained for the compliant group. (23) Considering first the "before/after 2000" regression, for the P- and S-segments the negative parameters are larger and statistically significant. Unsurprisingly perhaps, the negative preventive effects for the compliant are less robustly estimated for medium businesses where they are a smaller fraction of the audited. The point estimates suggest a behavioral response amounting to a tax loss of around 140-190 units per compliant taxpayer per year. The reverse is the case for the positive preventive effects for the noncompliant where these are largest (342 units per year) and most robustly estimated for medium businesses.
Results are broadly similar for the "post inquiry closure" regressions with four of the six parameters now statistically significant including a larger negative (and significant) effect by compliant medium businesses. This is also now similar in magnitude to the positive effect of the noncompliant, amounting to around 18 percent of declared tax by these groups on average. This larger parameter for the compliant group may reflect a greater capacity on the part of medium businesses on average, compared to small businesses or individuals, to change their compliance behavior once an audit is complete. They are also likely to have greater confidence that the compliance label means that another near-future audit is unlikely. (24)
With this exception, the results in Table 4 do not indicate that the magnitudes of preventive effects are necessarily larger on average, once inquiries are closed. This may partly reflect the dominance in the sample of the many small business/personal taxpayers whose audits closed quickly in 2001 or 2002 (Table 3). Hence, for those taxpayers our first post-2000 observation is also often the first "post inquiry closure" observation. (25)
C. How Big Are Preventive Effects?
The implied "preventive yield" from random audits can be calculated based on the parameters in Table 4. We use the "one year after closure" results–since these are the most robustly estimated–together with data on the numbers of taxpayers in each category. These are shown in Table 5 as a percentage of each category's declared tax, on average, in 2000. The table also shows the preventive yield obtained using parameters plus/minus one standard error from the mean point estimate. (26)
This shows, for example, that for Personal taxpayers identified as compliant, declared tax fell due to audit by about 7 percent on average over 2001-2003. For the noncompliant Personal group, the estimates are less precise (because the parameters are less well identified), but these predict on average a similar increase in the preventive yield. For medium and small businesses, the positive preventive effects are much larger, at around 18-24 percent of declared tax, while the negative preventive effects on the compliant group are also large on average for medium businesses (-17 percent), and for small businesses (-11 percent, but with a wider margin of error).
D. Controlling for the Amount of Discovered Tax Evasion
Thus far we have defined a taxpayer as "compliant" if the 2000 audit led to zero yield. Clearly, however, a small amount of positive yield could indicate mistakes that both the taxpayer and the revenue authority recognize as such, rather than deliberate evasion. In this case, such low-yield taxpayers may be regarded as, and behave like, zero-yield "compliant" taxpayers.
Hence some positive amount of tax evasion disclosed by the audit may provide a better threshold by which to distinguish the behavioral responses of "compliant" and "noncompliant" taxpayers. Using the data on the amount of direct yield from the year 2000 audits, we test: (1) whether taxpayers with a "small positive yield" (defined below) behave similarly to zero-yield taxpayers; and (2) whether in general larger yield values (perhaps above some threshold) signal a greater degree of noncompliance and hence a greater behavioral response. We first introduce in (2) an interaction between the dummy variable identifying the noncompliant taxpayers (positive yield) after the audit with the amount of total yield recovered by the tax authority from these taxpayers.
We examine this for the case where taxpayers know the outcome of the inquiry (when their yield is also known). To explore possible "threshold" levels of direct yield, we first examine the quintiles of yield distribution and then more arbitrary ranges of direct yield based on monetary values. The results are shown in Table 6 for the M-segment. (27) Results in the top half of the table use the quintiles of the direct yield distribution. These show that the impact of an audit is negative on those taxpayers with an amount of direct yield in the first quintile of the distribution, although the results are not statistically significant, as well as a stronger and positive effect on the (more obviously) noncompliant whose direct yield puts them in the fourth and fifth quintiles of the distribution.
Having reviewed yield data, we also examined a number of more arbitrary threshold yield levels. The lower half of the table reports results for yield ranges in pounds: 1-150, 151-350, 351-1,000, 1,001-3,000, and >3,000 (the omitted range being from 1-150). The results indicate that for those noncompliant taxpayers charged with a direct yield below 150 [pounds sterling], declared tax in the year following the inquiry closure decreased, similarly to the compliant taxpayers, although the result is not statistically significant. But for those taxpayers charged with a direct yield above 1,000 [pounds sterling], declared tax increased, with a larger effect for those in the range above 3,000 [pounds sterling]. Hence, our point (1) above, on whether a small positive yield might induce a similar behavior to other zero-yield taxpayers, receives some limited support, but with results that are not statistically robust. On point (2) above–whether higher yield generates a greater positive compliance reaction–there is some support for the M-segment.
Similar estimates for the S- and P-segments were not statistically significant. This may not be surprising once we look at the distribution of direct yield by quintile for the three (M, P, S) segments in Table 7. This shows an index of direct yield with the lowest quintile (Q1) for the P-segment set equal to 1. (The level of yield for this group was sufficiently small that it likely reflects mistakes rather than deliberate evasion.) It is clear that the M-segment demonstrates considerably greater detected evasion (yield) across each quintile and in total. Moreover, especially for the S- and P-segments, there is little yield except from the fifth quintile, and perhaps the fourth quintile for the S-segment. So it is perhaps not surprising that there is little evidence of statistically robust yield effects for the S- and P-segments–especially since the numbers of taxpayers in each quintile for those segments are also lower.
In conclusion, we find positive changes in declared tax for those taxpayers who have been discovered under-declaring tax, and negative changes for those found to be compliant over the three segments. We interpret these changes as changes in compliance. In an appendix available from the authors, we look for any changes in trends in total income and the shares of self-employment, and employment, income components–the most important sources of income for the M- and S-segments. We find very similar patterns across the three segments and across the control, compliant audited, and noncompliant audited groups. Hence, we conclude that the differences we find in declared tax are not explained by differences in income trends.
VI. SHOULD THE
The evidence above suggests there may be significant negative preventive effects associated with taxpayers identified by random audits as "compliant," especially for individual taxpayers where a large fraction of the randomly selected are found to be compliant. This might suggest that it would be preferable to substitute more risk-based audits for the random audit program. This would both target the "noncompliant" better, hence raising revenue per audited taxpayer, and avoid providing information to the so-called "compliant" audited, which allows them to reduce their compliance. However, these arguments must be weighed against the potentially important role of indirect deterrent effects from the impact of auditing on non-audited taxpayers. Alm and McKee's (2004) experimental evidence, for example, established that an enforcement agency would be more effective by using a combined risk-based and random audit program, for given compliance resources.
Information for our
During the period covered by our empirical analysis in Section V, there were approximately 27 million income taxpayers, of whom around 9 million were SA taxpayers. Most of the remaining 18 million were PAYE taxpayers, considered by HMRC to be sufficiently low risk that they are not obliged to complete a SA tax return–they are expected to be compliant or engage in only minor noncompliance. Of the 9 million SA taxpayers, many would similarly be regarded as "low-risk," as they are required to complete an annual return simply because they are liable for the top marginal tax rate which is not applied at source to all income (e.g., bank deposit interest). As a benchmark, consider the case where half the SA sample (= 4.5 million) form the pool of taxpayers who would be targeted by HMRC's risk-based audit program as potentially "high-risk." This probably exaggerates the numbers of taxpayers with a high risk of noncompliance, since our random samples considered above reveal only around one-third of this group as underpaying tax and for many of those the level of their tax underpayment was small (Table 7).
To illustrate the magnitudes involved, these data imply that, for every 10,000
Additional indirect revenue effects potentially arise from two sources: greater compliance from those feeling a threat of audit from the risk-based program and less compliance from those who consider themselves unlikely targets of that program. For the latter case, 22.5 million "low-risk" taxpayers would now face a zero probability of being audited. Though we cannot know the previous probability of audit perceived by those taxpayers, if based on publicly available information on the ratio of audited to non-audited taxpayers, this probability could be estimated as between 0.022 percent (6,000/27 million) and 1.39 percent (375,000/27 million). However, a range of studies find that taxpayers often over-estimate the probability of audit (Alm, McClelland, and Schultz, 1992; Snow and Warren, 2007). With no random program, it seems plausible that such "low-risk" taxpayers would now perceive a zero audit threat and engage in (greater) noncompliance.
On the other hand, a compensating increase in revenue may arise from the increased probability of risk-based audit for the 4.5 million "high-risk" taxpayers. However this increase is negligible–rising from 1.37 percent (369,000/27 million) to, at most, 1.39 percent (375,000/27 million). That is, we estimate a rise in probability of 0.02 percentage points for 4.5 million "high-risk" taxpayers, as compared to a fall of 0.02 percentage points or more (to 0 percent) for the remaining 22.5 million "low-risk taxpayers. While we cannot be sure how these opposing influences would net out, there is a strong possibility of reduced overall compliance and revenue. Many so-called low-risk taxpayers are clearly not innately compliant, but likely comply in response to incentives, degrees of risk-aversion, etc. With a large proportion of such taxpayers in the
This paper has argued that a random audit process provides income taxpayers with information that can be expected to alter their perceptions regarding, and hence their behavioral responses to, tax audits. In particular, taxpayers' perceived probabilities of audit in the future and the predicted extent of detected evasion can be expected to change. Three key determinants of these perceptions are likely to be: (1) whether the verdict of a previous random audit was that the taxpayer is compliant or noncompliant, (2) whether random audits are known or perceived to be conducted with or without replacement; and (3) the existence of a targeted audit program.
Comparing samples of randomly selected audited and non-audited
We also found that behavioral responses–both positive and negative–by small and medium businesses were greater than those for personal taxpayers. This is consistent with the argument of Slemrod, Blumenthal, and Christian (2001) that different opportunities to evade are associated with differing levels of compliance, with business taxpayers in the
The data available to assess the possible impact of replacing the
We are grateful to
Allingham, Michael G., and Sandmo, Agnar, 1972. "Income Tax Evasion: A Theoretical Analysis."
Aim, James, and
Alm, James, and
Benjamini, Yael, and Shlomo Maital, 1985. "Optimal Tax Evasion and Optimal Tax Evasion Policy Behavioral Aspects." In Gaertner, Wulf, and
Dubin, Jeffrey, 2007. "Criminal Investigation Enforcement Activities and Taxpayer Noncompliance." Public Finance Review 35 (4), 500-529.
Erard, Brian, 1992. "The Influence of Tax Audits on Reporting Behavior." In Slemrod, Joel (ed.), Why People Pay Taxes: Tax Compliance and Enforcement, 95-114.
Kim, Youngse, 2005. "Audit Misperception, Tax Compliance and Optimal Uncertainty."
Kirchler, Erich, 2007. The Economic Psychology of Tax Behavior.
Lochner, Lance, 2007. "Individual Perceptions of the Criminal Justice System." American Economic Review 97 (1), 444-460.
Long, Susan B., and
Mittone, Luigi, 1999. "Psychological Constraints on Tax Evasion: An Experimental Approach." Unpublished Ph.D. dissertation.
Mittone, Luigi, 2006. "Dynamic Behavior in Tax Evasion: An Experimental Approach."
National Audit Office, 2000. Appropriation Accounts 1998-99: Report of the Comptroller and Auditor General. HC 25-XVI. National Audit Office,
Plumley, Alan H., 1996. "The Determinants of Individual Income Tax Compliance." Internal Revenue Service Publication 1916 (Rev. 11-96).
Sandmo, Agnar, 2005. "The Theory of Tax Evasion: A Retrospective View."
Snow, Arthur, and
Tauchen, Helen V.,
Webley, Paul, 1987. "Audit Probabilities and Tax Evasion in a Business Simulation." Economics Letters 25 (3), 267-270.
(1) Aim, Jackson, and McKee (2009) prefer the term "direct deterrent effect" to refer to effects on those who are audited, and "indirect deterrent effect" to refer to spillover effects on the non-audited.
(2) Their findings suggest that the announcement of audits increases the compliance rates of those who are told that they will be audited, but decreases the compliance rate of those who know, by elimination, that they will not. An increase in the announced audit effectiveness has a positive effect on compliance only for taxpayers whose audit has been announced.
(3) See Sandmo (2005) for an overview.
(4) For example, according to Plumley (1996), who uses a rich dataset across U.S. states for the years 1982-1991, the indirect effects of audits produce I 1 out of
(5) In the empirical analysis below of randomly selected "audited" and "control" groups, this is analogous to assuming that any other relevant characteristics are equally reflected in both random groups. Empirical studies have attempted to shed some light on the factors affecting offenders' beliefs of being caught. Lochner (2007), for example, examines which factors influence individuals' perceptions about the probability of arrest among young offenders and shows that while beliefs are largely unresponsive to most outside influences–such as demographic characteristics, neighborhood conditions, local official arrest rate–they do respond to an individual's own experience with crime and the police.
(6) At least in the
(7) Of course, all taxpayers, even if they are taxed at source via PAYE, are expected to complete a tax return if they believe they are likely to be liable for additional tax.
(8) Segment definitions are as follows. Personal (P, Individuals/Non-Business/Non-Complex): Self-assessed taxpayers, but with no income from trade, land, property, or partnership. Individuals have no directorships and total non-business income less than 100,000 [pound sterling]. Small Businesses (S): Self-assessed taxpayers with income from trade and land property but no recorded partnerships; total annual turnover from all trades and property must be under 15,000 [pounds sterling]. Medium Businesses (M): The same as small businesses, but with total annual turnover in the range 15,000 [pounds sterling]-250,000 [pounds sterling].
(9) A taxpayer audited again after 2000 may respond differently to the 2000 audit because audits can take a number of years to complete and some taxpayers' responses may be conditioned by audit outcomes, as discussed below.
(10) For example, controls were selected from identical taxpayer segments and, like the audited, were deselected if they had been audited in 1997-1999 or 2001-2003.
(11) We eliminated from the original data set those taxpayers who left SA over the period 1997-2003. This raised some concerns about attrition issues. However, those taxpayers who left SA did not seem to have left due to the treatment. For example, for the M-segment, we initially had a total of 3,715 taxpayers, of which 80 percent were in the control group and 20 percent in the treatment group. After eliminating those who left SA over the period, we were left with 2770 observations, of which 79.06 percent were in the control group and 20.94 percent in the treatment group. Hence, despite having lost 25 percent of the original sample, the relative size of the control and treatment groups stayed the same, suggesting that leaving SA was not related to being subject to an audit.
(12) The full set of income components includes: self-employment income, employment income, partnership income,
(13) We do not consider the 2000 return since data for 2000 includes any additional tax mandated by the audit and penalties imposed for under-declaration. These represent direct effects of the investigations.
(14) Note that, though we use the term "evasion" here, this includes some very small amounts of direct yield that may result from unintended errors on tax returns. These are discussed further below.
(15) These results are available from the authors.
(16) The most problematic variable was trade description, for which we had 2,754 missing values (out of 16,464 observations) for the M-segment, 7,440 missing values (out of 15,942 observations) for the S-segment and 17,304 missing values (out of 18,150 observations) for the P-segment.
(17) For example, some types of income are likely to be seen as easier to hide or, once noncompliance is identified, more likely to be checked by HMRC in the future (e.g., via being added to the risk-based pool). Hence, the audited might be expected to change the amounts they declare. Further, if an audit reveals that, for example, self-employment income in particular is evaded, future reporting of this component could change.
(18) We use individual fixed effects because we need to allow for the fact that individuals will have correlated errors with themselves in our panel dataset. If we considered group fixed effects, we would obtain biased standard errors.
(19) In an appendix to an earlier version of this paper (available from the authors), we explore some of these issues. Since this did not identify any non-audit sources of difference, we regard the FE model as reliably identifying audit effects.
(20) Taxpayers are likely to acquire differing amounts of information during the audit process regarding the final audit outcome, and they may adjust their expectations accordingly before closure. It is therefore unclear how far audit "closure" provides additional information.
(21) For example, for the M-segment, we find positive as for 2001-2003 ([alpha] = 128) and for each year individually when they are included separately ([alpha]01 = 42, [alpha]02 = 77, [alpha]03 = 261). However, t-ratios are 1.13 (for 2001-2003), and 0.33, 0.57, 1.62 (for 2001-2003 respectively).
(22) Where the data indicate that a taxpayer's audit closed during, say, 2001 but he/she submitted their 2001 tax return prior to that date, their 2002 tax return is treated as the first "after closure" year. For the period before the inquiry we use the mean of declared tax (including NICs) over the period 1997-1999. After 2000, for those who were audited we consider the first tax return following the inquiry closure. For example, in the M-segment, 42 percent of the 2001 returns were submitted after the inquiry closed, so for those cases the "year after" is 2001. Similar numbers for 2002 and 2003 are 41 percent and 7 percent, respectively. For the control group, we construct a weighted average of their 2001-2003 returns using as weights the same proportion of returns as for the audited (42 percent, 41 percent, and 7 percent).
(23) We dropped five observations for the S-segment (three from the audited compliant group, one from the audited noncompliant, and one from the control group), which experienced a drastic decrease in declared tax in 2002 and 2003. Close examination of the data revealed that these cases experienced a very large increase in employment income in 2001 (above 100,000 [pounds sterling]), which was followed by a very sharp decrease (greater than 80,000 [pounds sterling] a year) in the next two years. These changes were due to either a change in employment in 2001 or to bonuses received in 2001, neither of which seem likely to be audit-related.
(24) In view of the relatively large compliance costs imposed on audited businesses, HMRC is generally reluctant to audit such businesses again, when previously found to be compliant, should they be randomly selected again soon thereafter.
(25) We noted earlier that the expected probability of a future audit might be expected to decline immediately following an audit, and perhaps rise again subsequently if the experimental results of Maciejovsky, Kirchler, and Schwarzenberger (2007) apply here. However, we are limited in our ability to examine this issue using the 2000 audit data since we only have complete data on our samples to 2003 and many audits, especially for medium business, were not completed until then.
(26) Publicly available data on total amounts of direct tax yield from audits are given in NAO (2000) for the 1998-1999 tax year. They show that self assessment inquiries yielded 81 million [pounds sterling] from around 248,000 non-business taxpayers (average = 327 [pounds sterling]) and 98 million [pounds sterling] from around 121,000 business taxpayers (average = 808 [pounds sterling]).
(27) Results for the S- and P-segments were not statistically significant.
(28) In addition, eliminating random audits could result in the loss of any benefits they provide in fine-tuning the "red flags" that guide risk-based auditing.
Table 1 Effects of Random Audits on Expected Audit Probabilities and Detection Rates Expected Expected Category Audit Detection (from Random Audit) Probability Rate Compliance Compliant FALL FALL FALL (non-replacement) or or CONSTANT CONSTANT CONSTANT (replacement) Noncompliant RISE RISE RISE (targeted in future) or or CONSTANT FALL (FALL?) (replacement) Table 2 Sample Numbers [and Percentages] of Audited and Controls Sample Medium (M) Small (S) Personal (P) Control 2,169  2,002  2,249  Audited 575  655  776  of which: "compliant" 255 447 640 "noncompliant" 320 208 136 Total 2,744 2,657 3,015 Table 3 Case Closure and Average Direct Yields by Year Percent of Returns Submitted after Audit was Closed Year Medium Small Personal 2001 42 65 79 2002 83 96 98 2003 90 99 99 2004-2005 100 100 100 Index of Average Direct Yield Per Noncompliant Taxpayer (1) (2001 = 100) Year Medium Small Personal 2001 100 100 100 2002 168 176 52 2003 174 249 35 2004-2005 426 312 4 (1) This value is based on average yield per noncompliant audited taxpayer. Table 4 Difference-in-Differences Parameters Audited Subgroup Medium Small Personal "Before/after 2000" regression All audited: 128 -50 -115 (1.13) (-0.70) (-1.29) Noncompliant 342 *** 166 219 (2.67) (1.50) (1.44) Compliant -143 -151 * -187 ** (-1.01) (-1.90) (-2.40) Number of observations 16,464 15,942 18,150 "Before 2000/after inquiry closure" regression Noncompliant 330 *** 321 *** 170 (3.18) (3.37) (1.23) Compliant -312 *** -95 -182 *** (-2.95) (-1.41) (-2.60) Number of observations 16,110 15,900 18,132 Notes: Regression parameters in this table have been rescaled by a common, undisclosed factor to preserve confidentiality. The t-ratios shown in parentheses are the actual estimated values. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels. Table 5 Preventive Yield (As Percentage of Declared Tax) Audited Subgroup Medium Small Personal Total corrective yield (using "after inquiry closure" parameters) Noncompliant [+ or -] 18 (1) [+ or -] 24 (1) [+ or -] 5 Compliant [+ or -] 17 (1) [+ or -] 11 [+ or -] 7 (1) +/- 1 standard error Noncompliant [+ or -] 6 [+ or -] 7 [+ or -] 4 Compliant [+ or -] 6 [+ or -] 7 [+ or -] 2 (1) These values are based on DID parameters significantly different from zero. Table 6 Difference-in-Differences Parameters: Controlling for Direct Yield Medium Businesses Audited Subgroup Parameter "Before 2000/after inquiry closure" regression Compliant -121 *** -312 *** (-2.95) (-2.95) Noncompliant -109 -258 (0.50) (-0.91) Noncompliant*Direct Yield Q2 421 (1.36) Noncompliant*Direct Yield Q3 320 (1.03) Noncompliant*Direct Yield Q4 793 *** (2.57) Noncompliant*Direct Yield Q5 670 ** (2.16) Noncompliant*Direct Yield Range 15-350 543 [pounds sterling] (1.50) Noncompliant*Direct Yield Range 351-1,000 462 [pounds sterling] (1.40) Noncompliant*Direct Yield Range 1,001-3,000 869 *** [pounds sterling] (2.50) Noncompliant Direct Yield Range >3,000 968 *** [pounds sterling] (2.55) Number of observations 16,110 16,110 Notes: Regression parameters in this table have been rescaled as in Table 4. The t-ratios shown in parentheses are the actual estimated values. Q indicates quintile. Asterisks denote significance at the 1% (***) and 5% (**) levels. Table 7 Average Direct Yield by Segment Quintiles Segment Quintile Medium Small Personal Q1 9 4 1 Q2 25 10 5 Q3 57 19 12 Q4 139 46 28 Q5 473 216 255 All 141 58 60 Number of noncompliant taxpayers: 320 209 136 Notes: Values have been scaled such that Personal Q1 = 1. Average yield (per noncompliant taxpayer) for this group was sufficiently small that it likely reflects mistakes rather than deliberate evasion.
|Copyright:||(c) 2012 National Tax Association|