THE EFFECTS OF IRS AUDITS ON EITC CLAIMANTS. Jason DeBacker, Bradley T. Heim, Anh Tran, and Alexander Yuskavage

Transcription

1 THE EFFECTS OF IRS AUDITS ON EITC CLAIMANTS Jason DeBacker, Bradley T. Heim, Anh Tran, and Alexander Yuskavage The Internal Revenue Service (IRS) devotes substantial resources to audit tax returns of Earned Income Tax Credits (EITC) claimants, but little is known about the deterrence effect of these audits. Our paper examines the impact of this tax enforcement on subsequent individual taxpaying among those who claimed an EITC. Using evidence from randomized IRS audits during the period, we find that EITC participants who are audited show much larger increases in reported income in subsequent years, both compared to a control group of EITC filers, and compared to audited filers who were not EITC claimants. We find behavioral impacts on the extensive margin as well, with the probability of a filer claiming an EITC dropping by over 6 percentage points within 4 years following the audit, as well as changes in filing status and the reported number of dependents. Keywords: individual income tax, tax audit, tax evasion, tax avoidance JEL Classifications: H24, H26 Jason DeBacker: Darla Moore School of Business, University of South Carolina, Columbia, SC, USA (jason.debacker@moore.sc.edu.com) Bradley T. Heim: School of Public and Environmental Affairs, Indiana University, Bloomington, IN, USA (heimb@indiana.edu); Anh Tran: School of Public and Environmental Affairs, Indiana University, Bloomington, IN, USA (trananh@indiana.edu) Alexander Yuskavage: Office of Tax Analysis, U.S. Department of the Treasury, Washington, DC, USA (alexander.yuskavage@treasury.gov)

2 I. INTRODUCTION The Earned Income Tax Credit (EITC) is a major component of the antipoverty efforts in the United States, distributing an estimated $66.7 billion to 27.5 million families in However, EITC claims suffer from a substantial amount of noncompliance, with the Treasury Inspector General for Tax Administration estimating that between $13.3 billion and $15.6 billion in claims were issued improperly in fiscal year 2012, representing 21 to 25 percent of total payments. 2 IRS s main countermeasure is to expend a substantial amount of enforcement resources auditing returns that claim the EITC. While only 20 percent of 2014 tax filers claimed the EITC, EITC claimants represent more than a third of all audited returns. 3 Though these audits may recapture credits that were erroneously claimed in the tax year that was audited, we do not know whether such audits are effective in deterring subsequent non-compliance. To address this important policy question, our study investigates the impact of audits on EITC claimants longer-term taxpaying behavior, on both intensive and extensive margins. The EITC functions as a wage supplement for low income workers. Conditional on the number of dependents on the filer s return, the credit initially increases at the phase-in rate with a taxpayer s earned income, up to a maximum amount. The credit stays flat over 1 See In addition, between 1990 and 2006, fourteen states adopted state-level EITCs, bringing the total number of state EITCs to nineteen. State level EITCs generally are figured as a fraction of the federal EITC, with the fraction in 2006 ranging across states from 5 percent to 35 percent. 2 See In contrast, the rate of noncompliance from the income tax system as a whole is about 15%. 3 Though the rate of audits among EITC claimants is higher than for the full population, both rates are still quite low. For example, in 2015, about 1.7 percent of EITC claimants were audited, compared to 0.6 percent of non-eitc filers. See 2

3 an additional range of earnings, and beyond a certain amount of earnings (or adjusted gross income, whichever is higher) the credit declines at the phase-out rate with additional earnings until it is completely taxed away. Saez (2010) and Mortenson and Whitten (2015) both document evidence of bunching around the first kink the EITC schedule. This suggests that individuals manipulate income to locate near the maximum EITC benefit. While such manipulation may come from changes in real economic activity, the fact that this bunching exists only for filers with self-employment income suggests that the manipulation of income is through reporting changes. Mortenson and Whitten (2015) find increases in the degree of bunching at the first kink on the EITC schedule over time. Single filers without dependents receive a relatively small credit ($510 in 2017), but the credit increases substantially with more dependents (reaching a maximum of $6,318 in 2017 with three or more dependents). Given this structure, the EITC provides incentives for filers to misreport income, dependents, and possibly filing status (for example, to split income and dependents across two different returns). Compliance problems with the EITC are well documented. Blumenthal, Erard, and Ho (2005) finds substantial noncompliance among EITC claimants and that the rates of improper claims are not lower when professional tax preparers are used. McCubbin (2000) finds that noncompliance often arises through the misreporting of the number of dependents, likely because the EITC is more generous when there are more dependents in the filing unit. McCubbin (2000) also finds that the probability of misreporting the 3

4 number of dependents increases with the EITC benefit, with about 28 percent of the qualifying-child errors being driven by systematic misreporting. 4 Joulfaian and Rider (1996) study the role of marginal tax rates on tax compliance. They focus on the presence of large changes in marginal tax rates at different points on the EITC schedule and find that despite significant negative marginal rates in the phasein region and large marginal tax rates in the phase-out region of the schedule, there is little effect on reported income. One exception, consistent with the bunching evidence, is that filers with sole proprietorship income tend to underreport income to a greater extent. The availability of administrative data on audits and tax returns led to a handful of recent studies that examine the impact of audits on subsequent taxpaying in the field. 5 Though DeBacker et al. (2015) find that corporations in the U.S. tend to pay less tax following an audit, other studies that focus on individual taxpayers tend to find that taxpayers report more income and pay more tax following an audit. Kleven et al. (2011) find that audits and the threat of audits led to an increase in reported income in Denmark, and that the effect of audit on incomes subject to third-party information reporting is much smaller than on incomes that are not subject to third-party reporting. 6 Using U.S. tax data and randomized IRS audits, DeBacker et al. (2016) find similar results. Advani et al. (2015) find self-employed individuals in the U.K. also pay more tax after an audit. 4 Holtzblatt (1991), using data from the Tax Compliance Measurement Program (TCMP), also finds the majority of returns whose EITC amount of was disallowed uncovered misreporting in the number of qualifying children or marital status, but that misreporting of income was the source of the EITC disallowance in about 30 percent of the cases. However, one should note that the TCMP data that Holtzblatt et al. (1994) uses come from the mid-1980 s, before the IRS required filers to report the Social Security Numbers of dependents. 5 For a comprehensive review of recent work in the vast literature on tax compliance and the effects of audit see Slemrod (2016). 6 This is consistent with Pomeranz (2015) and Slemrod et al. (2015) who find increases in compliance with third-party reporting. Note that much, but not all, income that determines EITC eligibility is subject to third-party reporting. 4

5 In light of these studies, it is likely that the effects of audits of EITC recipients differ from audits performed on other taxpayers. EITC recipients face relatively large benefits from misreporting and have many margins across which they can misreport. A key finding of both Advani, et al. (2015) and DeBacker, et al. (2016) is that taxpayer reporting behavior depends on what the taxpayer believes the tax authority know about them. For EITC recipients, it will also depend on how quickly they believe information about income, filing status, and dependents becomes outdated. To test for the long-term effects of audits on tax reporting, we use data from the Internal Revenue Service s (IRS) National Research Program (NRP). The NRP piloted some random audits of individual tax filers starting in tax year 2001 and began conducting annual random audits starting in To these data, we merge returns from the universe of filers from 2000 to 2012, allowing us to examine the impact of audits on individual taxpaying behavior for a period of up to six years after an audit. 7 Importantly, we are able observe the reported incomes, tax payments, and credits claimed both of those who were, and those who were not, subject to audit, enabling us to compare the behavior of the audited treatment group to a suitably chosen control group. Since the treatment, an IRS audit, is randomized through the NRP, our empirical strategy is straightforward. Using the NRP sampling weights from the IRS, we construct a nationally representative sample of audited individuals who were EITC claimants in the year of audit. We pair this with a random sample of individuals drawn from the same 7 The IRS defines an audit as a review/examination of an organization s or individual s accounts and financial information to ensure information is being reported correctly, according to the tax laws, to verify the amount of tax reported is correct. (Internal Revenue Service (2014)). 5

6 population of tax filers, those who filed a return in that year and claimed an EITC. Then, we compare the tax filings of these two groups before and after the audit year. Our results show that the adjusted gross income of EITC recipients increase by over 6 percent following audit, with this increase persisting for at least 6 years. We also find changes along other margins, with the rate at which filers claim the EITC falling by 6 percentage points as a result of audit and the likelihood of reporting head of household status or multiple dependents dropping at similar rates. The paper proceeds as follows. Section 2 describes the reporting incentives created by the EITC, the audit process, and potential responses to audits. Section 3 describes the data. Section 4 provides detail on tax compliance rates amongst EITC claimants and, for comparison, all filers. Section 5 outlines our empirical strategy, and Section 6 presents results on the impact of audits on reported income following an audit. Section 7 provides a framework to explain the relatively large changes we see from EITC participants and gives some evidence to support this theory. Section 8 explores changes along the margins of dependents and filing status. We conclude our study in Section 9. II. REPORTING INCENTIVES AND AUDIT RESPONSES OF EITC PARTICIPANTS A. Income Reporting The incentives for income reporting vary across EITC claimants. In particular, one s incentives to over or under-report income depend upon where one s true income falls along the EITC schedule. Figure 1 shows the EITC schedules for filers who file as 6

7 single or head of household (solid line) and those who file are married (dashed line) and with various numbers of dependents. [Figure 1 about here] If a filer has at least one dependent and has an income amount that lies along the phase-in region, she has an incentive to over-report income. In this case, she faces a negative marginal tax rate (even accounting for FICA/SECA taxes) and can thus, with the EITC, increase after-tax income by reporting additional income, up to the first kink in the EITC schedule. Without dependents, the phase-in rate for the EITC is low enough that it would not produce these incentives because the individual income taxes and FICA/SECA taxes that may potentially be levied on the additional reported income would outweigh the amount of the credit. Filers on the plateau of the EITC schedule have an incentive to underreport income because reporting additional income will not increase their EITC and will trigger additional income and FICA/SECA tax liabilities. The incentive would be to report income to maximize the credit and no more i.e., to report income right at the first kink in the EITC schedule. Filers on the phase out region also have an incentive to underreport their true income, as reporting more income would lower their EITC amount and increase other tax liabilities. Note that while taxpayers with dependents maximize after-tax income by reporting income right at the first kink of the EITC schedule, we should not necessarily expect a large fraction of all taxpayers to locate here. First, most taxpayers are compliant and report their true income regardless of the benefits from noncompliance. That is, they do not have a willingness to be noncompliant. Second, the incentives to under- or over- 7

8 report income may be irrelevant if taxpayers do not fully comprehend the EITC schedule. In the case of limited information or understanding, EITC claimants may follow a simple rule of thumb to under-report income, regardless of where they are on the EITC schedule. Third, ability to control earned income over the entire tax year is beyond the reach of many filers. Hence, the manipulation of income on a filer s return is most likely driven by reporting, rather than real economic, responses. Fourth, much of the earned income reported by taxpayers is documented on 3 rd party information returns and so would be difficult to misreport and avoid detection. Finally, if the taxpayer is audited, the penalties imposed are increasing in the amount of non-compliance. B. Reporting of Dependents and Filing Status In addition to misreporting income, there are other margins over which EITC claimants can affect their tax liability. One of these in the number of dependents reported. By reporting additional dependents (up to 3), one can increase the size of their credit. As Figure 1 shows, the schedules for filers with more dependents (up to 3) lie at or above the schedule for filers with fewer dependents. However, the largest jumps between credit amounts come when moving from 0 to 1 dependents and from 1 to 2 dependents. Another way, beyond income, to affect the credit amount is through one s reporting of filing status. As shown in Figure 1, married filing jointly filers have the same phase-in region as single or head-of-household filers but face a longer plateau region. The incentives to file jointly are dependent on the amounts of income earned by each household member, however married households typically achieve the lowest total tax liability by splitting their return into two returns that are filed as head-of-household. 8

9 Consider the case of married couple filing in tax year 2012 (to be consistent with Figure 1) with 2 dependents and $20,000 wages ($10,000 from each parent). If they file a joint tax return, their EITC amount is $5,236. If they misreport their marital status and both parents file as head-of-household with one dependent each, the total credit for the household is the sum of the $3,169 credit each parent receives, or $6,338 for the household. Consistent with this, Holtzblatt (1991) reports evidence that most changes to filing status amongst EITC claimants following audit were changes from filing statuses for which they qualified for the EITC (married filing or head of household) to statuses where they did not (married filing separately or single). 8 C, Audit Process and Responses Those selected for an NRP audit have three potential audit experiences. Their return can be accepted as filed; they may receive a correspondence audit, communicating with the IRS via telephone or mail; or, they might have an in-person audit, meeting with the IRS examiner. These last two experiences, the correspondence audit and in-person audit, are both intensive, where the examiner will thoroughly consider all items on the filers return, checking for accuracy and documentation on the income and expenses listed. Among taxpayers selected for an NRP audit, the vast majority received either a 8 Allowing for heterogeneity in the responses to audit according to who would benefit from filing jointly versus separately (either as head of household or single) across the population would be an interesting extension and in line with the vast literature on the marriage penalty (see, for example, Whittington and Alm (2001)) that finds significant variation in the penalty across the filing population. However, our data do not allow us to do such an analysis. In particular, we are only able to identify couples if they married filing jointly. Thus, for single individuals we cannot construct the counterfactual of how well off they would be filing jointly because we can t identify who the partner would be if filing jointly. 9

10 correspondence audit or an in-person audit. 9 While selection into the NRP is random, selection into the type of audit experienced is not, which we address in our empirical section. For the average audit, the filer is notified about her return being selected within eight months of filing and then notified of the audit results eight to 12 months after being notified of the audit. Thus, for an average length audit process, the filer knows that she is under audit when filing her next return, but is unlikely to know the audit results before that next return is filed 10. The audit experience results in a flood of new information to the filer. As a result of being selected into audit, filers will likely update their probability of being selected into future audits. From the audit process itself, filers learn about proper reporting of items on their return and about how stringent the IRS is in uncovering true tax liability. Theoretically, filers could respond to an audit by becoming either more or less compliant. They may become more compliant if they either perceive their audit risk as having increased after being selected for an audit, or if they think that increased compliance would decrease their chance of being selected again. However, they might learn through the audit process how to better evade income taxes, or they might perceive their probability of audit as declining if they have recently experienced an audit (see Kastlunger et al. (2009)), which might lead to decreased compliance. Thus, given that theoretical models of the responses to audit are ambiguous, we form a hypothesis based 9 Among all filers in our sample, 93 percent of those audited are contacted either in person or via correspondence. In the sample with AGI under $40,000, it is 86 percent, and among EITC filers it is 98 percent. After weights are applied, these amounts are 77 percent, 67 percent, and 98 percent respectively. 10 The timeline for the length of audits can be gleaned from the date the audit was opened and closed in our data. The timing of notification came from discussions with IRS personnel. 10

11 on the empirical literature. Given past research on responses to audit (DeBacker et al. (2017), Advani et al. (2017), Kleven et al. (2011)) we expect that individuals will respond to audits by becoming more compliant. Thus, in the case of EITC claimants who underreport their true income, we expect to see reported income increase following an audit, while claimants who over-report true income might decrease reported income. We also expect that EITC claimants might decrease the number of dependents claimed after audit and may change their filing status. We should note that the changes in reporting following audit may not reflect malicious non-compliance prior to audit. Filing an income tax return in the United States is complicated and the EITC rules add further complexity. Given the limited resources that many EITC claimants have, complying perfectly with the tax code might be difficult. In a model where compliance takes effort on behalf of the taxpayer, Phillips and Plumley (2015) show that taxpayers will only choose to dedicate more resources towards understanding how to file a return correctly if it is likely to reduce their tax liability. Thus, the costs associated with complying with EITC rules may lead individuals to forego effort that would increase the accuracy of their return while also increasing their tax liability. This non-compliance will result in an understated tax liability, but it may be incorrect to interpret this behavior as malicious non-compliance, since it reflects a misunderstanding of the tax code. However, the responses to audit will look similar. Following audit, the taxpayer will report in a way that increases tax liability as they, through the audit, learn how to better comply with the tax system. In sum, we expect reported income (and other misreported items) to be corrected following audit, with an increase in reported income most likely driven by increased compliance rates. 11

12 III. DATA We construct our estimation sample with data from three sources. First, we use data from the waves of the IRS s National Research Program (NRP). 11 The NRP conducts audits on a stratified, random sample of the filing population. Included in these data are taxpayer identifiers (the social security number (SSN) of the primary filer), the year of the audited return, and the resulting adjustment to the tax return by line on the Form Each of the NRP waves have approximately 15,000 observations. Since the NRP provides the sampling weights used, we are able to aggregate our treatment group up to a nationally representative level. 12 Our second source of data is the IRS s Audit Information Management System (AIMS). The AIMS data contain detailed information on all IRS audits (including NRP and non-nrp audits) from 1996 to present and allows us to observe the date the audit began and ended. To pair with this group of audited taxpayers, we create a control group for each of the NRP waves by randomly selecting a sample of filers from the IRS s population of individual income tax returns. To do so, we randomly pick SSN endings for each year from We then select all primary filers who had one of these endings from the universe of returns filed that year. We only retain those returns which meet the sample restrictions used by the IRS for selecting NRP audits. 14 By using the weights 11 Note that we exclude the NRP wave from Documentation suggests that the sampling frame and intent of the 2001 wave sufficiently different from later waves to treat them as comparable. 12 See Leibel (2014) for an in-depth look at the details of this data. 13 The sample size is dictated by computational constraints. 14 Most of the selection criteria are administrative in nature, such as the date when the return was filed. 12

13 implied by the sampling frequencies of our control group, we are able to aggregate the control group up to both a nationally representative sample that is also directly comparable to the weighted treatment group. Using the SSN of the primary filer, for each taxpayer in our treatment and control groups for the four years , we draw all returns they filed from 2000 through 2012 from the population of individual income tax returns in the IRS s Individual Return Transaction File (IRTF). These data include all items on the front page of Form 1040 and the main line items from most associated schedules. Our panel datasets are thus comprised of a control group of randomly selected non-audited filers from the years of the NRP waves and a treatment group of randomly selected audited filers from the NRP waves. 15 By creating the control group in a way that mirrors the selection of the treatment group, we can allow for natural rates of attrition across both treatment and control. 16 While some members of our samples do leave the population of filers over time, these rates are consistent for treatment and control. Taxpayers who leave the sample for any reason are able to re-enter it in any years a return is filed. Our primary aim in this paper is to estimate the impact of audits on EITC claimants. To put those responses in context, we compare the response of the EITC claimant population to two other groups. First, in order to test the hypothesis that EITC claimants respond differently to audit than other taxpayers, we compare the results from 15 In other words, to be in the 2006 EITC recipient sample, the taxpayer had to either be included in the 2006 wave of the NRP and claim an EITC in 2006 or had to be drawn into the 2006 control group and claimed an EITC in Some members of both the treatment and control groups were subjects of operational audits at some point during our sample period. Although they may have been affected by that audit process, both treatment and controls should have been affected in a similar manner, and so the effect should difference out. Further, omitting such individuals would result in a select sample that was no longer be nationally representative. 13

14 EITC claimants the entire tax filing population. However, since claimants must have income below a threshold in order to qualify for the credit, differences in behavior between EITC claimants and the entire population may reflect differential response by income level, rather than differential responses by whether the EITC was claimed. Thus, we extract three different samples from our data. The first sample is our main sample of interest, consisting of filers who claimed the EITC. 17 The second sample is randomly drawn from the entire population of taxpayers. 18 Finally, the third sample is comprised of taxpayers with Adjusted Gross Income (AGI) of at least -$1,000 and no more than $40, Although we have detailed information on the characteristics of the audit and the adjustments to tax returns following audit for audits that were closed by the time we pull data from the AIMS database, we lack information on audits that were not closed by October However, given that our last NRP wave is from 2009 and that well over 95 percent of audits are closed within two years, almost all audits have been closed. Table 1 summarizes our sample, noting weighted and unweighted observations in the base year (i.e., NRP wave year) and across all years The NRP Sample represents our treatment group and the Random Sample our control group. 17 That is, for a filer to be in the EITC panel, she must have claimed the EITC in the year of the return that was selected for the NRP audit (if in the treatment group) or claimed the EITC in the year of the return that was selected to be a member of the control group. The taxpayer may or may not claim the EITC in other years of the panel. 18 The control group was drawn as a 0.1 percent sample by picking ten four-digit SSN endings for each wave. 19 The control group for the first and third samples was drawn as a 0.2 percent sample by picking two fourdigit SSN endings for each wave. 20 We use weights for both our randomly sampled control group and the treatment group. We weight the control group by giving each filer equal weight to sum to the total population of filers in the base year. We weight the treatment groups using the NRP sampling weights. This gives us a number of weighted observations approximately equal to the population of filer in the base year for the NRP sample. We then apply these weights to the filing units for each year they are in the panel. 14

15 [Table 1 about here] While we do observe the date an audit was opened and closed, we do not know when the filer was notified of the audit or the results of the audit. Thus, we use as our timing convention the number of years since the audited return was filed. For example, filers selected in the 2006 NRP wave had their tax year 2006 return audited, which we consider their base year of treatment. We then count tax year 2007 return as being one year since the audited return was filed. As a result, one would not expect a sharp increase in reported income for all filers in a given NRP wave in the first year since audit, since the duration of audits and the time when filers were notified varies. However, since the vast majority of audits are closed within two years, we do expect the effects of audits to fully materialize two to three years after the audited tax year. Because selection into the type of audit experienced is not random, we do not distinguish between these types of audits in our results. Although the different audit experiences might give rise to different responses, because the audit experience is endogenous (for example, those whose returns are accepted as filed may be those who have simple returns with easily verifiable income), we can t estimate causal effects of the different experiences. 21 Rather, the treatment effect we estimate is an intent to audit effect. IV. CHARACTERISTICS OF EITC FILERS In order to validate our identification strategy, it is important to first show that our treatment and control groups are indeed comparable. Table 2 presents summary statistics, 15

16 comparing characteristics of those in the treatment group to those in the control group. We present statistics for our three samples: our primary sample of filers who were EITC claimants in the base year (i.e., the year of the NRP audit or the year from which the control group was determined), and our two comparison samples of all filers and filers with adjusted gross incomes (AGI) less than $40,000. We compute these statistics in the base year so as not to be affected by the treatment (i.e., the NRP audit). The differences in the means of variables across the treatment and control groups are small, as would be expected given the stratified random sampling done by the IRS for the NRP audit (the treatment group) and our random sampling to construct the control group. [Table 2 about here] Mean AGI in the EITC sample is $15,200, which is lower than our two other samples, but rates of reporting wage income and self-employment income (on Schedule C) are higher for the EITC sample. Looking down the table, we see that EITC filers are younger on average than those in the other two samples, which taken together with the income differences noted above, suggests that the group of lower income filers as a whole is disproportionately older filers with more retirement or capital income. EITC filers are also more likely to use a tax preparer, file as married or head of household, and report more than twice as many dependents as the low-income group as a whole. About 69 percent of EITC claimants use a paid preparer. Any responses to audit we measure will reflect the extent to which they also work with paid preparers when responding to the IRS. As noted, the IRS has had a particular interest in the compliance of filers with the EITC program. Table 3 highlights compliance rates across samples and across EITC 16

17 regions. The fraction non-zero are the number of returns that have a non-zero entry for that item. For example, 45.9 percent of EITC filers have some adjustment to their tax liability after audit and the mean adjustment amount is $4,848. What we find are that compliance rates among EITC claimants are lower than for the full population of filers; 45.9 percent of EITC claimants have an adjustment to tax liability after audit as compared to 43.5 percent of all filers. The size of adjustments is smaller for EITC claimants, $4,848, than for all filers, $5,561l. However, comparing EITC claimants to other low-income filers, we find that the EITC claimants rates of adjustment and underreported income are higher (45.9 percent and 40.2 percent compared to 38.8 percent and 32.2 percent, respectively), and amounts of average audit adjustments are about 15 percent higher as well. On average, about 47 percent of EITC claimants who are audited have some adjustment to their credit amount, with an audit overturning about $1,000 in EITC on average. However, the variation is large in the adjustment to the EITC amount, as seen in the standard deviations of the adjustments in Table 3. Some of the audited filers receive a positive adjustment to their credit amount upon audit. When comparing EITC claimants at different points on the EITC schedule, we find that those on the phase-out region have the highest levels of compliance. Rates of over-reporting and under-reporting income (conditional on an audit adjustment) are similar across the three regions, but the amount of under-reported and over-reported income tends to be greatest in the plateau region. These results do not match up entirely with the incentives outlined in Section 2, which suggested that those on the phase-in region are most likely to overreport income and those on the phase-out region are the most likely to under-report income. We find the rates of under-reporting and over- 17

18 reporting are highest on the plateau at 41.7 percent and 6.5 percent, compared to overall rates of 40.2 percent and 5.8 percent respectively. However, note that the incentives to over- or under- report income depend on the taxpayer s true income, not what is observed by the IRS. If taxpayers are responding rationally, then we might expect taxpayers whose true income is on the phase-in or phase-out regions to report an income that puts them on or near the plateau. [Table 3 about here] Table 4 presents a modified version of Table 3 which uses the EITC region as determined by the NRP audit (i.e., true location on the EITC schedule), as opposed to the reported region. Compared to Table 3, we now see stronger evidence of differential compliance behaviors across the three regions. In all regions, under-reporting income is still the dominant type of adjustment, but now the phase-out region has both the highest rate of under-reporting income and the largest magnitude understatements. Conversely, the phase-in region has the highest rate and magnitude of overstatement. This suggests that while not all EITC claimants are responding solely to EITC incentives, there is at least some response to them. It is also important to note that the sample in Table 4 is different than in Table 3. Almost a quarter of taxpayers initially claiming the EITC were found ineligible for some reason, with the largest driver involving a reduction in the number of qualified dependents. This is in line with other work that finds significant misreporting of dependents. In the rest of the paper, when we discuss taxpayers by their EITC region, we rely on their reported region. We do this for two reasons. First, we are only able to identify a measure of true income for filers who were audited and not for the control group. This 18

19 places limits on our ability to construct control groups or to follow taxpayers over time. Second, and more importantly, the reported region is what is observed by the tax authority. Inferences which rely on true income are of limited usefulness in helping tax authorities develop policies or procedures. However, knowing the distribution of noncompliance conditional on observed characteristics allows for better use of scarce auditing resources. [Table 4 about here] V. EMPIRICAL STRATEGY Our empirical approach first identifies the impact of audits on EITC claimants using fixed-effects regression models, which use within filer variation for precise estimates of the effects of audit. We then examine the persistent effects of the audit by including a series of indicator variables for each year around the audit. These regression models allow us to control for both unobserved individual characteristics and economywide annual variation which a simple difference-in-difference approach is unable to address. 22 Using a panel of tax returns, we examine changes in individuals behavior after an audit while controlling for time-invariant unobserved individual characteristics. We first estimate an equation of the form Income '( = βpostaudit '( + γ ' + η (,6 + ε '(, (1) 22 In theory, the fact that we randomly select members of our treatment and control groups should account for unobserved individual characteristics. In practice, there are minor deviations from random selection into the NRP made in special cases. We therefore prefer the fixed effects specification of our regression. 19

20 where Income '( denotes adjusted gross income (AGI) for individual (taxpayer) i in year t; PostAuditit equals one if the observation comes from a member of the treatment group in a year after they were audited, and γ ' denotes an individual (taxpayer) fixed effect. The parameter η (,6 denotes a year-wave fixed effect. Rather than a dummy variable for each calendar year, we include a dummy for each year relative to the base year (i.e., the wave of the NRP the filer was selected or the NRP wave they are a control group for). Thus, we have dummies for 9 years before audit (observations from year 2000 for those audited in 2009) through 6 years after audit (observations from 2012 for those audited in 2006). 23 This is necessary because of how we construct our samples. Typically, year dummies are included to control for issues that affect all observations in a given year equally. However, note that in order to be selected into our EITC or AGI < $40,000 samples (or into various NRP audit categories), a taxpayer has to have sufficiently low income in their base year. If there is any form of cyclicality to incomes (such as regression to the mean), then the dominant issues in common are not necessarily between households in the same year but rather households at the same point in the cycle. 24 In this specification, identification of the effects of audit come from within-filer changes in reported income between the pre- and post- audit periods, net of trends in income common across the treatment and control groups (which are picked up by the year-wave fixed effects). We then examine whether the effects of audit differ with the number of years since the audited tax year. To do so, we estimate equations of the form: 23 This is a slightly more restrictive specification than simply allowing each wave to have its own year dummies. Our specification as presented restricts all fixed effects for e.g. 3 years after the audit to be the same across waves. Relaxing this restriction has almost no effect on our main findings. 24 Regressions of each wave run separately support the case for the cyclically-defined year dummies. 20

21 Income '( = ; :<= β : (PostAudit '( ) (k Years Since Audit) + γ ' + η (,6 + ε '( (2) In this specification, the key explanatory variables are a series of dummies that show the difference between the audited and control group from Year 1 through (at most) Year 6 after the audited tax year. We also include coefficients on each of the two years prior to audit to test for any pre-trend differences. VI. EFFECTS OF AUDIT ON REPORTED INCOME Among the EITC filers, we see that reported income goes up on average after audit for filers in all three regions. In Table 5 we present the results of estimating Equation (1), using within-filer variation to identify the effect of an audit, after netting out trends common to the treatment and control groups. We see that the common trends assumption holds; the coefficients on the dummy variables for years prior to audit are statistically insignificant. The coefficient on the post audit period is positive and strongly significant for the EITC claimant sample, with a change in AGI of over $700. Comparing the magnitude of the effect of audit on AGI across these groups, we find that the estimated effect among EITC claimants is slightly larger than that for all filers ($691) and is substantially larger than that for low-income filers ($406). These responses are economically significant. Consider that average AGI for an EITC claimants in the NRP is $15,253. Thus, the change in AGI is 4.8 percent for these tax units. 21

22 When comparing the effects of audit across EITC claimants in this model, we see positive point estimates on the post-audit indicator variable for all groups, although it is not significant for those filers on the phase-in region. 25 [Table 5 about here] Table 6 presents estimates of the effects of audit over time on AGI in each year before and after audit, using the fixed effect regression model in Equation (2). 26 The pretrends again show that the common trends assumption is satisfied when using wave-year interaction terms and fixed effects. In the first year after audit, EITC claimants income increases $564, relative to claimants who were not audited. The increase in AGI rises to $874 in the third year after audit and remains high, at $1,013, six years after audit. These represent substantial and persistent effects on EITC claimants, who have a mean AGI of just over $15,000 in the year they were audited. The qualitative result that EITC filers are much more responsive than low income filers generally is reinforced in these models. In fact, the point estimates for most of the post-audit years show that EITC filers respond to audit by increasing reported income by almost twice as much as the group of filers with AGI less than $40,000 (which includes EITC filers, among other filers). For example, three years after audit EITC claimants AGI increases by $874, while the increase three years after audit for the low-income filers is $494. The effects on EITC claimants also appears more persistent than the effects on low-income filers as a whole. In the fifth and 25 The 2009 American Recovery and Reinvestment Act added additional income thresholds and credit amounts for filers with three or more qualified dependents. We ran an alternative specification of the models in Table 5 that included an interaction between post-2009 and the post-audit indicator variables, and the interaction was insignificant, suggesting no difference between the overall audit effect results and those after this policy change. 26 Note that in Table 6, we winsorize income variables at the 99 percent level for the all filers sample. The IRS data we use are unedited and several large outliers (which could be data entry errors on the part of the tax filer) introduced noise into this sample. 22

23 sixth years since audit, EITC claimants income is $767 and $1,013 higher than their preaudit levels. In comparison, for all filers with AGI less than $40,000, the point estimates on the increase in reported are not significantly different from zero. The results in Table 6 also show that the responses of EITC filers as a whole are almost entirely driven by those whose reported income lies on the plateau and phase-out regions. These filers report an average of more than $900 in additional income in the first five years after audit, relative to the control group of filers with reported income on the same ranges of the EITC schedule in the audit year who did not get audited. Filers on the phase-in region show no statistically significant changes in reported income in any of the post-audit years. Because we do not see EITC claimants on the phase-in region behaving in a significantly different manner following audit, we proceed with our analysis looking at all EITC claimants together. [Table 6 about here] Although EITC claimants are not much different in terms of age or income from the group of filers with AGI less than $40,000 (see Table 2), their changes in reported AGI following audit are 80 percent larger ($733 compared to $406, as reported in Table 5). We propose that these large responses are related to the distinct incentives of the EITC program and the large impact that audits can have under such conditions. We develop this further in the next section. After having seen the relatively large responses to AGI, and along several other margins, we now consider how different income sources change after audit. Kleven et al. (2011) present a model and provide empirical evidence that the ability of a taxpayer to change her reported income depends on factors such as third-party reporting. DeBacker et 23

24 al. (2016) have documented substantial differences in the magnitude and trends of the changes across different income sources post audit. They find large, but more transitory, effects on income less subject to third party reporting, such as self-employment income and more muted, but more persistent, effects on other income sources such as wages and salaries. [Figure 2 about here] Figure 2 presents the estimated coefficients on the year relative to audit indicator variables from our fixed effects model described in Equation (2) on the sample of EITC claimants, where we allow for differential effects of audit for different types of income. We estimate the effect of audits on wages and salaries (from line 7 of the Form 1040 and which is subject to third-party reporting on the W-2) and self-employment income(as reported on line 12 of Form 1040, which generally has no third-party reporting) and compare these effects to the overall effect on AGI. We find that wage income is much less responsive to an audit than AGI and, when considered in terms of percentages of mean income, Schedule C income. That is, reported amounts of AGI and sole proprietorship income increase more sharply than wage income following audit. The difference in these responses is to be expected given the withholding and third-party documentation associated with wage income. The transitory effects of audit on sole proprietorship income are less intuitive, but the pattern is consistent with that found in broader populations of filers by Advani et al. (2016) and DeBacker et al. (2016). VII. THE EFFECT OF AUDITS IN THE PRESENCE OF ANCHOR POINTS 24

25 This design of the EITC program has been shown to encourage labor force participation (see Holtzblatt et al. (1994) and Meyer (2002)). However, the sharp change in the marginal tax rate on earned income at the first kink in the EITC schedule provides an incentive for filers to report just enough income to obtain the maximum credit and not more (see Saez (2010) for a detailed discussion of the effect of the incentives at these kink points). The change in the filer s marginal tax rate at this first kink can be as high as 45 percentage points (for filers with three or more dependents), which may represent a substantial incentive to manipulate reported income in order to be just at that kink point. There are similar, albeit smaller, incentives to locate just to the left of the second kink point as well. Here, the phase-out of the EITC means that the increase in marginal tax rates at that point may be as high as 21.6 percentage points (for filers with two or more dependents). Thus, one possible cause of that large intensive margin changes in the reported income of EITC participants who have been audited may be that before the audit, filers are manipulating income in order to locate near the first or second kink point, but after the audit, the increase in compliance results in filers becoming unanchored from these kink points. Thus, one would see a change in reported income that is large and persistent as it now increases with the filers income growth over time rather than being tied to these threshold levels created by the incentives of tax policy. The most direct test of this hypothesis is to look at whether tax payers do in fact bunch at these kink points before audit and whether that behavior is changed as a result of the audit. Saez (2010) and Mortenson and Whitten (2015) both find evidence of bunching around the first kink point, mostly driven by filers with income that is less 25

26 verifiable by the IRS such as self-employment income. Using our sample, we test for this bunching pattern. Figure 3 provides evidence of this change in filer behavior by comparing the treatment and control groups in all years since audit. Panel (a) shows the density of filers with self-employment income in all years before the audit year for the control group. Recall that our control group is never audited and thus audit year is referring to the year of the NRP wave ( ) for which members of the control group were drawn. Panel (b) shows this density for all years these filers were observed after the audit year. For the control group, bunching at the first kink point in the EITC schedule is evident both before and after the audit. Panels (c) and (d) show the corresponding histograms for the treatment group, those selected for an NRP audit. For this group, the pre-audit period (Panel (c)) shows bunching similar to that of the control group around the first kink point. However, following the audit (Panel (d)), the distribution shows less of a mass at the first kink point and a more uniform density to the right of the first kink point on the EITC schedule. The elasticity of earned income implied from this bunching is 0.48 in the years before the audit and 0.23 in the years following audit. This decline in the elasticity of earned income, visible by the decline in the fraction of filers bunched near the first kink point on the EITC schedule in the treatment group, is statistically significant at the five percent level. [Figure 3 about here] VIII. OTHER MARGINS OF RESPONSE TO AUDIT The EITC also provides incentives for reporting behavior to change along other dimensions. In particular, the amount of the filer s credit is an increasing function of the 26

27 number of dependents (up to three) and the credit phase out is shifted to the right for married filers relative to single filers (see Figure 1). Thus, one might consider how audits affect reporting behavior along these dimensions, which represent notches in the EITC schedule. We present results on those responses below, using our sample of EITC filers. A. Dependents The EITC amounts increase significantly with the number of qualifying dependents. While related to other definitions of dependent (such as the one used for determining the number of tax exemptions on the form 1040) there are some differences. Importantly, the dependent must be a child who has a valid SSN, which requires the tax filer to interact with an additional government agency unrelated to the IRS. In contrast, dependents claimed for exemptions do not need to be eligible for a SSN, but may be a resident of Mexico or Canada who obtains an individual taxpayer identification number (ITIN) through the IRS. McCubbin (2000) provides evidence of compliance issues from EITC filers stemming from the misreporting of dependents in order to increase credit amounts, even after a 1987 change in tax administration that required filers to report the social security numbers of dependents over age 5. Indeed, Splinter et al. (2017) provide recent evidence suggesting that the number of qualifying children is an important margin along which EITC participants respond. Figure 4 shows how the number of EITC-qualifying dependents are affected by audits for our EITC sample, plotting the point estimates and confidence intervals from estimating the regression model in Equation (2) with the number of qualified dependents 27

28 as the dependent variable. Audit filers report about 0.1 less dependents two years after audit as compared to those not being audited. [Figure 4 about here] The jump in EITC benefits is largest between one and two dependents. 27 Thus, we also estimate a linear probability model with filer fixed effects where the dependent variable is an indicator variable for different numbers of dependents. 28 The results from estimating these models are presented in Table 7. These results show that the change in the number of children post audit from EITC filers is almost entirely due to a lower likelihood of reporting two dependents and a higher likelihood of reporting zero dependents. Two years after audit, the likelihood of an audited filer reporting two dependents drops by 3.4 percentage points relative to claimants who were not audited, while the likelihood of reporting zero dependents rises by 5 percentage points. [Table 7 about here] While we don t have sufficient information to determine how much of the total change in the EITC after audit is solely due to changes in the number of dependents, indirect evidence suggests this effect is large. When we look at the direct effects of NRP audits, we find that 53.0 percent of the change in EITC dollars on audited returns come 27 Note that prior to 2009 the EITC amount was the same for families with two or more dependents. The 2009 American Recovery and Reinvestment Act added additional income thresholds and credit amounts for filers with three or more qualified dependents. This means that the incentives to report additional children (in particular 3 instead of 2) changed beginning in This may affect our results, but as Table 7 shows, there is little response among those with more than two dependents, suggesting this is not quantitatively important for our overall results. To examine whether the results are affected by the policy change, we ran alternative specifications of the models in Table 7 (the effects of audits on dependents) that included an interaction between post-2009 and a post-audit indicator variable. The interaction was insignificant, implying no significant difference between the overall audit effect and the effect after this policy change. 28 Throughout the paper, when the dependent variable is an indicator variable, we use a linear probability model for our discrete choice regression models. We do this in order to allow for individual fixed effects which are not possible with a logit model or results in inconsistent estimates from a probit model. 28

29 from filers who were found to be correctly reporting income but misreporting the number of qualifying dependents. 29 B. Filing status The amount of EITC for which a taxpayer is eligible can depend on filing status, for two reasons. First, starting in 2002, the beginning and ending points of the phase-out range are higher for married taxpayers filing jointly than they are for heads of household. As a result, a married taxpayer with income in the phase-out range is eligible for a larger EITC than a head of household with the same amount of income. Second, a couple in which both partners earn income may be eligible for different EITC amounts depending on whether they were married and filed jointly, or unmarried and filed separately. These two factors suggest that filing status among EITC recipients may be affected by an audit. Table 8 shows the changes in the fraction of filers filing as married or head of household following audit, estimating a linear probability model of Equation (2) where the dependent variable is an indicator for filing status. Results for the sample of filers with AGI under $40,000 are presented together with the EITC sample to allow for comparison across these groups. These results show that the rate at which filers switch their status to married increases after audit for both groups. However, the increase for the EITC sample is about two to three times as large for the group of low income filers, peaking at a 2.5 percent increase in the likelihood of filing married six years after audit. [Table 8 about here] 29 This result is similar to what McCubbin (2000) finds using data from the 1990s, where she reports that 58.6 percent of EITC adjustments stem from returns where there were adjustments to the number of qualifying dependents and no other adjustments to the return. 29

30 Because head of household status requires one or more qualifying dependents, and because Table 7 shows an increase in taxpayers claiming zero dependents, we should expect to find a decrease in the number of unmarried taxpayers claiming this status. Columns 3 and 4 of Table 8 show the decline in head of household status post audit. Again, changes for the EITC sample are two to three times are larger for the low-income sample. Relative to EITC claimants who were not audited, the likelihood of filing as head of household for EITC claimants who were audited is 7.2 percent lower two years after audit. C. Extensive margin results We now turn to extensive margin changes in filing behavior. To do this, we use a linear probability model with filer fixed effects. This model is analogous to that presented in Equation (2) in Section 4, but with the dependent variable being an indicator variable for whether the filer claimed the EITC. Figure 5 summarizes the results from this model as estimated on the sample of EITC filers, plotting the point estimates and the 95 percent confidence interval. Filers who were audited have a 6-percentage point lower likelihood of claiming the EITC two years after audit and a 2-percentage point lower likelihood six years after audit, relative to the EITC claimants who were not audited. [Figure 5 about here] D. Overall Effect on EITC Claims These results highlight how audits can counteract incentives to misreport either intentionally or not along many dimensions. To understand the total effect of audit on 30

31 EITC claims, we estimate Equation (2) with the EITC claim amount as the dependent variable. We estimate this regression model for the EITC sample only. Figure 6 plots the point estimates for the coefficient on the time-since-audit indicator variables, along with the 95 percent confidence interval around the estimates. EITC claims by those who are audited drop by as much as $200 relative to those who are not audited. Even 6 years after audit, those who were audited claim about $100 dollars less of a credit relative to those who have not been audited. In total, over the first 6 years after audit, the fall in EITC claims is about $800, more than even the average amount of EITC overturned upon audit, $ In other words, each audit reduces total EITC claims by around $1,300. [Figure 6 about here] It is also important to note, however, that we cannot say that these behavioral changes in claiming behavior represent an increase in tax compliance. The EITC is not a credit that is simple to claim correctly. For example, it can involve relatively complicated documentation (such as obtaining a valid SSN) to establish dependents as qualifying. Some taxpayers who might in fact be eligible for EITC might rationally decide to reduce or forgo their claim in order to avoid having to do bookkeeping. Indeed, Bhargava and Manoli (2015) report that the rate of incomplete takeup of the EITC is about 25 percent. Lower take up rates increase tax revenues but would not represent an increase in tax compliance. We cannot therefore interpret all increases in revenue as increases in compliance. 30 The $501 average adjustment can be computed from Table 3. The average adjustment to the EIC for those with an adjustment is $1,066 and 47.1 percent of EITC filers have an adjustment to their claim, thus the mean without conditional on a non-zero adjustment is $

32 IX. CONCLUSION Our paper merges two streams of tax compliance literature. One considers the tax compliance patterns of EITC claimants. For example, do they overstate their EITC claims, do they bunch at EITC kinks, or do they manipulate family status to meet the EITC criteria. The second studies the dynamic effects of audits, such as whether they promote greater compliance in the future. 31 On the matter of incentives, we find evidence that EITC participants at least partially respond to the incentives built into the EITC. The largest amounts of over- and under-reporting of income occur in the regions of the EITC schedule where it is most beneficial to the taxpayer, and there is a large amount of bunching at a spot on the EITC schedule that maximizes the amount of the credit. There is also a tendency to over-report the number of qualified dependents, which drives around half of the overall overclaiming amount. However, taxpayers do not seem driven entirely by EITC incentives. The second kink point does not exhibit large amounts of bunching, and under-reporting of income is the dominant response even when there is clear incentive to over-report income. On the topic of audit response, the empirical exercise in this paper offers clear evidence for auditing effectiveness among EITC recipients. We find that audits of EITC participants result in changes along several margins, including income, EITC claim rates, filing status, and the number of dependents. These effects are quantitatively significant and persist for at least six years after audit. 31 We use an idea from an anonymous referee to frame this conclusion. 32

33 Finally, at the junction of these two topics we find evidence that audited taxpayers are dissuaded away from misreporting incentivized by the EITC. It is possible that some of the post-audit changes in reporting come from better understanding of the tax code and not malicious non-compliance behavior. However, the reduction in bunching at EITC kink points following audit suggest that compliance is likely improved following audit. Tax payers strategically locating at this point on the income distribution are unlikely to be driven by error or misunderstanding. In addition, about 69 percent of EITC claimants use professional preparers, who presumably have the expertise to navigate the complexity of filing a tax return. Thus, in light of the evidence presented here, we find audits to be an effective tool to better ensure that the benefits of an anti-poverty program go to the intended beneficiaries of that program. The policy relevance of the results is clear, as the effectiveness of audits affects the distribution of billions of dollars targeted at millions of taxpayers. On average, an audit of an EITC claimant reduced EITC claims by about $1,300 over 6 years; $500 dollars in credits that are overturned on average from the audit itself, and an $800 reduction in claims in the years following the audit. In 2014, there were 387,658 audits of EITC claimants (Internal Revenue Service 2014). If filers responses to operational audits is similar to their responses to NRP audits, our estimates suggest that the operational audits of EITC claimants in 2014 resulted in a reduction of EITC claims of about $500 million in total, $310 million of which is due to reductions in claims post-audit. These are funds (net of enforcement costs) that can be reused to better target the intended beneficiaries. 33

34 DISCLOSURES The authors have no financial arrangements that might give rise to conflicts of interest with respect to the research reported in this paper. ACKNOWLEDGEMENTS We thank Janet McCubbin and Mark Mazur for their helpful comments as well as seminar participants at the University of South Carolina, Institute for Fiscal Studies, and the National Tax Association. The views expressed in this paper are those of the authors and do not necessarily reflect the views of the U.S. Department of the Treasury or the Office of Tax Analysis. 34

35 REFERENCES: Advani, Arun, Jonathan Shaw, and William Elming "How long-lasting are the effects of audits?" Working Paper, Tax Administration Research Center, UK. Bhargava, Saurabh and Dayanand Manoli, "Psychological Frictions and the Incomplete Take-Up of Social Benefits: Evidence from an IRS Field Experiment," American Economic Review, November 2015, 105(11), Blumenthal, Marsha; Brian Erard and Chih-Chin Ho "Participation and Compliance with the Earned Income Tax Credit." National Tax Journal, DeBacker, Jason, Bradley Heim, Anh Tran, and Alexander Yuskavage, Legal Enforcement and Corporate Behavior: An Analysis of Tax Aggressiveness after an Audit, Journal of Law and Economics, May 2015, 58, ,,, and, Once Bitten, Twice Shy? The Impact of IRS Audits on Filer Behavior, Journal of Law and Economics, forthcoming. Holtzblatt, Janet, Administering Refundable Tax Credits: Lessons from the EITC Experience, Proceedings of the Annual Conference on Taxation Held under the Auspices of the National Tax Association-Tax Institute of America, 1991, 84, , Janet McCubbin, and Robert Gillette, Promoting Work Through the EITC, National Tax Journal, September 1994, 47 (3), Internal Revenue Service "IRS Releases New Tax Gap Estimates." Internal Revenue Service "IRS Audits." Internal Revenue Service "Internal Revenue Service Data Book, Joulfaian, David and Mark Rider, Tax Evasion in the Presence of Negative Income Tax Rates, National Tax Journal, 1996, 49 (4), Kastlunger, Barbara; Erich Kirchler; Luigi Mittone and Julia Pitters Sequences of Audits, Tax Compliance, and Taxpaying Strategies. Journal of Economic Psychology, 30(3): Kleven, Henrik Jacobsen; Martin B Knudsen; Claus Thustrup Kreiner; Søren Pedersen and Emmanuel Saez "Unwilling or Unable to Cheat? Evidence from a Tax Audit Experiment in Denmark." Econometrica, 79(3), Leibel, Kara Tax Compliance and Sources of Error for the Earned Income Tax Credit Claimed on Returns. Internal Revenue Service, Technical Paper. 35

36 McCubbin, Janet, EITC Noncompliance: The Determinants of the Misreporting of Children, National Tax Journal, 2000, 53 (4), Meyer, Bruce D., Labor Supply at the Extensive and Intensive Margins: The EITC, Welfare, and Hours Worked, The American Economic Review, 2002, 92 (2), Mortenson, Jacob A. and Andrew Whitten, Bunching to Maximize Tax Credits: Evidence from Kinks in the U.S. Tax Schedule, mimeo, Phillips, Mark and Alan Plumley, It s Hard to be Good: The Joint Determination of Taxpayer Effort and Compliance,, mimeo, Pomeranz, Dina No Taxation without Information - Deterrence and Self- Enforcement in the Value Added Tax, American Economic Review, 105(8), Saez, Emmanuel, Do Taxpayers Bunch at Kink Points?, American Economic Journal: Economic Policy, August 2010, 2 (3), Slemrod, Joel. Tax Compliance and Enforcement: New Research and its Policy Implications, Ross School of Business Paper No. 1302, Slemrod, Joel, Brett Collins, Jeffrey Hoopes, Daniel Reck, and Michael Sebastiani. Does Credit-Card Information Reporting Improve Small-Business Tax Compliance? National Bureau of Economic Research, NBER Working Paper 21412, David Splinter, Jeff Larrimore, and Jacob Mortenson, Whose Child Is This? Shifting of Dependents among EITC Claimants within the Same Household, National Tax Journal, December 2017, 70 (4), Whittington, Leslie A. and James Alm, Tax Reductions, Tax Changes, and the Marriage Penalty, National Tax Journal, September 2001, 54 (3),

37 Figure 1 Earned Income Tax Credit Schedules Notes: EITC Schedule for Dashed lines denote schedule for married-joint return filers. 37

38 Figure 2 Effects of Audit on Income by Source Notes: Point estimates from a fixed effects regression model described by Equation 3. The dependent variable is the amount of income of a given typ (adjusted gross income, wages, Schedule C income) reported on the filers return (in 2005$). We estimate the model on the sample of EITC filers described in Section 3. 38

39 Figure 3 Distribution of Taxpayers with Self-Employment Income (a) Control group, before audit (b) Control group, after audit (c) Treatment group, before audit (d) Treatment group, after audit 39