## 2. Differential Effects of Audits

## Source details

**Canonical URL:** [2. Differential Effects of Audits](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019223-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2019/wpiea2019223-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019223-print-pdf.pdf.json)

---

### I. Introduction
- Administrative capacity to audit returns is limited; the IRS audits only about 1 percent of all self-employed individual income taxpayers annually.
- In Fiscal Year 2018, IRS audits of self-employed individual income taxpayers resulted in almost $2 billion in recommended additional tax assessments.
- Audits produce:
  - Direct revenue from assessed additional taxes, interest, and penalties.
  - Indirect revenue effects via deterrence:
    - General deterrence: improvement in compliance in the general population in response to higher audit rates.
    - Specific deterrence: impact of an audit on the audited taxpayer’s subsequent compliance behavior.
- Study design and main aggregate finding:
  - Administrative panel dataset: 2,453 self-employed taxpayers (Schedule C filers) audited after filing their Tax Year 2007 returns and a comparison sample of 6,922 Schedule C filers who were not audited.
  - Statistical matching techniques select unaudited controls for each audited taxpayer to construct counterfactuals.
  - Difference-in-differences estimates indicate operational tax audits induce taxpayers to increase reported taxable income by roughly 10% one year after the examination; three years after the audit reported income remains modestly (2%) above pre-audit levels.
- Key nuance: audit outcomes matter — separate estimates were generated for taxpayers who did (determined noncompliant) and did not (determined compliant) receive an additional tax assessment.

### II. Heterogeneous Effects by Audit Outcome (summary of key estimates)
- Determined noncompliant (received an additional tax assessment):
  - Report approximately 64% more taxable income than matched controls in the year following the audit.
  - Report around 44% more income three years after the audit.
- Determined compliant (no additional tax assessment):
  - Report approximately 15% less income the year after they were audited than matched counterparts.
  - Report around 21% less three years later.
- Interpretation cautions:
  - A “no-change” audit is not a perfect measure of compliance; some noncompliance may go undetected.
  - Some tax adjustments may be overruled on appeal; examination outcome is an imperfect proxy for taxpayer intent.
  - Behavioral mechanisms (e.g., bomb-crater effect, erosion of trust) may explain counter-deterrent responses among determined compliant taxpayers.

### III. Methodology
- Empirical strategy:
  - Matched difference-in-differences approach to estimate average treatment effect on the treated (ATT): ATT = E[ΔY1 | D=1] – E[ΔY0 | D=1].
  - Construct matched control groups to satisfy conditional independence: treatment assignment independent of outcome ΔY0 after conditioning on covariates X.
  - Three baseline covariate sets:
    - Set I: DIF score ventile, pre-treatment profitability (profit as share of business income), pre-treatment total taxable income.
    - Set II: Set I plus three distance measures reflecting income sources, business expenses, and benefit schemes (constructed from indicator variables).
    - Set III: Set II plus lags of variables in Set II and interactions with immediate pre-treatment counterparts, totaling eighteen control variables.
  - Dependent variable: changes in reported income (Δ) to control for time-constant unobservables.
- Matching estimators (four specifications):
  - 3-nearest neighbor matching with replacement (benchmark).
  - Kernel matching estimator using the normal density.
  - Local linear ridge estimator (Seifert & Gasser 2000).
  - Kernel-matching estimator re-estimated using the pairwise Mahalanobis distance (MHD).
- Practical implementation:
  - For each audited taxpayer, statistical matching selects one or more unaudited taxpayers as matched controls; separate matched controls constructed for each audit outcome subgroup.

### IV. Data and Sample Construction
- Initial samples:
  - Audit sample: 6,451 randomly selected self-employed taxpayers audited for TY2007.
  - Initial comparison sample: 11,218 unaudited taxpayers.
- Stratification: DIF score ventile cut-offs computed within examination classes to ensure similar DIF distributions; unaudited taxpayers drawn randomly from each stratum.
- Exclusion restrictions and additional requirements to isolate effect of a single audit:
  - TY2007 audit must begin prior to filing TY2008 return (treatment group).
  - No audits permitted for TY2005 through TY2009 for members of either sample (with the one exception above).
  - Taxpayers required to have filed Schedule C in both TY2006 and TY2007.
  - Returns TY2005–TY2009 must be filed on time and in chronological order.
  - Exclude extreme incomes: top 2.5% and bottom 2.5% of income distribution.
- Sample selection steps (Table 1 summary):
  - Step 0 Initial sample: Comparison sample 11,218; Audit sample 6,451; Total 17,699.
  - Step 1 Incomplete data or TY2007 audit began after filing of TY2008 return: Comparison sample 9,651 (0.86); Audit sample 4,251 (0.66); Total 13,902 (0.79).
  - Step 2 Violation of no-audit restriction for TY2005-TY2009: Comparison sample 9,560 (0.99); Audit sample 3,768 (0.89); Total 13,328 (0.76).
  - Step 3 Failure to file timely and chronologically: Comparison sample 7,278 (0.76); Audit sample 2,619 (0.70); Total 9,897 (0.74).
  - Step 4 Outlier exclusion: Comparison sample 6,922 (0.95); Audit sample 2,453 (0.94); Total 9,375 (0.95).
- Final sample: 9,375 taxpayers (2,453 treated taxpayers and 6,922 comparison taxpayers).

- Covariate balance (Table 2 mean values in 2007; matched averages are average of the 4 matching procedures):
  - DIF score: Control group 9.79; Determined non-compliant Treatment 10.94; Matched Average 10.86; Determined compliant Treatment 10.31; Matched Average 10.23.
  - Log tax: Control group 7.92; Determined non-compliant Treatment 8.63; Matched Average 8.58; Determined compliant Treatment 7.93; Matched Average 7.89.
  - Profitability: Control group -672.47; Determined non-compliant Treatment -266.02; Matched Average -228.80; Determined compliant Treatment -888.52; Matched Average -893.94.
  - Income distance: Control group 11.23; Determined non-compliant Treatment 10.73; Matched Average 10.57; Determined compliant Treatment 10.97; Matched Average 10.91.
  - Business distance: Control group 7.90; Determined non-compliant Treatment 7.69; Matched Average 7.65; Determined compliant Treatment 8.69; Matched Average 8.59.
  - Other distance: Control group 5.50; Determined non-compliant Treatment 6.23; Matched Average 5.81; Determined compliant Treatment 5.63; Matched Average 5.27.

### V. Empirical Results — Aggregate Effects (high-level)
- Common-trends check:
  - Natural log of reported taxable income follows a similar rising trend for treatment and matched control groups up through TY2007, consistent with the common-trends assumption.
  - Reported income in both groups declines substantially between TY2007 and TY2009 (likely due to the financial crisis), but the decline is weaker for the treatment group, indicating a positive audit effect on subsequent reporting.
- Table 3 (aggregate estimates across matching methods and covariate sets):
  - Average estimate: audited taxpayers report around 12% more income the year after they were audited relative to matched controls (statistical significance noted in text).

### Aggregate estimated effects of audits (one year and three years)
- One year after the audit (average across matching estimators and covariate sets):
  - Set I: Average ATT = 0.119** (0.059)
  - Set II: Average ATT = 0.162*** (0.062)
  - Set III: Average ATT = 0.159*** (0.058)
  - Range across estimators (Set I): smallest 0.087 (Nearest Neighbor), largest 0.138** (Kernel Propensity). Smallest estimated impact 9% (nearest neighbor) and largest 14% (kernel propensity) as reported in text; ratio = 1.6.
  - Nearest-neighbor matching generally yields less precise estimates than local linear ridge or kernel matching.
  - Estimates remain statistically significant at the 5% level for more inclusive covariate sets (Set II and Set III).
- Three years after the audit (average across matching estimators and covariate sets):
  - Set I: Average ATT = -0.075 (0.072)
  - Set II: Average ATT = -0.056 (0.070)
  - Set III: Average ATT = -0.033 (0.069)
  - On average, audited taxpayers report around 3% to 8% less income than matched counterparts three years after the audit; effects are statistically insignificant across all matching estimators and covariate sets.

- Tabulated one-year ATT estimates (selected exact entries from Table 3):
  - Nearest Neighbor (one year): Set I = 0.087 (0.078); Set II = 0.136* (0.075); Set III = 0.159** (0.075)
  - Kernel Propensity (one year): Set I = 0.138** (0.058); Set II = 0.177*** (0.063); Set III = 0.168*** (0.057)
  - Local Ridge (one year): Set I = 0.129** (0.060); Set II = 0.181*** (0.065); Set III = 0.186*** (0.059)
  - Kernel MHD (one year): Set I = 0.122** (0.056); Set II = 0.157** (0.062); Set III = 0.123** (0.062)

### Differential effects by audit outcome (determined noncompliant vs. determined compliant)
- Summary of main findings:
  - Audits that result in an additional tax assessment (determined noncompliant):
    - On average, audited taxpayers in this group report around 64% more income the year after they were audited, relative to matched control samples.
    - Average ATT one year after audit (selected averages from Table 4):
      - Determined noncompliant, Average: Set I = 0.612*** (0.104); Set II = 0.654*** (0.100); Set III = 0.639*** (0.104)
    - Specific one-year estimator values (Set I examples):
      - Nearest Neighbor = 0.609*** (0.142)
      - Kernel Propensity = 0.626*** (0.111)
      - Local Ridge = 0.622*** (0.112)
      - Kernel MHD = 0.591*** (0.109)
    - Variation across estimators (Set I): smallest 59.1% (Kernel MHD), largest 62.6% (Kernel Propensity); ratio = 1.06 (relatively little model uncertainty).
    - Three years after the audit, reported income remains substantially higher for this group (increase of 44% reported in text).
    - Three-year average ATT (Table 5 averages):
      - Determined noncompliant, Average: Set I = 0.439*** (0.120); Set II = 0.424*** (0.111); Set III = 0.464*** (0.120)
    - Selected three-year estimator values (Set I examples):
      - Nearest Neighbor = 0.402** (0.166)
      - Kernel Propensity = 0.421*** (0.123)
      - Local Ridge = 0.418*** (0.123)
      - Kernel MHD = 0.515*** (0.119)

  - Audits that do not result in an additional tax assessment (determined compliant):
    - On average, audited taxpayers who were determined compliant report around 15% less income than their matched counterparts one year after the audit.
    - Average ATT one year after audit (selected averages from Table 4):
      - Determined compliant, Average: Set I = -0.154** (0.073); Set II = -0.154** (0.074); Set III = -0.150** (0.073)
    - Specific one-year estimator values (Set I examples):
      - Nearest Neighbor = -0.154* (0.091)
      - Kernel Propensity = -0.140* (0.074)
      - Local Ridge = -0.148** (0.075)
      - Kernel MHD = -0.173** (0.085)
    - Three years after the audit, reported income remains substantially lower for this group (decline of 21% reported in text).
    - Three-year average ATT (Table 5 averages):
      - Determined compliant, Average: Set I = -0.211** (0.085); Set II = -0.218** (0.087); Set III = -0.203** (0.085)
    - Selected three-year estimator values (Set I examples):
      - Nearest Neighbor = -0.241** (0.105)
      - Kernel Propensity = -0.201** (0.088)
      - Local Ridge = -0.209** (0.088)
      - Kernel MHD = -0.191** (0.091)
    - Ratio of largest to smallest three-year impacts (parsimonious covariate set): approximately 1.26 for determined compliant, 1.28 for determined noncompliant (text).

- Behavioral interpretation:
  - Determined noncompliant taxpayers: reported taxable income increases substantially after audits relative to matched controls — interpreted as a strong pro-deterrent effect on noncompliance.
  - Determined compliant taxpayers: reported taxable income decreases in years following the audit relative to matched controls — interpreted as a counter-deterrent effect.

### Conclusion, plausible mechanisms, and implications for policy and future work
- Main conclusion:
  - Overall moderate positive effect on reported taxable income, but substantial heterogeneity by audit outcome: very substantial pro-deterrent effects for those found noncompliant and large counter-deterrent effects for those found compliant.
- Possible explanations for decline in reported income among those found compliant:
  - Undetected dishonest taxpayers may infer audits are ineffective and underreport more (longer-term driver).
  - Overly compliant taxpayers may correct upward misreporting and report less income subsequently.
  - Erosion of tax morale: coercive enforcement reduces trust and intrinsic motivation to comply (Lederman, 2018; Erard et al., 2019).
  - "Bomb crater effect": taxpayers may believe risk of future examination is low after a no-adjustment audit, reducing compliance.
- Policy implications and recommendations:
  - Improve targeting of noncompliant returns and capacity to detect noncompliance to enhance deterrence among cheaters.
  - Improve understanding of psychological impacts of audits on compliant taxpayers to revise examination approaches that mitigate erosion of tax morale and maintain incentives to comply.
  - Future work should investigate taxpayers’ perceptions of the audit experience and its effect on motivational processes, and explore differential impacts by audit technique (face-to-face vs. correspondence) and by income groups.
- Limitations noted by authors:
  - Potential for unobserved confounding in quasi-experimental design; audit selection influenced by IRS examiner discretion ("classifiers") could bias estimated treatment effects.
  - Authors control for a broad range of variables (including DIF score ventile and prior reported values of income sources and offsets) and rely on propensity score matching to mitigate bias, but residual concerns remain.
  - Sample period subject to unusually high economic volatility; replication using a more stable sample period is desirable.
  - Additional exploration of alternative audit techniques and heterogeneous taxpayer responses recommended.

*Content drawn exclusively from "Section 2.1. All of these methods have been applied using our parsimonious set of explanatory variables (Set I)" in the supplied PDF content unit.*

### References .............................................................................................................

### References

### Tables
- 1. Sample Description ................................................................................................................................................... 11
- 2. Covariate Balance ...................................................................................................................................................... 12
- 3. Estimated Treatment Effect One Year and Three Years after the Audit ............................................... 14
- 4. Estimated Treatment Effect One Year after the Audit ................................................................................. 16
- 5. Estimated ATT Three Years after the Au dit ........................................................................................................ 17

### Figures
- 1. Aggregate Effect of Audits ..................................................................................................................................... 13

*Source: wpiea2019223-print-pdf - References — https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019223-print-pdf.pdf*

### 2. Differential Effects of Audits ......................................................................................

### 2. Differential Effects of Audits

### I. Introduction
- Administrative capacity to audit returns is limited; the IRS audits only about 1 percent of all self-employed individual income taxpayers annually.
- In Fiscal Year 2018, IRS audits of self-employed individual income taxpayers resulted in almost $2 billion in recommended additional tax assessments.
- Audits produce:
  - Direct revenue from assessed additional taxes, interest, and penalties.
  - Indirect revenue effects via deterrence:
    - General deterrence: improvement in compliance in the general population in response to higher audit rates.
    - Specific deterrence: impact of an audit on the audited taxpayer’s subsequent compliance behavior.

- Study design and main aggregate finding:
  - Administrative panel dataset: 2,453 self-employed taxpayers (Schedule C filers) audited after filing their Tax Year 2007 returns and a comparison sample of 6,922 Schedule C filers who were not audited.
  - Statistical matching techniques select unaudited controls for each audited taxpayer to construct counterfactuals.
  - Difference-in-differences estimates indicate operational tax audits induce taxpayers to increase reported taxable income by roughly 10% one year after the examination; three years after the audit reported income remains modestly (2%) above pre-audit levels.
- Key nuance: audit outcomes matter — separate estimates were generated for taxpayers who did (determined noncompliant) and did not (determined compliant) receive an additional tax assessment.

### II. Heterogeneous Effects by Audit Outcome (summary of key estimates)
- Determined noncompliant (received an additional tax assessment):
  - Report approximately 64% more taxable income than matched controls in the year following the audit.
  - Report around 44% more income three years after the audit.
- Determined compliant (no additional tax assessment):
  - Report approximately 15% less income the year after they were audited than matched counterparts.
  - Report around 21% less three years later.
- Interpretation cautions:
  - A “no-change” audit is not a perfect measure of compliance; some noncompliance may go undetected.
  - Some tax adjustments may be overruled on appeal; examination outcome is an imperfect proxy for taxpayer intent.
  - Behavioral mechanisms (e.g., bomb-crater effect, erosion of trust) may explain counter-deterrent responses among determined compliant taxpayers.

### III. Methodology
- Empirical strategy:
  - Matched difference-in-differences approach to estimate average treatment effect on the treated (ATT): ATT = E[ΔY1 | D=1] – E[ΔY0 | D=1].
  - Construct matched control groups to satisfy conditional independence: treatment assignment independent of outcome ΔY0 after conditioning on covariates X.
  - Three baseline covariate sets:
    - Set I: DIF score ventile, pre-treatment profitability (profit as share of business income), pre-treatment total taxable income.
    - Set II: Set I plus three distance measures reflecting income sources, business expenses, and benefit schemes (constructed from indicator variables).
    - Set III: Set II plus lags of variables in Set II and interactions with immediate pre-treatment counterparts, totaling eighteen control variables.
  - Dependent variable: changes in reported income (Δ) to control for time-constant unobservables.

- Matching estimators (four specifications):
  - 3-nearest neighbor matching with replacement (benchmark).
  - Kernel matching estimator using the normal density.
  - Local linear ridge estimator (Seifert & Gasser 2000).
  - Kernel-matching estimator re-estimated using the pairwise Mahalanobis distance (MHD).

- Practical implementation:
  - For each audited taxpayer, statistical matching selects one or more unaudited taxpayers as matched controls; separate matched controls constructed for each audit outcome subgroup.

### IV. Data and Sample Construction
- Initial samples:
  - Audit sample: 6,451 randomly selected self-employed taxpayers audited for TY2007.
  - Initial comparison sample: 11,218 unaudited taxpayers.
- Stratification: DIF score ventile cut-offs computed within examination classes to ensure similar DIF distributions; unaudited taxpayers drawn randomly from each stratum.
- Exclusion restrictions and additional requirements to isolate effect of a single audit:
  - TY2007 audit must begin prior to filing TY2008 return (treatment group).
  - No audits permitted for TY2005 through TY2009 for members of either sample (with the one exception above).
  - Taxpayers required to have filed Schedule C in both TY2006 and TY2007.
  - Returns TY2005–TY2009 must be filed on time and in chronological order.
  - Exclude extreme incomes: top 2.5% and bottom 2.5% of income distribution.
- Sample selection steps (Table 1 summary):
  - Step 0 Initial sample: Comparison sample 11,218; Audit sample 6,451; Total 17,699.
  - Step 1 Incomplete data or TY2007 audit began after filing of TY2008 return: Comparison sample 9,651 (0.86); Audit sample 4,251 (0.66); Total 13,902 (0.79).
  - Step 2 Violation of no-audit restriction for TY2005-TY2009: Comparison sample 9,560 (0.99); Audit sample 3,768 (0.89); Total 13,328 (0.76).
  - Step 3 Failure to file timely and chronologically: Comparison sample 7,278 (0.76); Audit sample 2,619 (0.70); Total 9,897 (0.74).
  - Step 4 Outlier exclusion: Comparison sample 6,922 (0.95); Audit sample 2,453 (0.94); Total 9,375 (0.95).
- Final sample: 9,375 taxpayers (2,453 treated taxpayers and 6,922 comparison taxpayers).

- Covariate balance (Table 2 mean values in 2007; matched averages are average of the 4 matching procedures):
  - DIF score: Control group 9.79; Determined non-compliant Treatment 10.94; Matched Average 10.86; Determined compliant Treatment 10.31; Matched Average 10.23.
  - Log tax: Control group 7.92; Determined non-compliant Treatment 8.63; Matched Average 8.58; Determined compliant Treatment 7.93; Matched Average 7.89.
  - Profitability: Control group -672.47; Determined non-compliant Treatment -266.02; Matched Average -228.80; Determined compliant Treatment -888.52; Matched Average -893.94.
  - Income distance: Control group 11.23; Determined non-compliant Treatment 10.73; Matched Average 10.57; Determined compliant Treatment 10.97; Matched Average 10.91.
  - Business distance: Control group 7.90; Determined non-compliant Treatment 7.69; Matched Average 7.65; Determined compliant Treatment 8.69; Matched Average 8.59.
  - Other distance: Control group 5.50; Determined non-compliant Treatment 6.23; Matched Average 5.81; Determined compliant Treatment 5.63; Matched Average 5.27.

### V. Empirical Results — Aggregate Effects (high-level)
- Common-trends check:
  - Natural log of reported taxable income follows a similar rising trend for treatment and matched control groups up through TY2007, consistent with the common-trends assumption.
  - Reported income in both groups declines substantially between TY2007 and TY2009 (likely due to the financial crisis), but the decline is weaker for the treatment group, indicating a positive audit effect on subsequent reporting.
- Table 3 (aggregate estimates across matching methods and covariate sets):
  - Average estimate: audited taxpayers report around 12% more income the year after they were audited relative to matched controls (statistical significance noted in text).

*Source: Excerpt from "2. Differential Effects of Audits" (chapter/section) of the IMF working paper PDF wpiea2019223-print-pdf.*

### Section 2.1. All of these methods have been applied using our parsimonious set of explanatory variables (Set I)

### Section 2.1. All of these methods have been applied using our parsimonious set of explanatory variables (Set I)

### Aggregate estimated effects of audits (one year and three years)
- One year after the audit (average across matching estimators and covariate sets):
  - Set I: Average ATT = 0.119** (0.059)
  - Set II: Average ATT = 0.162*** (0.062)
  - Set III: Average ATT = 0.159*** (0.058)
  - Range across estimators (Set I): smallest 0.087 (Nearest Neighbor), largest 0.138** (Kernel Propensity). Smallest estimated impact 9% (nearest neighbor) and largest 14% (kernel propensity) as reported in text; ratio = 1.6.
  - Nearest-neighbor matching generally yields less precise estimates than local linear ridge or kernel matching.
  - Estimates remain statistically significant at the 5% level for more inclusive covariate sets (Set II and Set III).

- Three years after the audit (average across matching estimators and covariate sets):
  - Set I: Average ATT = -0.075 (0.072)
  - Set II: Average ATT = -0.056 (0.070)
  - Set III: Average ATT = -0.033 (0.069)
  - On average, audited taxpayers report around 3% to 8% less income than matched counterparts three years after the audit; effects are statistically insignificant across all matching estimators and covariate sets.

- Tabulated one-year ATT estimates (selected exact entries from Table 3):
  - Nearest Neighbor (one year): Set I = 0.087 (0.078); Set II = 0.136* (0.075); Set III = 0.159** (0.075)
  - Kernel Propensity (one year): Set I = 0.138** (0.058); Set II = 0.177*** (0.063); Set III = 0.168*** (0.057)
  - Local Ridge (one year): Set I = 0.129** (0.060); Set II = 0.181*** (0.065); Set III = 0.186*** (0.059)
  - Kernel MHD (one year): Set I = 0.122** (0.056); Set II = 0.157** (0.062); Set III = 0.123** (0.062)

### Differential effects by audit outcome (determined noncompliant vs. determined compliant)
- Summary of main findings:
  - Audits that result in an additional tax assessment (determined noncompliant):
    - On average, audited taxpayers in this group report around 64% more income the year after they were audited, relative to matched control samples.
    - Average ATT one year after audit (selected averages from Table 4):
      - Determined noncompliant, Average: Set I = 0.612*** (0.104); Set II = 0.654*** (0.100); Set III = 0.639*** (0.104)
    - Specific one-year estimator values (Set I examples):
      - Nearest Neighbor = 0.609*** (0.142)
      - Kernel Propensity = 0.626*** (0.111)
      - Local Ridge = 0.622*** (0.112)
      - Kernel MHD = 0.591*** (0.109)
    - Variation across estimators (Set I): smallest 59.1% (Kernel MHD), largest 62.6% (Kernel Propensity); ratio = 1.06 (relatively little model uncertainty).
    - Three years after the audit, reported income remains substantially higher for this group (increase of 44% reported in text).
    - Three-year average ATT (Table 5 averages):
      - Determined noncompliant, Average: Set I = 0.439*** (0.120); Set II = 0.424*** (0.111); Set III = 0.464*** (0.120)
    - Selected three-year estimator values (Set I examples):
      - Nearest Neighbor = 0.402** (0.166)
      - Kernel Propensity = 0.421*** (0.123)
      - Local Ridge = 0.418*** (0.123)
      - Kernel MHD = 0.515*** (0.119)

  - Audits that do not result in an additional tax assessment (determined compliant):
    - On average, audited taxpayers who were determined compliant report around 15% less income than their matched counterparts one year after the audit.
    - Average ATT one year after audit (selected averages from Table 4):
      - Determined compliant, Average: Set I = -0.154** (0.073); Set II = -0.154** (0.074); Set III = -0.150** (0.073)
    - Specific one-year estimator values (Set I examples):
      - Nearest Neighbor = -0.154* (0.091)
      - Kernel Propensity = -0.140* (0.074)
      - Local Ridge = -0.148** (0.075)
      - Kernel MHD = -0.173** (0.085)
    - Three years after the audit, reported income remains substantially lower for this group (decline of 21% reported in text).
    - Three-year average ATT (Table 5 averages):
      - Determined compliant, Average: Set I = -0.211** (0.085); Set II = -0.218** (0.087); Set III = -0.203** (0.085)
    - Selected three-year estimator values (Set I examples):
      - Nearest Neighbor = -0.241** (0.105)
      - Kernel Propensity = -0.201** (0.088)
      - Local Ridge = -0.209** (0.088)
      - Kernel MHD = -0.191** (0.091)
    - Ratio of largest to smallest three-year impacts (parsimonious covariate set): approximately 1.26 for determined compliant, 1.28 for determined noncompliant (text).

- Behavioral interpretation from figure and text:
  - For determined noncompliant taxpayers: reported taxable income increases substantially after audits relative to matched controls — interpreted as a strong pro-deterrent effect on noncompliance.
  - For determined compliant taxpayers: reported taxable income decreases in years following the audit relative to matched controls — interpreted as a counter-deterrent effect.

### Conclusion, plausible mechanisms, and implications for policy and future work
- Main conclusion:
  - Overall moderate positive effect on reported taxable income, but substantial heterogeneity by audit outcome: very substantial pro-deterrent effects for those found noncompliant and large counter-deterrent effects for those found compliant.

- Possible explanations for decline in reported income among those found compliant:
  - Undetected dishonest taxpayers may infer audits are ineffective and underreport more (longer-term driver).
  - Overly compliant taxpayers may correct upward misreporting and report less income subsequently.
  - Erosion of tax morale: coercive enforcement reduces trust and intrinsic motivation to comply (Lederman, 2018; Erard et al., 2019).
  - "Bomb crater effect": taxpayers may believe risk of future examination is low after a no-adjustment audit, reducing compliance.

- Policy implications and recommendations:
  - Improve targeting of noncompliant returns and capacity to detect noncompliance to enhance deterrence among cheaters.
  - Improve understanding of psychological impacts of audits on compliant taxpayers to revise examination approaches that mitigate erosion of tax morale and maintain incentives to comply.
  - Future work should investigate taxpayers’ perceptions of the audit experience and its effect on motivational processes, and explore differential impacts by audit technique (face-to-face vs. correspondence) and by income groups.

- Limitations noted by authors:
  - Potential for unobserved confounding in quasi-experimental design; audit selection influenced by IRS examiner discretion ("classifiers") could bias estimated treatment effects.
  - Authors control for a broad range of variables (including DIF score ventile and prior reported values of income sources and offsets) and rely on propensity score matching to mitigate bias, but residual concerns remain.
  - Sample period subject to unusually high economic volatility; replication using a more stable sample period is desirable.
  - Additional exploration of alternative audit techniques and heterogeneous taxpayer responses recommended.

*Italic: Content drawn exclusively from "Section 2.1. All of these methods have been applied using our parsimonious set of explanatory variables (Set I)" in the supplied PDF content unit.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019223-print-pdf.pdf_