## Preface

## Source details

**Canonical URL:** [Preface](https://www.imf.org/-/media/files/publications/tar/2024/english/tarea2024088-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tar/2024/english/tarea2024088-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/tar/2024/english/tarea2024088-print-pdf.pdf.json)

---

### Mission context and acknowledgements
- Mission period: June 30 – July 04, 2024.
- Mission financed by: CCAMTAC.
- Purpose: Predict the Corporate Income Tax (CIT) gap based on operational audits for the income year 2023; follow-up to a March 2024 mission that estimated the CIT gap for 2020-2022 based on operational audits.
- Team: Mr. Soren Pedersen (FAD) and Mr. Tobias Gabel Christiansen (FAD short-term expert).
- Source data: CIT returns for income year 2023 provided by the Armenia State Revenue Committee (SRC). Report findings are based on data provided by SRC.
- Acknowledged SRC staff: Mr. Arsen Sarikyan; Mr. Mkhitar Ayvazyan; Mr. Martin Sandoyan; Grigor Hakobyan.
- Report submission: Final version of the draft report submitted to Mr. Ashot H. Muradyan and Mr. Arsen Sarikyan on July 10, 2024.
- Report composition: Executive Summary; (I) Introduction; (II) Key Findings; (III) Next Steps.

### Executive Summary — predicted CIT gap for income year 2023
- Approach: Bottom-up using SRC operational audits (2020-2022) and CIT returns for 2023; prediction uses previously developed model.
- Predicted CIT gap (2023), including audit adjustments with immediate effect:
  - 121.2 billion AMD.
  - 34.1 percent of potential CIT liability.
  - 1.5 percent of GDP.
- Predicted CIT gap (2023), excluding audit adjustments with no immediate effect on revenue:
  - 80.1 billion AMD.
  - 25.5 percent of potential CIT liability.
  - 1.0 percent of GDP.
- Historical estimated CIT gaps (based on audits):
  - 2020: 38.8 percent of potential CIT liability; 83.1 billion AMD; 1.3 percent of GDP.
  - 2021: 35.2 percent of potential CIT liability; 90.3 billion AMD; 1.4 percent of GDP.
  - 2022: 32.8 percent of potential CIT liability; 108.5 billion AMD; 1.5 percent of GDP.
- Historical estimated CIT gaps excluding adjustments with no immediate effect on revenue:
  - 2020: 29.2 percent; 54.0 billion AMD; 0.9 percent of GDP.
  - 2021: 26.4 percent; 59.6 billion AMD; 0.9 percent of GDP.
  - 2022: 24.6 percent; 72.4 billion AMD; 1.0 percent of GDP.
- Observed pattern:
  - CIT gap in terms of potential CIT is trending downward when the largest corporations are excluded.
  - When considering all corporations, the predicted CIT gap shows a slight increase in 2023, suggesting the increase is driven by large corporations.
- Recommendation (headline):
  - Measure the CIT gap once audits for 2023 are completed (expected in 2025) and compare estimated gap to the predicted CIT gap for 2023.
  - SRC can use provided R code and instructions to conduct analysis internally or seek IMF capacity development for assisted self-assessment.

### Note on terminology from Executive Summary
- “Predicted”: used when CIT gap is based on CIT returns before audits.
- “Estimated”: used when CIT gap is based on audited CIT returns.
- Analysis measures the CIT compliance gap (policy gap excluded); "gap" and "compliance gap" used indistinguishably.
- Excluding audit adjustments with no immediate impact on revenue refers to cases where audit adjustment reduced a loss or carry forward of losses.
- Potential CIT liability is defined as self-reported CIT plus the estimated CIT gap.

### Recommendations (Executive Summary condensed)
- Estimate the CIT gap when audits are completed in 2025.
  - Responsible: SRC.
  - Method: SRC data analysts can use provided R code and instructions or seek IMF CD for assisted self-assessment.
  - Timing: May 2025.

### I. Introduction — model, methodology, and limitations
- Methodology:
  - Prediction based on previously developed model that estimated CIT gap for income years 2020-2022 using operational audits.
  - Model corrects for non-random selection of audited corporations (sample selection bias) using the Heckman two-step procedure.
    - First stage: probit model estimating probability of being audited using 25 separate risk scores used by SRC.
    - Second stage: models audit outcome (tax uncovered) using company characteristics and a regressor accounting for selection.
- Data used for prediction: non-audited CIT declarations for income year 2023 (declaration deadline April 20, 2024).
- Caveats and limitations:
  - Suitability of Heckman estimator depends on audit practices; narrowly focused audits may underestimate undetected noncompliance.
  - Audits targeting specific sectors or firm types increase extrapolation and uncertainty.
  - Some noncompliance may remain undetected, potentially underestimating the CIT gap.
  - Prediction assumes audit selection strategy remains unchanged and no significant tax regulation changes or new evasion methods occur.

### II. Key Findings — CIT gap results (detailed)
- Model basis:
  - Heckman model estimated using 4,432 comprehensive operational audits covering income years 2020–2022.
  - Results adjusted for sample selection bias.
- Main predicted outcomes for 2023:
  - 121.2 billion AMD; 34.1 percent of potential CIT; 1.5 percent of GDP.
  - Excluding adjustments with no immediate effect on revenue: 80.1 billion AMD; 25.5 percent of potential CIT; 1.0 percent of GDP.
- Sectoral findings:
  - Predicted median CIT gap highest in sector labeled “Extraterritorial” (sector is small so contribution to total CIT gap is limited).
  - Second-highest median CIT gap is in the mining sector.
  - Lowest median CIT gap among top-10 sectors is in the accommodation sector.
  - Sector classification follows NACE coding; sector codes are self-reported by corporations; top-10 sectors displayed due to space limitations.
- Alternative model specifications and their 2023 predictions:
  - Excluding largest corporations:
    - Predicted CIT gap: 87.9 billion AMD; 27.2 percent of potential CIT; 1.1 percent of GDP.
    - Rationale: Top 0.1 percent largest companies (by turnover) excluded to reduce influence of outliers while accounting for most of the population.
  - Excluding largest corporations and excluding adjustments with no immediate effect on revenue:
    - Predicted CIT gap: 62.8 billion AMD; 21.1 percent of potential CIT; 0.8 percent of GDP.
- Interpretation:
  - Excluding largest corporations likely reduces noise from outliers and may provide a more accurate estimate for the remaining population, but reduces the population scope.
  - The increase in the predicted CIT gap for 2023 when including all corporations suggests large corporations drive the observed rise.

### III. Next Steps
A. Set up a team of data analysts and apply the developed model
- Recommendation:
  - SRC should appoint 2-3 data analysts responsible for compiling future CIT gap estimations.
  - Invest in data analytics capacity: at least 2-3 individuals trained in data analytics.
  - Appointed team should invest around 80 percent of their time on data analytics.
  - Team should learn necessary programming tools (R code and instructions provided) to carry out CIT gap analysis.

B. Recommendations (operational)
- SRC should carry out the CIT gap estimation after income year 2023 audits are completed in 2025 and compare to the predicted CIT gap for 2023 presented in this report.
- If SRC cannot perform the estimation in 2025, they could seek IMF assistance for an assisted self-assessment of the CIT gap to develop data analytical skills and move toward self-sufficiency.

### Annex I. Supplementary material

### Methodology: Heckman 2-step estimator and application
- Purpose: Correct sample selection bias from non-random risk-based audits by estimating both the selection process and the level of non-compliance (tax uncovered from audit) in the same model.
- Model specification:
  - Outcome equation: Yi = Xiβ + ui, where Yi measures tax uncovered from audit and Xi are company characteristics (lines from the CIT return, sector, number of employees etc.).
  - Selection equation: Si = 1[Ziδ + vi > 0], where Si = 1 denotes company i was audited, Zi are factors that determine audit selection (including 25 separate risk scores used by the SRC).
  - Assumptions: vi ~ N(0,1); E(ui | vi) = γ vi. The inverse Mills ratio λ(∙) = φ(∙)/Φ(∙) arises from vi ~ N(0,1).
- Estimation procedure:
  - Estimate selection equation using a probit to obtain δ̂.
  - Compute λ(Zi δ̂) and regress Yi on Xi and λ(Zi δ̂) by OLS to obtain consistent estimates of β and γ (Heckman 2-step).
  - Predicted unconditional expectation of non-compliance: E[Yi | Xi]̂ = Xi β̂.
- Practical notes and identification:
  - Use separate variables where possible in the selection model to satisfy an exclusion restriction.
  - If the same variables are used in both models, identification relies on the non-linearity of the inverse Mills-ratio, risking unstable results from high multicollinearity.
  - In this application, 25 SRC risk scores were used to predict audits; regressions indicate 16 out of the 25 risk scores are not statistically significantly related to the tax uncovered from audit, which supports the presence of potential exclusion restrictions.
- Data handling and estimation details:
  - Data used: CIT returns and operational audits from 2020, 2021 and 2022.
  - Prices: main estimation deflated to 2020-prices; predictions for 2023 were requested in 2023-prices (non-deflated) and are presented separately.
  - Trimming: 0.1 pct. of audit results were trimmed in top and bottom for each year. Re-audits are discarded. Only corporations with a reported CIT return are included.
  - Standard errors: Standard errors are wrong when manually estimating the 2-step Heckman estimator; correct standard errors are obtained using bootstrap (bootstrap resampling done at the TIN-level).

### Estimation results and diagnostics (summary)
- Sample sizes:
  - Selection model observations: 71,374
  - Outcome model observations: 4,324
- Goodness-of-fit:
  - Selection model 푹ퟐ (McFadden’s Pseudo 푹ퟐ): 0.26
  - Outcome model 푹ퟐ: 0.14
- Inverse Mills ratio (IMR):
  - IMR coefficient in outcome model: -1032,997* with Std. Error 578,905 (significance indicated: * denotes p < 0.01, ** p < 0.05, *** p < 0.10 per table note)
  - Interpretation: The significant IMR indicates the presence of sample selection bias.
- Notable covariate notes from regressions:
  - The selection model includes 25 risk scores; many risk scores have small coefficients (both positive and negative), and several are statistically significant at conventional levels.
  - The outcome model includes tax-line variables, number of employees, turnover per employee, total compensation, and sector dummies.
  - The outcome model R^2 of 0.14 reflects the challenge of explaining heterogeneous audit adjustments across diverse companies.

### Predicted CIT gap in 2023 (non-deflated) — Heckman model predictions
- Predicted CIT gap in 2023 (All, No loss, No large, No large / No loss) — values reported in 2023-prices:
  - All: 143.9 billion AMD
  - No loss: 95.1 billion AMD
  - No large: 104.3 billion AMD
  - No large / No loss: 74.5 billion AMD
- Corresponding shares:
  - Percent of potential CIT:
    - All: 34.1
    - No loss: 25.5
    - No large: 27.2
    - No large / No loss: 21.1
  - Percent of GDP:
    - All: 1.5
    - No loss: 1.0
    - No large: 1.1
    - No large / No loss: 0.8
- Notes on definitions and construction:
  - Predictions made using a Heckman model estimated on audits from 2020–2022.
  - “All” refers to the model using all remaining audits.
  - “No loss” excludes audit adjustments with no immediate impact on revenue.
  - “No large” excludes the top 0.1 percent of the largest companies, as measured by turnover.
  - Potential CIT liability is defined as self-reported CIT plus the estimated CIT gap.
  - The reported predictions for 2023 are in 2023-prices; the underlying model used data deflated to 2020-prices for estimation coherence.

### Key takeaways and practical implications
- Non-random risk-based audit samples generate sample selection bias; the Heckman 2-step estimator provides a consistent correction when selection and outcome processes are jointly modeled.
- The significant inverse Mills-ratio indicates selection bias is present in the audit data used.
- Prediction accuracy for non-compliance depends on how well the selection and outcome models explain the data; the outcome model’s 푹ퟐ = 0.14 highlights limits in explanatory power given company heterogeneity.
- Exclusion restrictions (variables that affect audit selection but not non-compliance) improve identification; the presence of many risk scores not significantly related to audit adjustments supports identification in this application. Changes to SRC targeting strategies could make finding valid exclusion restrictions more challenging.
- For inference, correct standard errors require bootstrap; reported standard errors in the outcome model are computed using bootstrap with TIN-level resampling.

*IMF Technical Assistance Report | Preface and Executive Summary (June 30 – July 04, 2024 mission to SRC).*

### Preface ____________________________________________________________________________ 5

### Preface

### Preface (mission context and acknowledgements)
- Mission period: June 30 – July 04, 2024.
- Mission financed by: CCAMTAC.
- Purpose: Predict the Corporate Income Tax (CIT) gap based on operational audits for the income year 2023; follow-up to a March 2024 mission that estimated the CIT gap for 2020-2022 based on operational audits.
- Team: Mr. Soren Pedersen (FAD) and Mr. Tobias Gabel Christiansen (FAD short-term expert).
- Source data: CIT returns for income year 2023 provided by the Armenia State Revenue Committee (SRC). Report findings are based on data provided by SRC.
- Acknowledged SRC staff: Mr. Arsen Sarikyan; Mr. Mkhitar Ayvazyan; Mr. Martin Sandoyan; Grigor Hakobyan.
- Report submission: Final version of the draft report submitted to Mr. Ashot H. Muradyan and Mr. Arsen Sarikyan on July 10, 2024.
- Report composition: Executive Summary; (I) Introduction; (II) Key Findings; (III) Next Steps.

### Executive Summary — predicted CIT gap for income year 2023
- Approach: Bottom-up using SRC operational audits (2020-2022) and CIT returns for 2023; prediction uses previously developed model.
- Predicted CIT gap (2023), including audit adjustments with immediate effect:
  - 121.2 billion AMD.
  - 34.1 percent of potential CIT liability.
  - 1.5 percent of GDP.
- Predicted CIT gap (2023), excluding audit adjustments with no immediate effect on revenue:
  - 80.1 billion AMD.
  - 25.5 percent of potential CIT liability.
  - 1.0 percent of GDP.
- Historical estimated CIT gaps (based on audits):
  - 2020: 38.8 percent of potential CIT liability; 83.1 billion AMD; 1.3 percent of GDP.
  - 2021: 35.2 percent of potential CIT liability; 90.3 billion AMD; 1.4 percent of GDP.
  - 2022: 32.8 percent of potential CIT liability; 108.5 billion AMD; 1.5 percent of GDP.
- Historical estimated CIT gaps excluding adjustments with no immediate effect on revenue:
  - 2020: 29.2 percent; 54.0 billion AMD; 0.9 percent of GDP.
  - 2021: 26.4 percent; 59.6 billion AMD; 0.9 percent of GDP.
  - 2022: 24.6 percent; 72.4 billion AMD; 1.0 percent of GDP.
- Observed pattern:
  - CIT gap in terms of potential CIT is trending downward when the largest corporations are excluded.
  - When considering all corporations, the predicted CIT gap shows a slight increase in 2023, suggesting the increase is driven by large corporations.
- Recommendation (headline):
  - Measure the CIT gap once audits for 2023 are completed (expected in 2025) and compare estimated gap to the predicted CIT gap for 2023.
  - SRC can use provided R code and instructions to conduct analysis internally or seek IMF capacity development for assisted self-assessment.

### Note on terminology from Executive Summary
- “Predicted”: used when CIT gap is based on CIT returns before audits.
- “Estimated”: used when CIT gap is based on audited CIT returns.
- Analysis measures the CIT compliance gap (policy gap excluded); "gap" and "compliance gap" used indistinguishably.
- Excluding audit adjustments with no immediate impact on revenue refers to cases where audit adjustment reduced a loss or carry forward of losses.
- Potential CIT liability is defined as self-reported CIT plus the estimated CIT gap.

### Recommendations (Executive Summary condensed)
- Estimate the CIT gap when audits are completed in 2025.
  - Responsible: SRC.
  - Method: SRC data analysts can use provided R code and instructions or seek IMF CD for assisted self-assessment.
  - Timing: May 2025.

### I. Introduction — model, methodology, and limitations
- Methodology:
  - Prediction based on previously developed model that estimated CIT gap for income years 2020-2022 using operational audits.
  - Model corrects for non-random selection of audited corporations (sample selection bias) using the Heckman two-step procedure.
    - First stage: probit model estimating probability of being audited using 25 separate risk scores used by SRC.
    - Second stage: models audit outcome (tax uncovered) using company characteristics and a regressor accounting for selection.
- Data used for prediction: non-audited CIT declarations for income year 2023 (declaration deadline April 20, 2024).
- Caveats and limitations:
  - Suitability of Heckman estimator depends on audit practices; narrowly focused audits may underestimate undetected noncompliance.
  - Audits targeting specific sectors or firm types increase extrapolation and uncertainty.
  - Some noncompliance may remain undetected, potentially underestimating the CIT gap.
  - Prediction assumes audit selection strategy remains unchanged and no significant tax regulation changes or new evasion methods occur.

### II. Key Findings — CIT gap results (detailed)
- Model basis:
  - Heckman model estimated using 4,432 comprehensive operational audits covering income years 2020–2022.
  - Results adjusted for sample selection bias.
- Main predicted outcomes for 2023 (repeated for clarity):
  - 121.2 billion AMD; 34.1 percent of potential CIT; 1.5 percent of GDP.
  - Excluding adjustments with no immediate effect on revenue: 80.1 billion AMD; 25.5 percent of potential CIT; 1.0 percent of GDP.
- Sectoral findings:
  - Predicted median CIT gap highest in sector labeled “Extraterritorial” (sector is small so contribution to total CIT gap is limited).
  - Second-highest median CIT gap is in the mining sector.
  - Lowest median CIT gap among top-10 sectors is in the accommodation sector.
  - Sector classification follows NACE coding; sector codes are self-reported by corporations; top-10 sectors displayed due to space limitations.
- Alternative model specifications and their 2023 predictions:
  - Excluding largest corporations:
    - Predicted CIT gap: 87.9 billion AMD; 27.2 percent of potential CIT; 1.1 percent of GDP.
    - Rationale: Top 0.1 percent largest companies (by turnover) excluded to reduce influence of outliers while accounting for most of the population.
  - Excluding largest corporations and excluding adjustments with no immediate effect on revenue:
    - Predicted CIT gap: 62.8 billion AMD; 21.1 percent of potential CIT; 0.8 percent of GDP.
- Interpretation:
  - Excluding largest corporations likely reduces noise from outliers and may provide a more accurate estimate for the remaining population, but reduces the population scope.
  - The increase in the predicted CIT gap for 2023 when including all corporations suggests large corporations drive the observed rise.

### III. Next Steps
A. Set up a team of data analysts and apply the developed model
- Recommendation:
  - SRC should appoint 2-3 data analysts responsible for compiling future CIT gap estimations.
  - Invest in data analytics capacity: at least 2-3 individuals trained in data analytics.
  - Appointed team should invest around 80 percent of their time on data analytics.
  - Team should learn necessary programming tools (R code and instructions provided) to carry out CIT gap analysis.

B. Recommendations (operational)
- SRC should carry out the CIT gap estimation after income year 2023 audits are completed in 2025 and compare to the predicted CIT gap for 2023 presented in this report.
- If SRC cannot perform the estimation in 2025, they could seek IMF assistance for an assisted self-assessment of the CIT gap to develop data analytical skills and move toward self-sufficiency.

*IMF Technical Assistance Report | Preface and Executive Summary (June 30 – July 04, 2024 mission to SRC).*

### Annex I. Supplementary material.

### Annex I. Supplementary material

### Methodology: Heckman 2-step estimator and application
- Purpose: Correct sample selection bias from non-random risk-based audits by estimating both the selection process and the level of non-compliance (tax uncovered from audit) in the same model.
- Model specification (as described):
  - Outcome equation: Yi = Xiβ + ui, where Yi measures tax uncovered from audit and Xi are company characteristics (lines from the CIT return, sector, number of employees etc.).
  - Selection equation: Si = 1[Ziδ + vi > 0], where Si = 1 denotes company i was audited, Zi are factors that determine audit selection (including 25 separate risk scores used by the SRC).
  - Assumptions: vi ~ N(0,1); E(ui | vi) = γ vi. The inverse Mills ratio λ(∙) = φ(∙)/Φ(∙) arises from vi ~ N(0,1).
- Estimation procedure:
  - Estimate selection equation (2) using a probit to obtain δ̂.
  - Compute λ(Zi δ̂) and regress Yi on Xi and λ(Zi δ̂) by OLS to obtain consistent estimates of β and γ (Heckman 2-step).
  - Predicted unconditional expectation of non-compliance: E[Yi | Xi]̂ = Xi β̂.
- Practical notes and identification:
  - Use separate variables where possible in the selection model to satisfy an exclusion restriction (a variable that affects audit selection but not non-compliance levels).
  - If the same variables are used in both models, identification relies on the non-linearity of the inverse Mills-ratio, risking unstable results from high multicollinearity.
  - In this application, 25 SRC risk scores were used to predict audits; regressions indicate 16 out of the 25 risk scores are not statistically significantly related to the tax uncovered from audit, which supports the presence of potential exclusion restrictions.
- Data handling and estimation details:
  - Data used: CIT returns and operational audits from 2020, 2021 and 2022.
  - Prices: main estimation deflated to 2020-prices; predictions for 2023 were requested in 2023-prices (non-deflated) and are presented separately.
  - Trimming: 0.1 pct. of audit results were trimmed in top and bottom for each year. Re-audits are discarded. Only corporations with a reported CIT return are included.
  - Standard errors: Standard errors are wrong when manually estimating the 2-step Heckman estimator; correct standard errors are obtained using bootstrap (bootstrap resampling done at the TIN-level).

### Estimation results and diagnostics (summary of Table A1)
- Sample sizes:
  - Selection model observations: 71,374
  - Outcome model observations: 4,324
- Goodness-of-fit:
  - Selection model 푹ퟐ (McFadden’s Pseudo 푹ퟐ): 0.26
  - Outcome model 푹ퟐ: 0.14
- Inverse Mills ratio (IMR):
  - IMR coefficient in outcome model: -1032,997* with Std. Error 578,905 (significance indicated: * denotes p < 0.01, ** p < 0.05, *** p < 0.10 per table note)
  - Interpretation: The significant IMR indicates the presence of sample selection bias.
- Notable covariate notes from regressions:
  - The selection model includes 25 risk scores; many risk scores have small coefficients (both positive and negative), and several are statistically significant at conventional levels.
  - The outcome model includes tax-line variables, number of employees, turnover per employee, total compensation, and sector dummies.
  - The outcome model R^2 of 0.14 reflects the challenge of explaining heterogeneous audit adjustments across diverse companies.

### Predicted CIT gap in 2023 (non-deflated) — Heckman model predictions (Table A2)
- Predicted CIT gap in 2023 (All, No loss, No large, No large / No loss) — values reported in 2023-prices:
  - All: 143.9 billion AMD
  - No loss: 95.1 billion AMD
  - No large: 104.3 billion AMD
  - No large / No loss: 74.5 billion AMD
- Corresponding shares:
  - Percent of potential CIT:
    - All: 34.1
    - No loss: 25.5
    - No large: 27.2
    - No large / No loss: 21.1
  - Percent of GDP:
    - All: 1.5
    - No loss: 1.0
    - No large: 1.1
    - No large / No loss: 0.8
- Notes on definitions and construction:
  - Predictions made using a Heckman model estimated on audits from 2020–2022.
  - “All” refers to the model using all remaining audits.
  - “No loss” excludes audit adjustments with no immediate impact on revenue.
  - “No large” excludes the top 0.1 percent of the largest companies, as measured by turnover.
  - Potential CIT liability is defined as self-reported CIT plus the estimated CIT gap.
  - The reported predictions for 2023 are in 2023-prices; the underlying model used data deflated to 2020-prices for estimation coherence.

### Key takeaways and practical implications
- Non-random risk-based audit samples generate sample selection bias; the Heckman 2-step estimator provides a consistent correction when selection and outcome processes are jointly modeled.
- The significant inverse Mills-ratio indicates selection bias is present in the audit data used.
- Prediction accuracy for non-compliance depends on how well the selection and outcome models explain the data; the outcome model’s 푹ퟐ = 0.14 highlights limits in explanatory power given company heterogeneity.
- Exclusion restrictions (variables that affect audit selection but not non-compliance) improve identification; the presence of many risk scores not significantly related to audit adjustments supports identification in this application. Changes to SRC targeting strategies could make finding valid exclusion restrictions more challenging.
- For inference, correct standard errors require bootstrap; reported standard errors in the outcome model are computed using bootstrap with TIN-level resampling.

*Source: IMF calculations based on data from SRC (Annex I. Supplementary material).*

---


_Source: https://www.imf.org/-/media/files/publications/tar/2024/english/tarea2024088-print-pdf.pdf_