## wp17255

## Source details

**Canonical URL:** [wp17255](https://www.imf.org/-/media/files/publications/wp/2017/wp17255.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2017/wp17255.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2017/wp17255.pdf.json)

---

### Background and motivation
- States that do not raise adequate revenues are unable to build institutions to support economic development (Besley and Persson, 2014).
- Once the tax-to-GDP ratio reaches around 15 percent, GDP per capita increases sharply (Gaspar, Jaramillo, and Wingender, 2016).
- Corruption is repeatedly identified as a factor behind lagging revenue mobilization.

### Data, scope, and novel contributions
- Comprehensive dataset for 147 countries spanning 1995-2014 (WoRLD).
- Disaggregated revenue data by: personal and corporate income, social security, property, goods and services, international trade.
- First use of IMF tools:
  - Revenue Administration–Fiscal Information Tool (RA-FIT)
  - Revenue Administration Gap Program (RA-GAP)
- Novel compilation of establishment dates for specialized taxpayer offices and anti-corruption agencies:
  - 109 LTOs
  - 62 MTOs
  - 77 ACAs
- Robustness tested against ICTD Government Revenue Dataset (GRD).

### Stylized facts — correlations between corruption, tax revenue, and compliance
- Total tax revenue negatively correlated with Transparency International Corruption index and World Bank Control of Corruption index.
- Disaggregated period-average correlations (1995–2014):
  - Individual income tax more strongly negatively correlated with corruption than corporate income tax.
  - Social security contributions negatively correlated with corruption.
  - International trade taxes negatively correlated with corruption.
- VAT-specific measures:
  - VAT C-efficiency (108/94 countries depending on panel): lower C-efficiency associated with higher perceived corruption.
  - RA-GAP VAT compliance gap (23 countries; Panel F uses 19 countries): higher perceived corruption associated with larger VAT compliance gaps.
  - Days to clear imports (119 countries): higher perceived corruption associated with more days to clear imports.
  - VAT refunds (102 countries): more corrupt countries face problems obtaining VAT refunds.
- Panel counts and spans:
  - Panel E – VAT C-Efficiency Ratio: 94 countries (data spans 1995–2014).
  - Panel F – Compliance Gap: 19 countries (4–8 observations per country in 2004–2013).
  - Panel G – Days to Clear Imports: 119 countries (1995–2014).
  - Panel H – VAT Refunds: 102 countries (1995–2014).

### Empirical methodology — identification and estimation
- Pooled IV framework estimated equation:
  - y_it = alpha_i + lambda_t + beta * Corruption_it + sum(theta_k * X_it_k) + epsilon_it
  - y_it: tax revenue-to-GDP ratio
  - Corruption_it: TI CPI or World Bank CCI, rescaled so higher = more corruption
  - X_it_k: trade openness, inflation, share of agriculture, GDP per capita PPP and squared
  - alpha_i, lambda_t: country and time fixed effects
- Instruments for corruption:
  - Ethnolinguistic fragmentation (Taylor and Hudson, 1972)
  - Religious fragmentation index (Mauro, 1995)
- Estimation choices:
  - Pooled IV used to avoid collinearity with fixed effects and instruments.
  - Corrections for autocorrelation and heteroskedasticity.
  - GMM estimators attempted but inconsistent due to over-identification.

### Empirical results — main quantitative findings
- Baseline pooled IV:
  - A one standard deviation (SD) increase in corruption perception corresponds on average to a 12 percentage points lower tax revenue-to-GDP ratio.
  - Example CPI averages reported: 26.3 (advanced), 62.5 (emerging), 74 (low income).
- Table 2 reported coefficients (IV, all countries; p-values shown in source):
  - TI Corruption Perception Index (CPI): -0.1246*** (0.000)
  - World Bank Control of Corruption Index (CCI): -0.1218*** (0.009)
  - Additional CPI column: -0.2175*** (0.023)
  - CCI coefficient in advanced-economies interacted column: -0.1664*** (0.022)
  - Observations: 1,934
  - Cragg-Donald statistics: 94.50; 106.3; 3.877; 2.536
- Heterogeneity:
  - Interaction with advanced-economy dummy: adverse effect of corruption on tax collection larger in advanced economies (coefficients negative and significant at the 1 percent level).
- Tax-component results (Table 3 highlights and textual interpretations):
  - PIT (personal income tax): an increase of one SD in corruption perception would decrease the PIT ratio by 0.2 percentage points per year.
  - Aggregate impact: corruption can reduce aggregate revenue performance by as much as 0.6 percent of GDP annually.
  - Component coefficients (WORLD data, TI Perception Index and WB Control of Corruption where reported):
    - Total Income: -0.0227*** ; -0.0236***
    - Individual Income: -0.0363*** ; -0.0386***
    - Corporate Income (CIT): 0.0136*** ; 0.0143*** (positive)
    - Tax on Trade: -0.0408*** ; -0.0384***
    - Tax on Goods and Services: -0.0272*** ; -0.0319***
    - Excise Tax: -0.0045 (insignificant) ; -0.0097***
    - VAT: -0.0340*** ; -0.0373***
    - Property Tax: 0.0023 (insignificant) ; 0.0016 (insignificant)
    - Social Security Contributions (SSCs): -0.0997*** (0.023) ; -0.1049*** (0.014)
  - Interpretations:
    - SSCs: "An increase of one SD in corruption would decrease SSCs by 0.5 percentage points per year."
    - International trade tax revenue: "An increase of one SD in corruption perception would decrease the international trade tax revenue by 0.2 percentage points per year."
    - Excise taxes tend to be unaffected.
    - Positive CIT association suggests substitution toward formal sector taxation in cross-country regressions.
- Appendix III (ICTD results) — selected coefficients and diagnostics (exact figures as reported):
  - First block examples:
    - Total Income: -0.0249*** (0.000); Observations 981; Cragg-Donald 43.45
    - Tax on Trade: -0.0272*** (0.000); Observations 1,033; Cragg-Donald 42.65
    - Social Security Contributions: -0.0612*** (0.023); Observations 1,014; Cragg-Donald 30.14
  - Second block examples:
    - Total Income: -0.0259*** (0.005); Observations 981; Cragg-Donald 49.61
    - Tax on Goods and Services: -0.0520*** (0.008); Observations 1,065; Cragg-Donald 32.25
    - Social Security Contributions: -0.0583*** (0.011); Observations 1,014; Cragg-Donald 37.84

### Institutional reforms, PSM and difference-in-difference findings
- Institutional dataset: establishment dates for LTOs, MTOs, ACAs (Appendix IV); first wave of LTOs in mid- to late 1990s; some countries lack establishments or have special cases (notes a–d).
- PSM methodology:
  - Propensity scores from Probit with development dummy, share of agriculture, real per capita GDP, openness, population, time; expanded baseline adds inflation, latitude, oil exporter dummy, total tax revenue/GDP, external debt/GDP.
  - Matching estimators: Stratification, Kernel, local linear regression. Common support enforced.
- PSM baseline ATEs (examples from Kernel, Log linear, Stratification methods):
  - Corruption (CPI change (T+3) - (T-1)):
    - LTO: ATE ≈ -2.87** (Kernel), -2.77*** (Log linear), -3.07*** (Stratification) — LTOs reduce perceived corruption (ATE about -3).
    - MTO, ACA: insignificant impacts on corruption perception in baseline.
  - Tax Revenue (change (T+2) - T):
    - LTO: ATE ≈ 0.006** (Kernel), 0.006** (Log linear), 0.007** (Stratification) — interpreted as about 0.6 percent of GDP increase in tax revenue post-LTO.
    - MTO, ACA: positive but insignificant impacts on tax revenue in baseline.
  - Example observation counts (Kernel method):
    - Corruption tests: Treatment Group = 32 (LTO), 12 (MTO), 41 (ACA); Control Group = 286, 432, 534.
    - Revenue tests: Treatment Group = 52 (LTO), 20 (MTO), 33 (ACA); Control Group = 492, 601, 534.
- Robustness (Table 5, Kernel examples):
  - Corruption ATEs: -2.1* (Expanded Baseline), -3.33*** (No Fragile States), -2.35*** (Full Sample)
  - Revenue ATEs: 0.007*** (Expanded Baseline), 0.006** (No Fragile States), 0.004** (Full Sample)
  - Interpretation: LTO effects on perceived corruption and tax revenue are robust across specifications; effects strengthen when fragile states are excluded.

### Mechanisms and interpretation
- Compliance channel: poor compliance mediates corruption’s adverse effect on revenue; evidence includes VAT compliance gaps, delays in VAT refunds, and days to clear imports correlating with perceived corruption.
- LTOs likely improve revenue by targeted taxpayer service and enforcement for large taxpayers, improving compliance and reducing perceived corruption.
- MTOs often newer and regionally organized; average tailored compliance programs less implemented — may explain insignificant average effects.
- Cross-country CIT positive association reconciled by institutional effects: LTO establishment can reduce perceived corruption and increase CIT and other tax revenues, even if cross-country regressions show higher CIT with higher perceived corruption.

### Policy-relevant implications and recommendations
- Perceived corruption meaningfully reduces tax revenue; aggregate impact up to 0.6 percent of GDP annually.
- VAT, taxes on income, international trade taxes, and social security contributions are particularly adversely affected.
- Strengthening institutions that improve tax compliance is a viable channel to mitigate corruption’s negative effect on revenue mobilization.
- Practical administrative reforms:
  - Establish specialized taxpayer offices (LTOs, and where appropriate MTOs) to focus on segment-specific compliance risks, build specialist skills, and improve service/monitoring.
  - Monitor large taxpayers as an important compliance strategy and a signal of enforcement commitment.
  - Use RA-GAP and RA-FIT diagnostics (VAT gaps, refund issues, days to clear imports) to guide targeted administrative fixes.

### Appendices — instruments, variance, and summary statistics (selected exact figures)
- Variance decomposition:
  - TI Perception Index: 95.77 (fraction of variance due to country fixed effects %)
  - WB Control of Corruption: 97.47
  - Ethnolinguistic fragmentation: 100.00
  - Religion Fragmentation: 100.00
  - Tax Revenue WoRLD: 93.99
  - Tax Revenue ICTD: 95.27
- First-stage instrument results (Appendix II):
  - Ethnolinguistic Fragmentation: 18.3044*** (TI Perception Index), 0.9831*** (WB Control of Corruption)
  - Religion Fragmentation: -9.8781***, -0.4166***
  - Observations: 1,934; R-squared: 0.672 (TI), 0.663 (WB)
  - Partial R-squared: 0.262 (TI), 0.275 (WB)
  - F(2, 1931) = 424.40; Prob > F = 0.0000
- Appendix V selected summary statistics (exact values):
  - Total Tax - WoRLD: Observations 1934; Mean 19.06776; Stadanrd Deviation 8.04348; Minimum 0.70334; Maximum 58.11479
  - Total Income Tax - WoRLD: Observations 1562; Mean 7.57159; Stadanrd Deviation 5.18073; Minimum 0.00253; Maximum 33.195
  - Individual Income Tax - WoRLD: Observations 1309; Mean 4.89952; Stadanrd Deviation 4.61933; Minimum 0.001; Maximum 27.674
  - Corporate Income Tax - WoRLD: Observations 1388; Mean 3.02241; Stadanrd Deviation 2.41465; Minimum 0.000555; Maximum 25.50648
  - Trade Taxes - WoRLD: Observations 1347; Mean 1.96975; Stadanrd Deviation 3.21967; Minimum 0.001; Maximum 39.90194
  - Excise Taxes - WoRLD: Observations 1451; Mean 2.29456; Stadanrd Deviation 1.44299; Minimum 0.00042; Maximum 21.88961
  - Tax on goods and services - WoRLD: Observations 1531; Mean 6.09664; Stadanrd Deviation 2.89537; Minimum 0.00242; Maximum 25.87794
  - VAT Tax - WoRLD: Observations 1310; Mean 6.17329; Stadanrd Deviation 2.34868; Minimum 1.34E-07; Maximum 19.20702
  - Propoerty Tax - WoRLD: Observations 1388; Mean 1.00759; Stadanrd Deviation 1.01473; Minimum 6.12E-07; Maximum 7.299
  - Social Security Contribution - WoRLD: Observations 1076; Mean 6.59905; Stadanrd Deviation 5.07973; Minimum 4.42E-06; Maximum 18.042
  - Corruption Perception Index - TI: Observations 1934; Mean 55.8348; Stadanrd Deviation 22.65745; Minimum 0; Maximum 92
  - Control of Corruption - WB: Observations 1934; Mean -0.1120269; Stadanrd Deviation 1.082339; Minimum -2.59; Maximum 1.64
  - PPP Per Capita: Observations 1934; Mean 9.104392; Stadanrd Deviation 1.176914; Minimum 6.097523; Maximum 11.916
  - Share of Agricuture: Observations 1934; Mean 12.29981; Stadanrd Deviation 11.88534; Minimum 0.03447; Maximum 58.20515
  - Inflation: Observations 1934; Mean -3.195819; Stadanrd Deviation 1.065253; Minimum -9.782633; Maximum 1.178743
  - Ethnic Fragmentation: Observations 1934; Mean 0.4220083; Stadanrd Deviation 0.2551487; Minimum 0; Maximum 0.9302
  - Religion Fragmentation: Observations 1934; Mean 0.4274414; Stadanrd Deviation 0.2360986; Minimum 0.0035; Maximum 0.8603
  - Trade Openness: Observations 1934; Mean 84.9161; Stadanrd Deviation 49.21485; Minimum 15.58034; Maximum 439.6567
  - Compliance Gap % of GDP: Observations 109; Mean 2.798532; Stadanrd Deviation 1.664916; Minimum 0.38; Maximum 6.88
  - C-Efficiency: Observations 1213; Mean 0.5020849; Stadanrd Deviation 0.1863975; Minimum 0.0611281; Maximum 1.193815

*Source: wp17255 (PDF chapter/section).*

### References ________________________________________________________________30

### wp17255 - References ________________________________________________________________30

### Background and motivation
- States that do not raise adequate revenues are unable to build institutions to support economic development (Besley and Persson, 2014).
- Gaspar, Jaramillo, and Wingender (2016) show that once the tax-to-GDP ratio reaches around 15 percent, GDP per capita increases sharply, a threshold that several developing countries have not yet crossed.
- The literature repeatedly and increasingly identifies corruption as one of the factors behind lagging revenue mobilization.

### Data, scope, and novel contributions
- Comprehensive dataset for 147 countries spanning 1995-2014, compiled by the IMF.
- Data are disaggregated by revenue type: taxes on personal and corporate income, social security, property, goods and services, and international trade.
- First paper to make use of data generated by two IMF tools:
  - Revenue Administration–Fiscal Information Tool (RA-FIT)
  - Revenue Administration Gap Program (RA-GAP)
- Novel compilation of a comprehensive dataset for the establishment of “specialized” taxpayer offices—both large and medium-sized—and anti-corruption agencies (LTO, MTO, and ACA establishment dates).
- Robustness of results is tested against another data source.

### Key findings and empirical results
- Corruption is negatively associated with overall tax revenue, and most of its components.
- The negative relationship is predominantly influenced by the way corruption interacts with tax compliance.
- When properly implemented, segmentation by taxpayers could lead to better focus on managing the unique compliance risks presented by the respective segments, thereby reducing the perception of corruption and increasing taxpayer compliance.
- Results indicate that strengthening institutions focusing on taxpayer segmentation can improve tax compliance.

### Implications for policy and institutional design
- Establishing specialized taxpayer offices and anti-corruption agencies can be part of an anti-corruption strategy (IMF 2016).
- Taxpayer segmentation—when properly implemented—offers a pathway to reduce perceived corruption and increase compliance by focusing enforcement and services on segment-specific risks.
- Strengthening institutions that improve tax compliance is a viable channel to mitigate the negative effect of corruption on revenue mobilization.

*Source: wp17255 - References ________________________________________________________________30*

### Section III describes a series of key stylized facts. The dataset is presented in the Section IV.

### wp17255 — Sections III–IV (stylized facts and dataset)

### Literature review — empirical regularities on corruption and taxation
- Widespread corruption harms the culture of compliance, thereby increasing tax evasion (Aghion et al., 2016). Tax exemptions perceived to be the result of a bribe undermine trust in government and compliance with tax laws (Dreher and Herzfeld 2005).
- Corruption fosters the development of the informal sector and erodes the potential tax base (Schneider, 2005; Schneider and Denste, 2000).
- Empirical studies mostly find a negative relationship between corruption and tax revenue (Sen Gupta, 2007; Besley and Persson, 2014).
- Indirect taxes requiring frequent interactions between taxpayers and authorities are more prone to corruption (Hwang 2002; Attila et al. 2008; Thornton 2008; Imam and Jacobs 2014). International trade taxes and taxes on goods and services (VAT and excise) are negatively and statistically significantly affected by corruption in several studies.
- Estimates for direct taxes are mixed: personal and corporate income taxes are often found insignificantly impacted by corruption in some studies (Attila et al. 2008; Thornton 2008), while Imam and Jacobs (2014) find a negative impact on corporate taxes in the Middle East.
- Revenue administration reforms that segment taxpayers (large, medium, small) and build specialized taxpayer offices can improve compliance, reduce perceptions of corruption, and raise tax yield. IMF (2015) lists effective monitoring of large taxpayers as one way to address compliance deficits.
- Evidence on large taxpayer offices (LTOs) is mixed: Ebrill et al. (2001) show better VAT performance with LTOs; Ebeke et al. (2016) find no significant direct impact on tax revenues but view LTOs as a pre-condition for other revenue-increasing measures. This paper extends the PSM methodology and expands the dataset to include MTOs and ACAs.

### Dataset — coverage and sources
- Primary revenue dataset: World Revenue Longitudinal Database (WoRLD), compiled at the IMF, covering 147 countries between 1995 and 2014. Revenue data are presented as a percentage of nominal GDP.
- Robustness check: International Centre for Tax and Development (ICTD) Government Revenue Dataset (GRD).
- Additional administrative data sources: IMF Revenue Administration Gap Program (RA-GAP) and Revenue Administration–Fiscal Information Tool (RA-FIT).
  - RA-GAP applied to 23 countries to evaluate the VAT potential-actual collection between 2003 and 2013.
  - RA-FIT includes over 140 countries with hundreds of variables on revenue collection services, staffing, registration, dispute resolution, auditing, verification, etc.
- LTOs, MTOs, ACAs dates of establishment identified for more than 100 countries:
  - 109 LTOs
  - 62 MTOs
  - 77 ACAs
- Corruption measures used:
  - Transparency International Corruption Perception Index (TI CPI)
  - World Bank Control of Corruption Index (CCI)
  - Both indices measure perceptions of corruption and are rescaled so higher values indicate greater corruption.
- Control variables used in empirical work (from World Bank Development Indicators, 1995–2014): trade openness (exports+imports as percent of GDP), inflation, share of agriculture, GDP per capita (PPP).

### Stylized facts — correlations between corruption, tax revenue, and compliance
- Total tax revenue is negatively correlated with the Transparency International Corruption index (and similar results with the World Bank Control of Corruption index).
- Disaggregated correlations (1995–2014, period averages):
  - Individual income tax is more strongly negatively correlated with corruption than corporate income tax.
    - Explanation offered: large corporate enterprises often fall under LTO supervision and are in the formal sector; discretion and perceptions of misuse are more pronounced for personal income tax, lowering compliance.
  - Social security contributions show a negative correlation with corruption (Panel C, Figure 2), attributed to informality and noncompliance among individuals and SMEs.
  - International trade taxes are negatively correlated with corruption, reflecting delays and bribe payments at the border.
- VAT-specific measures:
  - VAT C-efficiency ratio (actual VAT revenue relative to potential under a uniform rate) for 108 countries shows a strong correlation: lower C-efficiency associated with higher perceived corruption (Figure 3, Panel E).
  - RA-GAP VAT compliance gap (data from RA-GAP for 23 countries): Panel F shows a positive relationship between perceived corruption and VAT compliance gaps (data represents between 4 and 8 observations in 2004–2013 for each country; Panel F uses 19 countries).
  - Days to clear imports (119 countries, Panel G): countries with higher perceived corruption require more days to clear imports.
  - VAT refunds (102 countries, Panel H): negative relationship between corruption and VAT refunds — more corrupt countries face problems obtaining VAT refunds.
- Panel counts and data spans in figures:
  - Panel E – VAT C-Efficiency Ratio: 94 countries (data spans 1995–2014; points are period averages).
  - Panel F – Compliance Gap: 19 countries (data represents 4–8 observations per country in 2004–2013).
  - Panel G – Days to Clear Imports: 119 countries (data spans 1995–2014).
  - Panel H – VAT Refunds: 102 countries (data spans 1995–2014).

### Empirical methodology — equation and identification
- Estimated equation (pooled IV framework):
  - y_it = alpha_i + lambda_t + beta * Corruption_it + sum(theta_k * X_it_k) + epsilon_it
  - y_it: tax revenue-to-GDP ratio for country i at time t
  - Corruption_it: corruption measure (TI CPI or World Bank CCI), rescaled so higher = more corruption
  - X_it_k: control variables (trade openness, inflation, share of agriculture, GDP per capita PPP and squared)
  - alpha_i, lambda_t: country and time fixed effects
- Endogeneity concerns: potential reverse causality between corruption and taxation. The literature uses instrumental variables (IV) strategies.
  - This paper follows Hwang (2002), Attila et al. (2008), and Thornton (2008) using ethnolinguistic fragmentation (Taylor and Hudson, 1972) and an index capturing religious fragmentation (Mauro, 1995) as instruments.
- Estimation choices:
  - Pooled IV estimations are applied (pooled methods do not control for fixed effects but avoid collinearity between corruption indexes, IV variables, and country fixed effects).
  - All estimates correct for autocorrelation and heteroskedasticity.
  - GMM estimators were attempted for robustness but did not produce consistent estimates due to many instruments relative to cross-section (over-identification; Sargan-Hansen test probabilities high; probability of Hansen J-statistic test equals 1 in most cases following Roodman (2009) strategy).

### Empirical results — main quantitative findings
- Baseline pooled IV results (Table 2):
  - The relationship between corruption and tax revenue (percent of GDP) is negative and significant.
  - Estimated magnitude: an increase of one standard deviation (SD) in corruption perception corresponds on average to a 12 percentage points lower tax revenue-to-GDP ratio.
    - Note: the paper interprets one SD in corruption perception as the differential between relatively highly corrupt and relatively less corrupt countries; example averages: Corruption Perception Index average = 26.3 for advanced, 62.5 for emerging, and 74 for low income economies.
- Advanced economies vs developing economies:
  - Interaction of corruption with an advanced-economy dummy shows the adverse effect of corruption on tax collection is larger in advanced economies.
  - Coefficients associated with corruption are negative and statistically significant at the 1 percent level; interactive terms are significantly negative.
  - Interpretation: a given change in corruption is more detrimental in advanced economies, likely because advanced economies have lower baseline corruption.
- Tax-component results (Table 3 summary):
  - PIT (personal income tax), VAT, social security contributions, and trade taxes all show negative and statistically significant relationships with corruption (negative at the 1 percent level).
  - Magnitudes:
    - PIT: an increase of one SD in corruption perception or control of corruption would decrease the PIT ratio by 0.2 percentage points per year.
    - Table 2 coefficients for overall tax revenue specifications (IV, all countries):
      - TI Corruption Perception Index (CPI): -0.1246*** (p-value indicator shown as (0.000))
      - World Bank Control of Corruption Index (CCI): -0.1218*** (p-value shown as (0.009))
      - Additional column with CPI coefficient: -0.2175*** (p-value shown as (0.023))
      - CCI coefficient in advanced-economies interacted column: -0.1664*** (p-value shown as (0.022))
    - Observations: 1,934
    - Cragg-Donald statistics: 94.50; 106.3; 3.877; 2.536 (as reported in Table 2)
  - Control variables (baseline signs and significance):
    - Trade openness: positive coefficient but significance varies (positive and significant only when advanced economies dummy added in baseline).
    - Inflation: positive and significant in baseline (unanticipated inflation increases nominal tax revenue).
    - Share of agriculture: negative and significant in baseline.
    - Log PPPPC (per capita GDP PPP): positive and significant; squared term negative and significant (nonlinear effect).

### Policy-relevant implications and administrative reforms highlighted
- Corruption undermines tax collection across multiple tax bases, with particularly strong effects on VAT, PIT, social security, and trade taxes.
- Improving compliance and reducing perceptions of corruption can raise tax yields. Practical reforms discussed:
  - Establish specialized taxpayer offices (segmentation by size and possibly by economic sector) to address distinct compliance risks and build specialist skills, improving service, monitoring, and perceptions of corruption.
  - Focus on monitoring large taxpayers as an important compliance strategy and a signal of government commitment to enforcing tax laws (IMF 2015).
  - LTOs may improve VAT performance and support other revenue-enhancing policies, though empirical results on direct revenue impact are mixed.
- Administrative data tools (RA-GAP, RA-FIT) provide actionable diagnostics (VAT gaps, refund issues, days to clear imports) that correlate with perceived corruption and can guide targeted administrative fixes.

*Source: wp17255 — Sections III–IV.*

### 0.2 percentage points per year. Excise taxes tend to be unaffected. Social security

### wp17255 - 0.2 percentage points per year. Excise taxes tend to be unaffected. Social security

### Effects of corruption on tax revenue components
- Aggregate finding: Corruption has considerable impact on aggregate revenue performance—as much as 0.6 percent of GDP annually.
- Specific tax-component impacts (coefficients from Table 3, as reported):
  - Total Income: -0.0227*** (WORLD data, TI Perception Index), -0.0236*** (WB Control of Corruption)
  - Individual Income: -0.0363***, -0.0386***
  - Corporate Income (CIT): 0.0136***, 0.0143*** (positive effect noted)
  - Tax on Trade: -0.0408***, -0.0384***
  - Tax on Goods and Services: -0.0272***, -0.0319***
  - Excise Tax: -0.0045 (insignificant), -0.0097*** (second specification)
  - VAT: -0.0340***, -0.0373***
  - Property Tax: 0.0023 (insignificant), 0.0016 (insignificant)
  - Social Security Contributions (SSCs): -0.0997*** (0.023), -0.1049*** (0.014)
- Interpretation highlights:
  - SSCs are strongly affected: "An increase of one SD in corruption would decrease SSCs by 0.5 percentage points per year" (textual statement).
  - International trade tax revenue: "An increase of one SD in corruption perception would decrease the international trade tax revenue by 0.2 percentage points per year."
  - Excise taxes tend to be unaffected (text).
  - Property taxes show no significant impact; often levied sub-nationally where compliance may be weak.
  - Positive CIT coefficient: "An increase of one SD in corruption perception is associated with an increase in the CIT ratio by 0.1 percentage points per year" (textual statement), suggesting substitution of tax burden toward the formal sector.

### Compliance, institutions, and expected effects
- Institutional dataset and scope:
  - Collected dates of establishment for LTOs, MTOs and ACAs; database described as the largest of its kind (Appendix IV referenced).
  - First wave of LTO establishment in mid- to late 90s; minority of countries still have not established them.
- Expected institutional effects:
  - LTOs, MTOs, ACAs expected to dampen corruption and positively influence tax revenue via improved compliance.

### Propensity score matching (PSM) and difference-in-difference approach
- Methodology summary:
  - PSM (Rosenbaum and Rubin, 1983) used to estimate treatment effect of establishing institutions (LTO, MTO, ACA).
  - Propensity scores estimated with a Probit model using: a dummy for state of development, share of agriculture, real per capita GDP, openness, population, and time. Expanded baseline adds CPI inflation, latitude, dummy for oil exporters, total tax revenue as a share to GDP, and the ratio of external debt to GDP.
  - Matching estimators: Stratification, Kernel, local linear regression matching. Common support enforced; unmatched treated observations with propensity scores higher than max non-treated are left unmatched.
- Difference-in-difference specification (ATE formula provided in text) used to compare treated vs control over time and eliminate bias from common time trends.

### PSM baseline results (Table 4)
- Corruption Perception Index impact (Corruption Perception Index: (T+3) - (T-1)):
  - LTO: ATE ≈ -2.87** (Kernel), -2.77*** (Log linear), -3.07*** (Stratification) — LTOs lead to decline in perceived corruption (ATE about -3).
  - MTO and ACA: insignificant impacts on corruption perception in baseline.
- Tax revenue impact (Tax Revenue Change: (T+2) - T):
  - LTO: ATE ≈ 0.006** (Kernel), 0.006** (Log linear), 0.007** (Stratification) — interpreted as about 0.6 percent of GDP increase in tax revenue following LTO establishment.
  - MTO and ACA: positive but insignificant impacts on tax revenue in baseline.
- Observations (examples from Kernel method):
  - Corruption tests: Observations Treatment Group = 32 (LTO), 12 (MTO), 41 (ACA); Observations Control Group = 286, 432, 534 respectively.
  - Revenue tests: Observations Treatment Group = 52 (LTO), 20 (MTO), 33 (ACA); Observations Control Group = 492, 601, 534 respectively.

### Robustness of LTO impact on revenue and corruption (Table 5)
- Robustness tests:
  - Expanded baseline (adds CPI inflation, latitude, oil exporter dummy, total tax revenue/GDP, external debt/GDP)
  - No fragile states (deleting fragile states)
  - Full sample (do not drop observations after establishment)
- Key robust results (examples from Kernel method):
  - Corruption ATEs: -2.1* (Expanded Baseline), -3.33*** (No Fragile States), -2.35*** (Full Sample)
  - Revenue ATEs: 0.007*** (Expanded Baseline), 0.006** (No Fragile States), 0.004** (Full Sample)
- Interpretation: Effects on perceived corruption and tax revenue are consistent across specifications; perception declines and revenue increases significantly when fragile states are excluded; results hold for the entire sample including post-establishment years.

### Mechanisms and interpretation
- Compliance channel:
  - Strong correlation between VAT compliance and perceived corruption; higher perceived corruption associated with slower VAT refunds (impairing VAT functioning).
  - LTOs likely improve management of key taxpayer segments via targeted taxpayer service and improved enforcement programs, raising revenue.
- MTOs:
  - MTOs are newer and often regionally organized; tailored compliance programs less implemented on average—may explain insignificant average effects.
- Reconciliation with cross-country CIT result:
  - Cross-country regression shows higher perceived corruption associated with higher CIT revenues on average; however, LTO establishment can reduce perceived corruption and increase CIT and other tax revenues without altering that structural cross-country comparison.

### Policy conclusions (Section VII)
- Main conclusions:
  - Perceived corruption meaningfully reduces tax revenue; aggregate impact up to 0.6 percent of GDP annually.
  - VAT, taxes on income, international trade taxes, and social security contributions are adversely affected by corruption.
  - Poor compliance appears to mediate the effect of corruption on revenue performance.
  - LTOs reduce perceived corruption (ATE about -3 on CPI change measure) and enhance revenue performance (about 0.6 percent of GDP increase in tax revenue relative to non-LTO countries).
  - MTOs and ACAs do not show statistically significant effects on corruption or revenue performance in the presented specifications.

### Appendices — key instrument and variance findings
- Variance decomposition (Appendix I):
  - Corruption Index:
    - TI Perception Index: 95.77 (fraction of variance due to country fixed effects in %)
    - WB Control of Corruption: 97.47
  - Instruments:
    - Ethnolinguistic fragmentation: 100.00
    - Religion Fragmentation: 100.00
  - Tax Revenue:
    - WoRLD: 93.99
    - ICTD: 95.27
- First-stage instrument results (Appendix II):
  - Ethnolinguistic Fragmentation: 18.3044*** (TI Perception Index), 0.9831*** (WB Control of Corruption)
  - Religion Fragmentation: -9.8781***, -0.4166***
  - Observations: 1,934; R-squared: 0.672 (TI), 0.663 (WB)
  - Partial R-squared: 0.262 (TI), 0.275 (WB)
  - F(2, 1931) = 424.40; Prob > F = 0.0000
  - Note: "All instruments are highly significant in explaining the corruption index. Ethnolinguistic fragmentation has a positive and significant impact on corruption, whereas religious fragmentation has a negative and significant effect. The partial R2 of the excluded instruments is between 0.26 and 0.28. Moreover, the F-statistics of the first stage are significant at 1 percent."

*Source: wp17255 (PDF chapter/section).*

### Appendix III. Results with ICTD

### Appendix III. Results with ICTD

### Regression results: Corruption Index and Tax Types
- Data sources noted: ICTD Data; Transparency International Corruption Perception Index; World Bank Control of Corruption Index.
- Regression coefficients, significance (p-values in parentheses), observations, and Cragg-Donald statistic (where reported) for tax categories (first block and second block reported separately in the source):

First reported block
- Total Income: Coefficeint -0.0249*** (0.000); Observations 981; Cragg-Donald 43.45
- Individual Income: Coefficeint -0.0136 (0.000); Observations 913; Cragg-Donald 16.98
- Corporate Income: Coefficeint 0.0023 (0.000); Observations 866; Cragg-Donald 18.92
- Tax on Trade: Coefficeint -0.0272*** (0.000); Observations 1,033; Cragg-Donald 42.65
- Tax on Goods and Services: Coefficeint -0.0363*** (0.000); Observations 1,065; Cragg-Donald 29.82
- Excise Tax: Coefficeint -0.0045* (0.000); Observations 981; Cragg-Donald 26.64
- Indirect Taxes: Coefficeint -0.0748*** (0.023); Observations 1,031; Cragg-Donald 38.83
- Direct Taxes: Coefficeint -0.0385*** (0.000); Observations 1,036; Cragg-Donald 40.97
- Property Tax: Coefficeint -0.0045** (0.000); Observations 944; Cragg-Donald 26.94
- Social Security Contributions: Coefficeint -0.0612*** (0.023); Observations 1,014; Cragg-Donald 30.14

Second reported block
- Total Income: Coefficeint -0.0259*** (0.005); Observations 981; Cragg-Donald 49.61
- Individual Income: Coefficeint -0.0182** (0.009); Observations 913; Cragg-Donald 18.12
- Corporate Income: Coefficeint 0.0065 (0.004); Observations 866; Cragg-Donald 20.48
- Tax on Trade: Coefficeint -0.0273*** (0.010); Observations 1,033; Cragg-Donald 52.41
- Tax on Goods and Services: Coefficeint -0.0520*** (0.008); Observations 1,065; Cragg-Donald 32.25
- Excise Tax: Coefficeint -0.0126*** (0.003); Observations 981; Cragg-Donald 31.63
- Indirect Taxes: Coefficeint -0.0886*** (0.011); Observations 1,031; Cragg-Donald 45.02
- Direct Taxes: Coefficeint -0.0390*** (0.006); Observations 1,036; Cragg-Donald 47.34
- Property Tax: Coefficeint -0.0029 (0.002); Observations 944; Cragg-Donald 31.87
- Social Security Contributions: Coefficeint -0.0583*** (0.011); Observations 1,014; Cragg-Donald 37.84

### Implicit findings from reported coefficients
- Multiple tax categories show negative coefficients with respect to corruption indices (e.g., Total Income, Tax on Trade, Tax on Goods and Services, Indirect Taxes, Direct Taxes, Social Security Contributions) with varying levels of statistical significance as reported.
- Some corporate income coefficients are small and positive in the reported blocks (0.0023 and 0.0065) with p-values shown in parentheses as (0.000) and (0.004) respectively.
- Cragg-Donald statistics reported for each regression range from 16.98 to 52.41 indicating instrument strength diagnostics for the IV estimations as presented in the source.

### Appendix IV: LTO, MTO, and ACA Establishment Dates (selected entries)
- The source provides country-level indicators of whether a Large Taxpayer Office (LTO), Medium Taxpayer Office (MTO), and Anti-Corruption Agency (ACA) exist and the year since established. Examples included in the appendix:
  - Afghanistan: LTO Since Y 2005; MTO Since Y 2010; Anti-Corruption Agency Since Y 2008
  - Albania: LTO Since Y 1998; MTO Since N; Anti-Corruption Agency Since N
  - Argentina: LTO Since Y 1997; MTO Since Y 1997; Anti-Corruption Agency Since Y 1999
  - Bangladesh: LTO Since Y 2003; MTO Since N; Anti-Corruption Agency Since Y 2004
  - Brazil: LTO Since Y 2010; MTO Since N; Anti-Corruption Agency Since N
  - China: LTO Since N; MTO Since N; Anti-Corruption Agency Since Y 1978
  - Kenya: entries include Y (c) 2006 and Y (c) 1998 with explanatory notes (see notes below)
  - United States: LTO Since Y 1975
  - United Kingdom: Anti-Corruption Agency Since Y 1983
- Notes provided in the source explaining special cases:
  - (a) Georgia - The LTO was dismantled in 2010.
  - (b) Iraq - An LTO does exist, but it is staffed with 4 people only. We therefore classify it as not-functional.
  - (c) Kenya - Two dates are presented, because LTO was formed as an operations unit in 1998 to provide one-stop shop services in the administration of income tax and VAT matters affecting large taxpayers. The LTO was developed into a fully-fledged department in the year 2006 with the sole purpose of administering domestic tax matters affecting large taxpayers. Both dates are kept due to a likely impact from both reforms.
  - (d) Uganda - The LTO was closed in 2001 and re-opened in 2004.

### Appendix V: Summary Statistics (variables with Observations, Mean, Stadanrd Deviation, Minimum, Maximum)
- Total Tax - WoRLD: Observations 1934; Mean 19.06776; Stadanrd Deviation 8.04348; Minimum 0.70334; Maximum 58.11479
- Total Income Tax - WoRLD: Observations 1562; Mean 7.57159; Stadanrd Deviation 5.18073; Minimum 0.00253; Maximum 33.195
- Individual Income Tax - WoRLD: Observations 1309; Mean 4.89952; Stadanrd Deviation 4.61933; Minimum 0.001; Maximum 27.674
- Corporate Income Tax - WoRLD: Observations 1388; Mean 3.02241; Stadanrd Deviation 2.41465; Minimum 0.000555; Maximum 25.50648
- Trade Taxes - WoRLD: Observations 1347; Mean 1.96975; Stadanrd Deviation 3.21967; Minimum 0.001; Maximum 39.90194
- Excise Taxes - WoRLD: Observations 1451; Mean 2.29456; Stadanrd Deviation 1.44299; Minimum 0.00042; Maximum 21.88961
- Tax on goods and services - WoRLD: Observations 1531; Mean 6.09664; Stadanrd Deviation 2.89537; Minimum 0.00242; Maximum 25.87794
- VAT Tax - WoRLD: Observations 1310; Mean 6.17329; Stadanrd Deviation 2.34868; Minimum 1.34E-07; Maximum 19.20702
- Propoerty Tax - WoRLD: Observations 1388; Mean 1.00759; Stadanrd Deviation 1.01473; Minimum 6.12E-07; Maximum 7.299
- Social Security Contribution - WoRLD: Observations 1076; Mean 6.59905; Stadanrd Deviation 5.07973; Minimum 4.42E-06; Maximum 18.042
- Corruption Perception Index - TI: Observations 1934; Mean 55.8348; Stadanrd Deviation 22.65745; Minimum 0; Maximum 92
- Control of Corruption - WB: Observations 1934; Mean -0.1120269; Stadanrd Deviation 1.082339; Minimum -2.59; Maximum 1.64
- PPP Per Capita: Observations 1934; Mean 9.104392; Stadanrd Deviation 1.176914; Minimum 6.097523; Maximum 11.916
- Share of Agricuture: Observations 1934; Mean 12.29981; Stadanrd Deviation 11.88534; Minimum 0.03447; Maximum 58.20515
- Inflation: Observations 1934; Mean -3.195819; Stadanrd Deviation 1.065253; Minimum -9.782633; Maximum 1.178743
- Ethnic Fragmentation: Observations 1934; Mean 0.4220083; Stadanrd Deviation 0.2551487; Minimum 0; Maximum 0.9302
- Religion Fragmentation: Observations 1934; Mean 0.4274414; Stadanrd Deviation 0.2360986; Minimum 0.0035; Maximum 0.8603
- Trade Openness: Observations 1934; Mean 84.9161; Stadanrd Deviation 49.21485; Minimum 15.58034; Maximum 439.6567
- Corruption Index - ICRG: Observations 1651; Mean 0.5241551; Stadanrd Deviation 0.2138007; Minimum 0; Maximum 1
- Compliance Gap % of GDP: Observations 109; Mean 2.798532; Stadanrd Deviation 1.664916; Minimum 0.38; Maximum 6.88
- C-Efficiency: Observations 1213; Mean 0.5020849; Stadanrd Deviation 0.1863975; Minimum 0.0611281; Maximum 1.193815

*Appendix III. Results with ICTD, wp17255 (source content).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2017/wp17255.pdf_