## Appendix I. Data Descriptions and Sources

## Source details

**Canonical URL:** [Appendix I. Data Descriptions and Sources](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025219-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025219-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025219-source-pdf.pdf.json)

---

### Major themes and objectives
- Tax revenue mobilization remains a central challenge; in many developing countries tax revenue persists below 15 percent of GDP.
- Key objective: improve accuracy of estimated revenue gains from tax administration reform by addressing empirical challenges (measurement, simultaneity, omitted variable bias, time-variation).
- Five methodological contributions:
  - IV strategy using intensity of the IMF’s Fiscal Affairs Department (FAD) Capacity Development (CD) programs in tax administration (measured by cumulative Full-Time Equivalent (FTE) over the past four or five to two years prior to observation) to address simultaneity bias.
  - Unique survey of 30 tax administration experts from FAD using a Delphi method to improve measures of tax administration strength.
  - Exploitation of within-country variation via fixed effects models to relax time-invariance of tax administration strength.
  - Improved control for tax policy changes using the Global Tax Policy and Revenue Evaluation Database (TAPRED) built with a narrative approach and large language models (LLM).
  - Robustness checks and sensitivity analyses addressing measurement error in the Operational Strength Index (OSI) and heterogeneous effects by development level, informality, financial development, institutions, governance quality, and tax types.

### Data sources, OSI construction, and expert weighting
- Primary source for tax administration features: International Survey on Revenue Administration (ISORA) vintages used: ISORA 2016, ISORA 2018, and ISORA 2023.
  - ISORA vintages cover years: ISORA 2016 (2014–15), ISORA 2018 (2016–17), ISORA 2020 (2018–19), ISORA 2021 (2020), ISORA 2022 (2021), ISORA 2023 (2022).
  - ISORA 2020 and ISORA 2021 vintages excluded due to major revision; ISORA 2023 follows same structure as first two vintages.
  - Coverage includes 38 AEs, 78 EMs, and 50 LICs; missing survey responses filled using closest available vintage.
  - ISORA latest vintage contains 606 questions; data collected via voluntary self-assessment (online questionnaire).
- TAPRED (Global Tax Policy and Revenue Evaluation Database):
  - Collects tax policy changes and revenue yields from 5,200 IMF Staff Reports and MEFPs across 189 countries (1998–2024).
  - Captures attributes: type, timing, base vs. rate, rationale, implementation status, estimated revenue impact (in percent of GDP), direction, one-off nature, and whether based on forecasts.
- Delphi method and expert weighting:
  - Iterative three-step Delphi with a group of five IMF tax administration experts to select and classify ISORA indicators and design an expert weighting survey.
  - Pilot Tax Administration Yield and Assessment Tool (TAYAT) in eight countries by 13 IMF experts; pilot indicated need for greater granularity.
  - Expert weighting survey administered to 30 expert respondents; experts distributed 100 points among three layers (nine indices; sub-categories; sub-sub-categories).
  - Global weights obtained by multiplying layer 1–3 weights produce relative importance for each of the 193 questions; average layer 1 weights reported in Figure 1.
  - Average layer 1 weights: highest-rated practices CRM, UTD, and DIG; followed by ENF, LTO and HNWI, SOR, HRM, AUT, and PAC.
  - Response rate on larger expert survey: 42 percent (pilot by five experts).

### Operational Strength Index (OSI) structure and data adjustments
- OSI composed of nine sub-indices:
  1. Compliance Risk Management (CRM)
  2. Use of Third-Party Data (UTD)
  3. Degree of Digitalization (DIG)
  4. Service Orientation (SOR)
  5. Public Accountability (PAC)
  6. Autonomy (AUT)
  7. Large Taxpayers Office and High-Net-Worth Individuals (LTO and HNWI)
  8. Tax Enforcement (ENF)
  9. Human Resources Management and Development (HRM)
- New approach uses 193 questions (versus 66 in prior studies), structured into:
  - 41 disaggregated indices (sub-categories)
  - 152 more detailed indices (sub-subcategories)
- Expert weighting survey specifics:
  - 30 expert respondents; global weights formed by multiplying layer 1 × layer 2 × layer 3 weights to produce weights for each of the 193 questions.
  - Full disaggregated global weights used to construct OSI; only average weights at layer 1 presented in main figures.
- Data adjustments: for questions absent in 2016 and 2017 vintages, gaps filled using 2023 ISORA vintage; ISORA 2023 introduced stricter verification and validation reducing potential overreporting in earlier surveys.

### Dependent variable and controls
- Dependent variable: Tax-to-GDP ratio, excluding trade taxes and social security contributions; compiled from the IMF’s WoRLD and WEO databases.
- Controls included:
  - Macroeconomic: per capita GDP and its square, inflation, trade openness, external debt to GDP, terms of trade, real GDP growth (lagged).
  - Structure/institutions: oil exports to GDP, share of agriculture to GDP, control of corruption.
  - Novel control: revenue yields from tax policy changes from TAPRED (expressed in percent of GDP).

### Selected descriptive statistics (preserve reported values)
- Non-Trade Tax Revenue (Percent of GDP), consolidated WoRLD and WEO:
  - Obs.: 901; Mean: 0.152; Median: 0.142; Std.Dev.: 0.079; Min: 0.002; Max: 0.496
- Operational Strength Index (OSI), Expert Weights:
  - Obs.: 741; Mean: 0.609; Median: 0.627; Std.Dev.: 0.150; Min: 0.075; Max: 0.879
- Selected sub-indices (Obs.: 747 unless noted):
  - CRM: Mean: 0.627; Median: 0.649; Std.Dev.: 0.191; Min: 0.000; Max: 0.967
  - UTD: Mean: 0.409; Median: 0.385; Std.Dev.: 0.266; Min: 0.000; Max: 1.000
  - DIG: Mean: 0.659; Median: 0.781; Std.Dev.: 0.294; Min: 0.000; Max: 1.000
  - SOR: Mean: 0.693; Median: 0.733; Std.Dev.: 0.217; Min: 0.000; Max: 1.000
  - PAC: Mean: 0.564; Median: 0.561; Std.Dev.: 0.179; Min: 0.000; Max: 1.000
  - AUT: Mean: 0.718; Median: 0.772; Std.Dev.: 0.247; Min: 0.000; Max: 1.000
  - LTO: Mean: 0.584; Median: 0.622; Std.Dev.: 0.243; Min: 0.000; Max: 1.000
  - ENF: Mean: 0.575; Median: 0.571; Std.Dev.: 0.160; Min: 0.000; Max: 0.914
  - HRM: Mean: 0.665; Median: 0.696; Std.Dev.: 0.296; Min: 0.000; Max: 1.000
- IMF CD intensity and FTE measures (Obs.: 960):
  - FTE used in IMF tax administration CD (in years): Mean: 0.199; Median: 0.004; Std.Dev.: 0.403; Min: 0.000; Max: 4.476
  - Cumulative FTE used in IMF tax admin CD over 5 to 2 years earlier (in years): Mean: 0.644; Median: 0.311; Std.Dev.: 1.065; Min: 0.000; Max: 10.660

### Stylized facts on OSI dynamics and cross-group patterns
- ISORA vintages cover 2014, 2015, 2016, 2017, and 2022.
- Cross-country/temporal patterns:
  - EMs and LICs: OSI has significantly improved over time, starting from very low initial OSI in 2014; many EMs and LICs have 2022 OSI above 2014 OSI indicating convergence.
  - AEs and EMs: limited improvement or slight decline in 2022.
  - LICs: steady strengthening over the past decade; since 2016 distribution of OSI (median, 25th, 75th percentiles) shifted upward, including post-COVID-19 period.
- Drivers:
  - Digitalization (DIG) main driver of improvement in EMs and LICs.
  - Use of third-party data (UTD) slightly improved in LICs but remains low.
  - SOR, PAC, HRM improved slightly.
- Data-quality caveat: ISORA 2023 shows noticeable declines in service orientation, HRM, PAC, and ENF likely reflecting stricter verification and reduced overreporting in earlier rounds.

### Correlations and persistence of OSI over time
- Correlation matrix (Operational Strength Index across years):
  - OSI2014–OSI2015: 0.997
  - OSI2014–OSI2016: 0.961
  - OSI2014–OSI2017: 0.956
  - OSI2014–OSI2022: 0.893
  - OSI2015–OSI2016: 0.961
  - OSI2015–OSI2017: 0.956
  - OSI2015–OSI2022: 0.893
  - OSI2016–OSI2017: 0.997
  - OSI2016–OSI2022: 0.916
  - OSI2017–OSI2022: 0.915
- Implication: strong temporal stability of OSI with modestly lower correlation between 2014–2017 block and 2022.

### Empirical identification and estimation approach (data-related points)
- IV constructed from lagged cumulative FAD CD FTE in tax administration over two to five years earlier to instrument OSI.
- Fixed effects models exploit within-country variation and allow OSI to vary over time.
- TAPRED used to control for tax policy changes with detailed estimated revenue impacts.
- Key empirical challenges:
  1. Limited time coverage: panel restricted to five years (2014–2017, 2022) with a significant gap between 2017 and 2022.
  2. Limited within-country variation: OSI is slow-moving (see Appendix Table AII.2).
  3. Simultaneity bias: tax revenue and reforms may be endogenous.
  4. Omitted variable bias: unobserved factors (tax morale, geography, historical legacies, social trust).
  5. Measurement error: ISORA self-reported data may overstate administration strength, biasing estimates downward.

### Key empirical findings (estimates and diagnostics preserved exactly)
- IV estimates: increasing OSI from the 33rd to 67th percentile raises tax-to-GDP by 1.2 percentage points in the full sample and 1.7 percentage points in EMDEs; stated as enough to close 25–35 percent of the total tax gap identified by Benitez et al. (2023).
- Illustration: moving from the 33rd to the 67th percentile implies a 0.14 increase of the index; between 2017 and 2022 the median change was near zero and the 75th percentile only 0.04; a 0.14 rise corresponds to the 95th percentile of observed changes.
- First-stage (instrument → OSI) point estimates (Table 2 summary):
  - Column (1) 0.0172*** (0.0037)
  - Column (2) 0.0182*** (0.0053)
  - Column (3) 0.0150*** (0.0032)
  - Column (4) 0.0158*** (0.0044)
  - Interpretation: a one-year increase in cumulative FAD CD intensity increases OSI by up to 0.015–0.018.
  - Within R-squared for first stage: 0.3343 (Col 1) to 0.3637 (Col 4).
- Baseline IV coefficients on OSI (Table 1 and robustness tables):
  - Operational Strength Index [0,1]:
    - Column (3) 0.1015*** (0.0059)
    - Column (4) 0.1428*** (0.0100)
    - Column (5) 0.0821*** (0.0053)
    - Column (6) 0.1177*** (0.0183)
  - Alternate table highlights:
    - IV All [-4 to -2]: 0.0859*** (0.0059); IV EMDEs [-4 to -2]: 0.1236*** (0.0127); IV All [-5 to -2]: 0.0692*** (0.0050); IV EMDEs [-5 to -2]: 0.1026*** (0.0209) (Appendix Table AIII.5).
    - Robustness tables report ranges such as 0.082 to 0.102 (full sample) and 0.118 to 0.143 (EMDEs).
- First-stage and instrument diagnostics (selected reported values):
  - Cragg-Donald Wald F statistic: 36.230 (Col 3); 24.179 (Col 4); 43.229 (Col 5); 29.060 (Col 6).
  - Kleibergen-Paap rk Wald F statistic: 10.769 (Col 3); 11.742 (Col 4); 12.843 (Col 5); 14.545 (Col 6).
  - Ho: OSI is exogenous — rejected (reported as 0 for IV columns).
- Tax policy yield effect:
  - Tax Policy Yield, exc. Trade Tax and SSC: ranges 0.3020*** (0.0408) to 0.3196*** (0.0122); Column (3) 0.3320*** (0.0084); typical IV estimates ~0.32 ppt of GDP increase in actual tax revenue per 1 ppt GDP expected revenue from policy changes.
- Staffing and nonlinear effects:
  - #Tax Staff/LaborForce: Column (3) 0.5339*** (0.0812); Column (4) 0.5943*** (0.1028).
  - Sq(#Tax Staff/Labor Force): Column (3) -2.2606*** (0.3449); Column (4) -2.5973*** (0.4755).
- Other control variable highlights (exact reported signs and approximate magnitudes preserved):
  - Real GDP growth, lagged: coefficients ~0.0006–0.0008***.
  - Trade openness (% of GDP), lagged: coefficients ~0.0169–0.0224*** with SEs (0.0032)–(0.0055).
  - Inflation, lagged: negative and significant ~ -0.0249*** to -0.0328***.
  - Oil exports (% of GDP), lagged: negative and significant ~ -0.0948*** to -0.1415***.
- Sample sizes and model fit:
  - Observations: 529 (full sample) and 424 (EMDEs) in baseline.
  - Number of countries: 121 (full sample) and 99 (EMDEs).
  - AIC and BIC: lower for lag window [-5 to -2] (Columns 5 and 6) than for [-4 to -2] (Columns 3 and 4); reported values include AIC -3490.8 (Col 5), -3472.9 (Col 3); BIC -3473.7 (Col 5), -3455.8 (Col 3).

### Sensitivity and robustness highlights
- Alternative lag structures of the instrument:
  - Tested (-5/-4 to -1) and (-5/-4 to -3); IV estimates strong and significant across lag structures.
  - (-5/-4 to -1) yields largest estimates; (-5/-4 to -3) slightly smaller; baseline balances relevance and exogeneity.
- Alternative OSI measures:
  - Lagged OSI and equal-weighted OSI tested.
  - Lagged OSI has larger positive effect; equal-weighted OSI yields higher coefficients than expert-weighted OSI.
  - Baseline (contemporaneous, expert-weighted OSI) preferred by AIC/BIC.
- Income Group × Year fixed effects:
  - When included, IV OSI coefficients decrease in magnitude but remain statistically significant.
- Alternative samples (excluding outliers, FCS, small islands, resource-rich, top/least reformers):
  - Baseline results remain consistent and significant at 1 percent level.
  - Excluding top 10 percent OSI, bottom 10 percent tax-to-GDP, FCS, resource-rich, or top/least reformers generally strengthens OSI–tax revenue relationship.
  - Excluding highest tax-to-GDP, lowest OSI, and small islands yields coefficients slightly smaller or comparable to baseline IV.
- Alternative control variables and added controls (informality, urbanization, age dependency, education and health spending, financial development):
  - OSI positive and significant in most specifications.
  - Informality and education/health spending generally no effect; urbanization and age dependency ratio show tax revenue decreases with higher values.
  - Financial development unexpectedly negatively associated with tax revenue in several specifications.

### Sensitivity: heterogeneous effects and mechanisms (preserved numeric thresholds)
- LICs vs EMs:
  - Revenue gains from strengthening tax administration are more than twice as large in EMs compared to LICs (Appendix Table AIII.25).
- Informality interactions:
  - Interaction OSI × Informality negative and significant at 5 percent in some columns.
  - Quantitative thresholds: when self-employment exceeds 48 and 75 percent of total employment (top 30 and 14 percent of countries with highest self-employment), countries experience lower revenue gains relative to baseline for full and EMDE samples, respectively (Appendix Table AIII.26).
- Financial development interactions:
  - Interaction between OSI and financial development significant and large.
  - Thresholds: 60 percent of countries in full sample (financial development index below 0.32) experience lower revenue gains than baseline; 40 percent of countries in EMDE sample (index below 0.17) experience lower gains than baseline (Appendix Table AIII.27).
- Institutions and governance:
  - Interaction OSI × governance indicators positive and significant; depending on indicator between 0 and 17 percent of countries may experience negative revenue impacts despite OSI improvements, and between 21 and 58 percent experience less-than-baseline gains due to weaker institutions (Appendix Tables AIII.28–AIII.33).
- Differentiated effects by tax type (Section 5):
  - Largest gains for taxes on sales and production, including VAT.
  - Property taxes benefit modestly; income and profit taxes increase less.
  - Taxes not elsewhere classified and non-tax revenue: no significant effect (coefficients statistically insignificant at 10 percent).

### Policy implications (drawn from empirical results and mechanisms)
- Strengthening tax administration significantly increases tax revenue; IV estimates indicate economically meaningful gains, especially for EMDEs.
- Sustained, long-term capacity development (CD) efforts (IMF FAD and development partners) are important: cumulative FTE CD intensity predicts OSI improvements and translates into revenue gains.
- Policy design:
  - Integrate tax administration reforms with tax policy, financial sector development, and institutional/legal reforms to maximize revenue gains.
  - Target indirect tax administration (VAT and sales/production taxes) where administration improvements yield the largest returns.
  - Account for country-specific constraints: informality, weak governance, low financial development, and resource dependence attenuate revenue gains and warrant complementary reforms (formalization policies, governance strengthening, financial sector deepening).
- Monitoring and measurement:
  - Regular evaluations (ISORA, TADAT) and improved narrative/LLM-based tax policy tracking (TAPRED) can better distinguish tax policy impacts from administration effects and improve implementation monitoring.

*Italic: Source — Appendix I. Data Descriptions and Sources, wpiea2025219-source-pdf*

### Appendix I. Data Descriptions and Sources ..............................................................................

### Appendix I. Data Descriptions and Sources

### Major themes and objectives
- Tax revenue mobilization remains a central challenge; in many developing countries tax revenue persists below 15 percent of GDP.
- The paper’s key objective is to improve accuracy of estimated revenue gains from tax administration reform by addressing empirical challenges (measurement, simultaneity, omitted variable bias, time-variation).
- Five methodological contributions:
  - Use of an IV strategy with the intensity of the IMF’s Fiscal Affairs Department (FAD) Capacity Development (CD) programs in tax administration (measured by cumulative Full-Time Equivalent (FTE) over the past four or five to two years prior to observation) to address simultaneity bias.
  - Development and administration of a unique survey of 30 tax administration experts from FAD to improve measures of tax administration strength using a Delphi method.
  - Exploiting within-country variation via fixed effects models to relax the assumption that tax administration strength is time-invariant.
  - Improved control for tax policy changes using the Global Tax Policy and Revenue Evaluation Database (TAPRED), built using a narrative approach and large language models (LLM).
  - Robustness checks and sensitivity analyses addressing measurement error in the Operational Strength Index (OSI) and heterogeneous effects by development level, informality, financial development, institutions, governance quality, and tax types.

### Key empirical findings (as reported)
- IV estimates: increasing OSI from the 33rd to 67th percentile raises the tax-to-GDP ratio by 1.2 percentage points in the full sample and 1.7 percentage points in EMDEs.
  - This is stated to be enough to close 25–35 percent of the total tax gap identified by Benitez et al. (2023).
- Illustration of OSI change: moving from the 33rd to the 67th percentile implies a 0.14 increase of the index; between 2017 and 2022 the median change was near zero and the 75th percentile only 0.04; a 0.14 rise corresponds to the 95th percentile of observed changes.
- Intensity of past FAD CD in tax administration is a strong predictor of current OSI.
- Control variables with expected strong relationships to tax revenue include: tax policy changes, tax staff, and macro and structural factors.
- Sensitivity findings:
  - Revenue gains from strengthening tax administration are more than twice as large in Emerging Markets (EMs) compared to Low-income countries (LICs).
  - Gains are lower in Advanced Economies (AEs).
  - Revenue gains increase as informality declines.
  - Revenue gains are mediated by financial development.
  - Revenue gains increase with institutions and governance quality.
  - The largest gains are observed in indirect taxation; direct taxes and non-tax revenue show more muted responses.
- Robustness: findings are robust to alternative OSI measures (including lagged and equal-weighted OSI), income group-year fixed effects, alternative samples excluding outliers and specific country groups (fragile and conflict-affected states, small island countries, resource-rich countries), and alternative specifications of control variables.

### Data sources and construction of tax administration strength measures
- Primary source for tax administration features: International Survey on Revenue Administration (ISORA) vintages used: ISORA 2016, ISORA 2018, and ISORA 2023.
  - Note: ISORA vintages cover years as follows: ISORA 2016 (2014–15), ISORA 2018 (2016–17), ISORA 2020 (2018–19), ISORA 2021 (2020), ISORA 2022 (2021), ISORA 2023 (2022).
  - The analysis excludes ISORA 2020 and ISORA 2021 vintages due to a major revision of the survey structure; ISORA 2023 follows the same structures as the first two vintages.
  - Coverage has increased over time and includes 38 AEs, 78 EMs, and 50 LICs.
  - For countries missing a particular ISORA survey, missing values are filled using data from the closest available ISORA vintages.
- ISORA survey characteristics:
  - ISORA surveys tax administration features through both numerical and categorical questions.
  - The latest vintage contains 606 questions.
  - Data are collected via voluntary self-assessment (online questionnaire).
- Novel TAPRED database:
  - Global Tax Policy and Revenue Evaluation Database (TAPRED) built using a narrative approach and LLM by Atsebi et al. (forthcoming).
  - TAPRED collects tax policy changes and revenue yields from 5,200 IMF Staff Reports and MEFPs across 189 countries (1998–2024).
  - TAPRED captures attributes of tax policy measures: type, timing, base vs. rate, rationale, implementation status, estimated revenue impact (in percent of GDP), direction, one-off nature, and whether based on forecasts.

### Delphi method and expert weighting
- Iterative three-step Delphi Method involving a group of five IMF tax administration experts to:
  - Select and classify relevant ISORA indicators to improve practicality, granularity, and consistency across vintages.
  - Design and pilot an expert weighting survey to assess relative importance of each practice or structural foundation in raising tax revenue.
  - Support administration of the larger expert survey.
- Pilot of an initial Tax Administration Yield and Assessment Tool (TAYAT) was conducted in eight countries (Albania, Cabo Verde, Georgia, Malawi, Maldives, Mongolia, Sri Lanka, Uzbekistan) by 13 IMF tax administration experts; pilots indicated need for greater granularity and coverage.
- Expert weighting survey:
  - Piloted by the five experts to refine structure and questions.
  - Administered to a larger group; a 42 percent response rate was obtained.

### Operational Strength Index (OSI) structure
- OSI is based on nine sub-indices reflecting specific tax administration practices and structural foundations:
  1. Compliance Risk Management (CRM)
  2. Use of Third-Party Data (UTD)
  3. Degree of Digitalization (DIG)
  4. Service Orientation (SOR)
  5. Public Accountability (PAC)
  6. Autonomy (AUT)
  7. Large Taxpayers Office and High-Net-Worth Individuals (LTO and HNWI)
  8. Tax Enforcement (ENF)
  9. Human Resources Management and Development (HRM)
- The OSI measures overall strength of tax administration practices, institutional frameworks, and structural foundations.

### Empirical identification and estimation approach (data-related points)
- IV constructed from lagged cumulative FAD CD FTE in tax administration over two to five years earlier to instrument OSI.
- Fixed effects models exploit within-country variation and allow OSI to vary over time.
- TAPRED used to control for tax policy changes with detailed estimated revenue impacts.

*Source: Appendix I. Data Descriptions and Sources, wpiea2025219-source-pdf*

### Appendix Table AI.1. These sub-indices are compiled using responses to a series of mostly “yes/no” survey

### Appendix Table AI.1. These sub-indices are compiled using responses to a series of mostly “yes/no” survey questions.

### Construction of the OSI and Expert Weighting Survey
- The new approach incorporates 193 questions, compared with the 66 questions used by Chang et al. (2020) and Adan et al. (2023).
- The index structure:
  - 41 disaggregated indices (sub-categories).
  - 152 more detailed indices (sub-subcategories).
- Expert weighting survey:
  - 30 expert respondents.
  - Experts distributed 100 points among:
    - the nine indices (layer 1),
    - sub-categories within each of the nine indices (layer 2),
    - sub-sub-categories within each sub-category (layer 3).
  - Global weights obtained by multiplying weights from layers 1 to 3 to produce relative importance for each of the 193 questions.
  - The full disaggregated global weights for 193 questions were used to construct the OSI; only average weights at layer 1 are presented in Figure 1.
- Average weight results (layer 1, as reported):
  - Highest-rated practices: CRM, UTD, and DIG.
  - Followed by: ENF, LTO and HNWI, SOR, HRM, AUT, and PAC.
- Variation in expert opinion:
  - Variations are moderate overall.
  - Lower-rated practices such as PAC and ENF show the greatest divergence (highest coefficients of variation).
- Data adjustments for comparability:
  - Changes across ISORA vintages (added/removed questions, wording refinements, enhanced guidance, structural adjustments) introduce inconsistencies.
  - For questions absent in 2016 and 2017 vintages, gaps are filled using responses from the 2023 ISORA vintage.

### Other Variables (Dependent Variable and Controls)
- Dependent variable:
  - Tax-to-GDP ratio, excluding trade taxes and social security contributions.
  - Compiled from the IMF’s WoRLD and WEO database.
  - Trade taxes and social security contributions excluded to abstract from commodity price volatility and resource revenues, and because these revenues may be collected by agencies other than the tax administration.
- Controls included (based on prior literature):
  - Macroeconomic variables: per capita GDP and its square, inflation, trade openness, external debt to GDP, and terms of trade.
  - Structure and institutions: oil exports to GDP, the share of agriculture to GDP, and control of corruption.
  - A novel measure: revenue yields from tax policy changes from the TAPRED by Atsebi et al. (forthcoming).
- Data descriptions and sources are in Appendix Table AI.3; summary statistics in Appendix Table AII.1.

### Novel Stylized Facts (OSI Dynamics and Cross-Group Patterns)
- Coverage:
  - ISORA vintages cover the years 2014, 2015, 2016, 2017, and 2022.
- Cross-country and temporal patterns:
  - EMs and LICs: tax administration strength (OSI) has significantly improved over time, with a very low initial OSI score in 2014.
  - Many EMs and LICs have 2022 OSI located above their 2014 OSI (compared with the 45-degree line indicating no change), indicating convergence in tax administration performance.
  - AEs and EMs: limited improvement over time or even a slight decline in 2022.
  - LICs: steady strengthening of tax administration practices over the past decade; since 2016 the overall distribution of the OSI (median as well as 25th and 75th percentile values) has shifted upward, including the period after the COVID-19 pandemic.
- Drivers and component performance:
  - Digitalization (DIG) has been a main driver of improvement in OSI for EMs and LICs over the past decade.
  - LICs have relied more heavily on digital technologies in recent years, catching up to EMs especially after the COVID-19 pandemic (references: Amaglobeli et al., 2023; Nose and Mengistu, 2023; Okunogbe and Tourek, 2024).
  - Use of third-party data (UTD), closely related to digitalization, has slightly improved in LICs but remains particularly low in LICs and EMs.
  - Other practices—service orientation (SOR), public accountability, human resource management and development (HRM)—have improved slightly.
- Data-quality caveat:
  - ISORA 2023 vintage documents a noticeable decline in most performance scores (service orientation, human resource management, public accountability, and tax enforcement).
  - This likely reflects improvements in data quality: the 2023 round introduced stricter verification and validation of country responses, reducing potential overreporting in earlier surveys and complicating direct comparisons across vintages.
- Additional analysis:
  - Appendix II presents correlations between tax revenue and tax administration strength across different contexts (Appendix Tables AII.2 to AII.4).

### Empirical Methodology — Challenges and Bias
- Key empirical challenges that may bias revenue gain estimates from OSI:
  1. Limited time coverage of OSI:
     - Panel restricted to five years: 2014–2017, 2022, with a significant gap between 2017 and 2022.
  2. Limited within-country variation in OSI:
     - OSI is a slow-moving variable (Appendix Table AII.2), constraining identification strategies.
     - Note: the latest ISORA release (2023 vintage) introduces greater variation, which the authors exploit to estimate a fixed effects model.
  3. Simultaneity bias:
     - Tax revenue can influence tax administration reforms, and reforms can also impact tax revenue.
     - Example: Ebeke et al. (2016) show reforms (LTOs and SARAs) are more likely in countries with low revenue and IMF-supported programs, creating a negative feedback loop that could underestimate the OSI’s true effect.
  4. Omitted variable bias:
     - Unobserved factors (tax morale, geographical characteristics such as resource endowment, historical legacies, social trust) may jointly affect OSI and revenue.
     - Omitting time-invariant characteristics correlated with OSI and revenue can bias results.
  5. Measurement error:
     - ISORA is computed based on self-reported data; some countries may overrate their tax administration strength.
     - Overstatement of perceived strength leads to downward bias in estimating the true effect of tax administration on revenue.
- Identification and estimation implications:
  - (i) and (ii) limit the use of within-country changes in OSI as the key variable of interest, but the 2023 vintage’s greater within-country variation allows estimation of fixed effects models.
  - Simultaneity (iii) may cause underestimation of OSI’s positive impact on revenue.
  - Measurement error (v) likely leads to a downward bias in estimated effects.

*Source: IMF Working Paper — Appendix content as provided in the content unit.*

### 2. Empirical Strategy

### 2. Empirical Strategy

### Identification and estimation approach
- Use panel fixed effects exploiting within-country variations to relax the assumption that OSI is time-invariant and to control for time-invariant unobservable factors that affect tax revenue.
- Include country fixed effects (휇i) to control for time-invariant unobservable factors correlated with both OSI and tax revenue, and year fixed effects (휏t) to account for common shocks.
- Control for several macroeconomic, structural, and institutional factors identified in prior studies, and incorporate a novel LLM-powered narrative database capturing revenue yields from tax policy measures extracted from 5,200 IMF staff reports (Atsebi et al., forthcoming) to better distinguish tax policy impacts from tax administration effects.
- Conduct outlier analysis to identify countries that may underreport or overreport OSI; perform robustness checks excluding specific country groups (top/bottom 10 percent of OSI or tax-to-GDP ratio, FCS, small island countries, resource-rich countries) and top/least reformers (top and bottom 10 percent of OSI changes).

### Instrumental variable strategy and simultaneity
- Address simultaneity bias by instrumenting OSI with the intensity of the IMF’s FAD CD programs in tax administration.
- Instrument definition: total FTE used in tax administration CD over a single year, aggregated over lagged windows (e.g., past four to two years and past five to two years).
- Rationale: sustained IMF support for tax administration reforms is correlated with subsequent OSI improvements which affect tax revenue; CD programs focus on institutional and administrative reforms rather than directly setting tax policy.
- Use lagged CD intensity (cumulative FTE over the past four or five to two years) to provide temporal distance and to mitigate reverse causality concerns.
- Note: the model is just-identified (one endogenous variable, one instrument), so exclusion restriction cannot be empirically tested in this specification.

### Estimation equations
- First stage (Eq. 1):
  - OSI_{i,t} = α + δ1 FTE_RA_{i,t} + β1 X_{i,t} + τ_t + μ_i + ξ_{i,t}
  - FTE_RA_{i,t} denotes instrument (intensity of FTE used in tax administration CD).
- Second stage (Eq. 2):
  - Tax_{i,t} = α + δ2 OSÎ_{i,t} + β2 X_{i,t} + τ_t + μ_i + ε_{i,t}
  - Tax_{i,t} is tax-to-GDP ratio excluding trade taxes and social security contributions; OSÎ_{i,t} is predicted OSI from first stage; δ2 captures the revenue gain from OSI.
- X_{i,t} is a set of control variables (macroeconomic, structural, institutional variables cited in literature).

### Baseline IV and OLS findings (summary of Table 1)
- OLS: weak and insignificant relationship between OSI and tax-to-GDP for the full sample (Columns 1 and 2).
- IV estimates (Columns 3–6) are substantially larger and statistically significant at the 1 percent level.
- IV estimated OSI coefficients (panel fixed-effects IV):
  - Full sample: 0.0821, 0.1015, 0.0821, 0.1015 (reported ranges in text: 0.082 to 0.102).
  - EMDEs: 0.1177, 0.1428 (reported ranges in text: 0.118 to 0.143).
- Interpretation examples:
  - An increase in OSI from 33rd to 67th percentile (lower third to upper third) leads to:
    - Full sample: increase in tax-to-GDP by 1.2 ppts.
    - EMDEs: increase in tax-to-GDP by 1.7 ppts (enough to close 25–35 percent of the total tax gap identified by Benitez et al. (2023)).
- Significance: IV estimates are highly significant at the 1 percent level (***).

### Exact coefficient and statistic highlights from Table 1 (preserved values)
- Operational Strength Index [0,1]: Column (3) 0.1015*** (0.0059); Column (4) 0.1428*** (0.0100); Column (5) 0.0821*** (0.0053); Column (6) 0.1177*** (0.0183).
- Tax Policy Yield, exc. Trade Tax and SSC: ranges 0.3020*** (0.0408) to 0.3196*** (0.0122); Column (3) 0.3320*** (0.0084).
- #Tax Staff/LaborForce: Column (3) 0.5339*** (0.0812); Column (4) 0.5943*** (0.1028).
- Sq(#Tax Staff/Labor Force): Column (3) -2.2606*** (0.3449); Column (4) -2.5973*** (0.4755).
- Real GDP growth, lagged: coefficients ~0.0006–0.0008*** (very small positive effect).
- Trade openness (% of GDP), lagged: coefficients ~0.0169–0.0224*** with standard errors (0.0032)–(0.0055).
- Inflation, lagged: negative and significant coefficients ~ -0.0249*** to -0.0328***.
- Oil exports (% of GDP), lagged: negative and significant coefficients ~ -0.0948*** to -0.1415***.
- Observations: 529 (full sample) and 424 (EMDEs).
- Number of countries: 121 (full sample) and 99 (EMDEs).
- AIC and BIC: lower for lag window [-5 to -2] (Columns 5 and 6) than for [-4 to -2] (Columns 3 and 4); reported values include AIC -3490.8 (Col 5), -3472.9 (Col 3); BIC -3473.7 (Col 5), -3455.8 (Col 3).
- First-stage diagnostic statistics (from Table 1): Cragg-Donald Wald F statistic: 36.230 (Col 3), 24.179 (Col 4), 43.229 (Col 5), 29.060 (Col 6). Kleibergen-Paap rk Wald F statistic: 10.769 (Col 3), 11.742 (Col 4), 12.843 (Col 5), 14.545 (Col 6).
- Ho: OSI is exogenous — rejected (reported as 0 for IV columns).

### First-stage results (summary of Table 2)
- Intensity of FTE used in RA (time in years), instrument effect on OSI:
  - Column (1) 0.0172*** (0.0037)
  - Column (2) 0.0182*** (0.0053)
  - Column (3) 0.0150*** (0.0032)
  - Column (4) 0.0158*** (0.0044)
- Interpretation: a one-year increase in cumulative FAD CD intensity over the specified lag window increases OSI by up to 0.015–0.018.
- Within R-squared for first stage: 0.3343 (Col 1) to 0.3637 (Col 4), indicating reasonable explanatory power.
- Observations: 529 (full sample) and 424 (EMDEs); Number of countries: 121 and 99.

### Findings on control variables and mechanisms
- Tax policy yield: on average, a 1 ppt of GDP increase in expected revenue from tax policy changes raises actual tax revenue by around 0.32 ppt of GDP (less than one-for-one), consistent across specifications.
- Tax staff per labor force: positive association with tax revenue with diminishing returns (positive linear term and negative squared term).
- Macroeconomic factors:
  - Real GDP growth: small positive effect (~0.0006–0.0008).
  - Trade openness: significant positive effect.
  - Terms of trade: positive and significant, especially in EMDEs.
  - GDP per capita: positive in EMDEs with diminishing returns (negative squared term).
  - External debt: negative association with tax revenue, particularly in EMDEs.
  - Inflation: consistently negative effect on tax revenue.
- Structural/institutional factors:
  - Oil-exporting countries: significantly lower tax-to-GDP ratios (supports resource curse hypothesis).
  - Agricultural dependence: negatively correlates with tax revenue.
  - No robust association found between size of active taxpayers or control of corruption and tax revenue in this sample.
- Narrative on tax policy yield: potential reasons for the <1 effect include overestimated revenue yields in staff reports, incomplete or ineffective implementation of tax policy measures, and behavioral responses by taxpayers (avoidance/evasion).

### Interpretation and implications of baseline results
- OLS likely underestimates the true effect of OSI due to simultaneity bias; IV estimates provide strong evidence that strengthening tax administration significantly increases tax revenue.
- Effects are particularly pronounced in EMDEs where there is more room to improve tax administration.
- The preferred instrument lag window is the past five to two years ([-5 to -2]) based on lower AIC and BIC.
- The first-stage F-statistics and Kleibergen-Paap statistics exceed conventional thresholds, supporting instrument strength.
- Findings underscore the importance of sustained, long-term capacity development (CD) efforts—by IMF FAD and likely by broader development partners—in strengthening tax administration and enhancing domestic revenue mobilization.

### Robustness checks (overview)
- Alternative lag structures for the instrument including cumulative FTE over past five or four years (years -5/-4 to -1) and past five or four to three years (years -5/-4 to -3).
- Alternative measures of OSI: lagged OSI and equal-weighted OSI.
- Inclusion of income group-year fixed effects to account for differential macroeconomic shocks and policy responses across income groups.
- Alternative samples excluding specific country groups (top/bottom 10 percent OSI or tax-to-GDP, FCS, small islands, resource-rich countries, top and least reformers).
- Alternative control specifications: retain only one among strongly correlated controls; add controls for informality, urbanization, age dependency ratio, education and health spending, and financial development.

_Italic: Source — IMF Working Paper chapter "2. Empirical Strategy" (wpiea2025219-source-pdf)._

### 1. Alternative Lag Structures of the Instrument

### 1. Alternative Lag Structures of the Instrument

### Key points
- Baseline instrument: lagged values of the intensity of FAD CD in tax administration measured by the cumulative FTE used over the past five or four to two years (years -5/-4 to -2).
- Alternative lag structures tested:
  - cumulative FTE over the past five or four years (-5/-4 to -1)
  - cumulative FTE over the past five or four to three years (-5/-4 to -3)
- Empirical outcomes:
  - IV estimates (Appendix Tables AIII.1 and AIII.2) show a strong and statistically significant relationship between OSI and tax revenue across all lag structures.
  - Using CD intensity over the past five or four years (-5/-4 to -1) yields the largest estimates (Appendix Table AIII.1).
  - Using CD intensity over the past five or four to three years (-5/-4 to -3) yields slightly smaller estimates (Appendix Table AIII.2).
  - Instrument diagnostics:
    - Cragg-Donald and Kleibergen-Paap rk Wald F statistics indicate the (-5/-4 to -1) lag offers the strongest instrument relevance.
    - Model fit (AIC and BIC) is better for the (-5/-4 to -3) lag.
  - Baseline choice rationale: strikes a balance between instrument relevance and exogeneity, combining the highest precision (lowest standard errors) with a sound trade-off.

### 2. Alternative Measures of OSI

### Measures examined
- Baseline: contemporaneous OSI and expert-weighted OSI.
- Alternatives tested:
  - lagged OSI
  - equal-weighted OSI (assigns equal importance to all tax administration practices and structural foundations)

### Findings
- OLS estimates are generally downward biased.
- IV regressions: OSI is strongly and positively associated with tax revenue at the 1 percent significance level (Appendix Tables AIII.3 and AIII.4).
- Specific effects:
  - lagged OSI has a larger positive effect on tax revenue.
  - equal-weighted OSI yields higher coefficients than expert-weighted OSI.
- Model selection:
  - AIC and BIC and F-test indicate baseline specifications using contemporaneous OSI and expert-weighted OSI are preferred.

### 3. Controlling for Income Group-Year Fixed Effects

### Purpose
- Control for macroeconomic shocks, structural differences, and policy responses that vary across income groups over time (notably pandemic impacts).

### Findings
- Results in Appendix Table AIII.5:
  - When controlling for income group-year fixed effects, OSI coefficients in IV regressions decrease in magnitude but remain statistically significant.
  - Interpretation: some variation previously attributed to OSI may reflect broader income group-specific trends (e.g., fiscal impact of the COVID-19 pandemic, structural differences in economic resilience and policy responses).

### 4. Alternative Samples

### Sample tests (three counts)
- (i) Exclude countries in the top or bottom 10 percent of OSI or tax-to-GDP ratio to prevent extreme cases or outliers from distorting results.
- (ii) Exclude FCS, small island countries, and resource-rich countries because they often have distinct tax revenue levels, volatility, and mobilization strategies.
- (iii) Exclude countries among the top or bottom 10 percent in terms of OSI changes between 2017 and 2022 to avoid influence from extreme reformers, non-reformers, or back-sliding countries.

### Findings
- Results presented in:
  - Appendix Tables AIII.6 to AIII.9 for (i)
  - Appendix Tables AIII.10 to AIII.12 for (ii)
  - Appendix Tables AIII.13 and AIII.14 for (iii)
- Overall robustness:
  - Baseline results are consistent and robust across alternative samples; magnitude varies with sample exclusions.
  - Coefficients associated with OSI are statistically significant at the 1 percent level in all specifications and are generally of higher magnitude.
  - Excluding countries with highest OSI scores (top 10 percent of OSI), lower tax-to-GDP ratios (bottom 10 percent), FCS, resource-rich countries, and top or least reformers (top or bottom 10 percent in OSI change) makes the OSI–tax revenue relationship stronger (higher magnitude coefficients).
  - Excluding countries with the highest tax-to-GDP ratios (top 10 percent), the lowest OSI (bottom 10 percent of OSI), and small island states yields OSI coefficients slightly smaller or comparable to baseline IV estimates.
- Interpretation: extreme cases and countries with unique tax revenue features can dilute baseline estimated effects of tax administration strength.

### 5. Alternative Control Variables

### Strategy
- Two refinements:
  - retain only one variable from groups of strongly correlated controls
  - incorporate additional controls: informality, urbanization, age dependency ratio, education and health spending, and financial development

### Findings (Appendix Tables AIII.15 to AIII.24)
- Results closely align with baseline findings in both significance and magnitude for OSI in most specifications.
- Interpretation of control variables largely unchanged.
- Additional control variable outcomes:
  - informality (proxied by self-employment as a share of total employment) and education and health spending: generally no effect on tax revenue (coefficients small and insignificant).
  - urbanization and age dependency ratio: tax revenue decreases with higher urbanization and higher age dependency ratio.
  - financial development: unexpectedly negatively associated with tax revenue.

### VII. Sensitivity

### Methodology
- Assess how the impact of OSI on tax revenue varies by country characteristics: level of development (LICs vs. EMs), informality, financial development, institutions and governance quality.
- Compute the average of each conditioning variable over 2017–22 and interact it with OSI.
- Instrument both OSI and the interaction term using the intensity of FAD CD in tax administration and its interaction with the average variable to avoid forbidden regression issues.
- Averages are absorbed by country fixed effects.

### 1. LICs vs. EMs
- Context:
  - LICs generally lag EMs in practices and structural foundations; high informality and volatility can limit revenue gains from stronger tax administration.
  - EMs have more developed administrative structures and translate efficiency improvements into higher tax revenue more effectively.
- Results (Appendix Table AIII.25):
  - Revenue gains from strengthening tax administration are more than twice as large in EMs compared to LICs.
  - Intensity of FAD CD missions is less correlated with OSI in EMs; correlation is stronger in LICs (indicated by Kleibergen-Paap rk Wald F statistic).
  - Differentiated control variable effects:
    - In EMs: tax revenue negatively associated with inflation and oil exports; positively associated with terms of trade and tax policy yield.
    - In LICs: tax revenue negatively related to share of agriculture in GDP; positively associated with better control of corruption.

### 2. Informality
- Context:
  - Informality reduces the effectiveness of tax collection, particularly in LICs.
- Results (Appendix Table AIII.26):
  - OSI remains positive and significant at the 1 percent level.
  - Interaction between OSI and informality (self-employment as share of total employment) is negative and significant at the 5 percent level in columns 5 and 6, though smaller in magnitude.
  - Interpretation:
    - OSI impact on tax revenue remains positive regardless of informality level.
    - Revenue gains from improved tax administration decline as informality increases.
    - Quantitative thresholds:
      - When self-employment exceeds 48 and 75 percent of total employment (corresponding to the top 30 and 14 percent of countries with the highest self-employment), countries will experience lower revenue gains compared to baseline estimates for the full and EMDE samples, respectively.

### 3. Financial Development
- Context:
  - Financial development can provide third-party data and transaction transparency that aid tax enforcement.
- Results (Appendix Table AIII.27):
  - OSI coefficients are statistically significant only in the EMDE sample.
  - Interaction between OSI and financial development is significant and large in magnitude across all specifications.
  - Interpretation:
    - Financial development mediates the impact of OSI on revenue gains.
    - Quantitative thresholds:
      - 60 percent of countries in the full sample (financial development index below 0.32) experience lower revenue gains than baseline.
      - 40 percent of countries in the EMDE sample (financial development index below 0.17) experience lower revenue gains than baseline.

### 4. Institutions and Governance Quality
- Context:
  - Stronger institutions and governance quality amplify revenue gains from improved tax administration.
- Measures used:
  - control of corruption, rule of law, government effectiveness, regulatory quality, political stability and absence of violence or terrorism, voice and accountability.
- Results (Appendix Tables AIII.28 to AIII.33):
  - OSI coefficients remain significant at the 1 percent level across specifications.
  - Interaction terms between OSI and governance indicators are positive and significant at least at the 5 percent level.
  - Interpretation:
    - Stronger institutions and governance enhance the translation of OSI improvements into higher tax revenue.
    - Depending on the governance indicator, between 0 and 17 percent of countries experience negative revenue impacts despite OSI improvements due to weaker institutions and governance.
    - Between 21 and 58 percent of countries experience less than baseline revenue gains due to weaker institutions and governance.

*Source: IMF Working Paper section "1. Alternative Lag Structures of the Instrument" (source PDF).*

### 5. Differentiated Effects on Tax Types

### 5. Differentiated Effects on Tax Types

### Key findings on tax-type impacts
- Taxes on sales and production, including VAT:
  - The largest revenue gains are observed for taxes on sales and production, including VAT.
  - Interpretation: improvements in tax administration are particularly effective in enhancing indirect tax collection, consistent with a broader tax base and higher enforceability of indirect taxes relative to direct taxes.
- Property taxes:
  - Property taxes also benefit from stronger tax administration, though the revenue gains are generally smaller than for taxes on sales and production.
  - Note: lower coefficients for property taxes may partly reflect that, in many countries, property taxes are not administered by the national tax authority but are instead collected directly by local governments.
- Income and profit taxes:
  - Income and profit taxes increase, but to a smaller extent than indirect taxes.
  - Interpretation: constraints beyond tax administration—such as informality and tax avoidance—continue to limit revenue collection for direct taxes.
- Taxes not elsewhere classified and non-tax revenue:
  - Taxes not elsewhere classified and non-tax revenue are not affected by tax administration strength; coefficients are statistically insignificant in all specifications at a 10 percent significance level.
  - Rationale for non-tax revenue: inherently more volatile and influenced by commodity prices, contractual agreements, donor commitments; collection mechanisms often differ and may fall under agencies other than tax authorities.

### Reasons for heterogeneous effects across tax types
- Structural and institutional constraints:
  - Widespread poverty, inequality, informality, governance issues (including vested interest groups), and lack of property registers limit direct tax revenue gains.
- Policy trade-offs:
  - Competing objectives in many developing countries (e.g., increasing revenue while promoting investment) can constrain moves toward greater reliance on direct taxes.
- Design of indirect taxes:
  - Exemptions and reduced rates can limit indirect-tax potential, yet VAT’s broader base and enforceability make it more responsive to administration improvements.
- Non-tax revenue volatility and institutional differences:
  - Non-tax revenue depends on commodity prices, contractual arrangements, donor commitments, and different collection agencies, reducing sensitivity to tax-administration reforms.

### Empirical approach and measurement
- Instrumental variable (IV) strategy:
  - Used the intensity of past IMF’s FAD CD in tax administration as an instrument for the strength of tax administration.
- Novel indices of tax administration strength:
  - Developed measures based on a Delphi method and a weighting survey to refine indices’ practicality, granularity, consistency across ISORA vintages, and incorporate relative importance of practices and structural foundations.
- Robustness:
  - Results presented in Appendix Tables AIII.34 to AIII.38 (not reproduced here) show variation across tax types and statistical significance patterns noted above.

### Policy implications relevant to differentiated effects
- Adopt a holistic and institutional strategy:
  - Integrate tax administration, tax policy, and financial, institutional, or legal reforms to support revenue mobilization.
  - Targeted reforms should be based on a country’s level of development and structural constraints, taking into account the starting point and local context.
- Invest in capacity development (CD):
  - Sustained investment in CD programs is essential for building effective tax administration.
  - Evidence of a positive relationship between the IMF’s past CD and improvements in recipient country tax administration strength suggests the importance of long-term political commitment and continued CD interventions.
- Monitor and evaluate continuously:
  - Regular evaluations of tax administration performance (including ISORA and TADAT) can allow governments to adjust reform strategies and strengthen implementation.

### Cross-country heterogeneity and additional conclusions
- Magnitude differences by country group:
  - Revenue gains from tax administration reforms are more than twice as large in EMs compared to LICs, implying structural impediments and institutional weaknesses in LICs may limit benefits.
  - Estimated revenue gains are relatively smaller in AEs, where tax systems are mature, institutions are well functioning, and tax compliance is high.
- Factors associated with smaller revenue gains:
  - Higher levels of informality, weaker institutions, lower financial development, and a higher share of indirect taxes are associated with smaller estimated revenue gains.
- Overall assessment:
  - Strengthening tax administration significantly increases tax revenue; gains are lower than in some prior studies due to improved measurement and empirical strategies, but remain sizable and economically meaningful.

### Selected descriptive statistics (Appendix Table AII.1)
- Non-Trade Tax Revenue (Percent of GDP), consolidated WoRLD and WEO:
  - Obs.: 901; Mean: 0.152; Median: 0.142; Std.Dev.: 0.079; Min: 0.002; Max: 0.496
- Operational Strength Index (OSI), Expert Weights:
  - Obs.: 741; Mean: 0.609; Median: 0.627; Std.Dev.: 0.150; Min: 0.075; Max: 0.879
- Selected sub-indices (Obs.: 747 unless noted):
  - Compliance Risk Management (CRM): Mean: 0.627; Median: 0.649; Std.Dev.: 0.191; Min: 0.000; Max: 0.967
  - Use of Third-Party Data (UTD): Mean: 0.409; Median: 0.385; Std.Dev.: 0.266; Min: 0.000; Max: 1.000
  - Digitalization (DIG): Mean: 0.659; Median: 0.781; Std.Dev.: 0.294; Min: 0.000; Max: 1.000
  - Service Orientation (SOR): Mean: 0.693; Median: 0.733; Std.Dev.: 0.217; Min: 0.000; Max: 1.000
  - Public Accountability (PAC): Mean: 0.564; Median: 0.561; Std.Dev.: 0.179; Min: 0.000; Max: 1.000
  - Autonomy (AUT): Mean: 0.718; Median: 0.772; Std.Dev.: 0.247; Min: 0.000; Max: 1.000
  - LTO and HNWI (LTO): Mean: 0.584; Median: 0.622; Std.Dev.: 0.243; Min: 0.000; Max: 1.000
  - Tax Enforcement (ENF): Mean: 0.575; Median: 0.571; Std.Dev.: 0.160; Min: 0.000; Max: 0.914
  - Human Resource Management (HRM): Mean: 0.665; Median: 0.696; Std.Dev.: 0.296; Min: 0.000; Max: 1.000
- IMF CD intensity and FTE measures (Obs.: 960):
  - FTE used in IMF tax administration CD (in years): Mean: 0.199; Median: 0.004; Std.Dev.: 0.403; Min: 0.000; Max: 4.476
  - Cumulative FTE used in IMF tax admin CD over 5 to 2 years earlier (in years): Mean: 0.644; Median: 0.311; Std.Dev.: 1.065; Min: 0.000; Max: 10.660

*Source: Authors’ calculations using ISORA and WEO*

### Appendix Table AII.2. Correlations of the Operational Strength Index over Time

### Appendix Table AII.2. Correlations of the Operational Strength Index over Time

### Correlation matrix (Operational Strength Index across years)
- OSI2014 with:
  - OSI2014: 1
  - OSI2015: 0.997
  - OSI2016: 0.961
  - OSI2017: 0.956
  - OSI2022: 0.893
- OSI2015 with:
  - OSI2014: 0.997
  - OSI2015: 1
  - OSI2016: 0.961
  - OSI2017: 0.956
  - OSI2022: 0.893
- OSI2016 with:
  - OSI2014: 0.961
  - OSI2015: 0.961
  - OSI2016: 1
  - OSI2017: 0.997
  - OSI2022: 0.916
- OSI2017 with:
  - OSI2014: 0.956
  - OSI2015: 0.956
  - OSI2016: 0.997
  - OSI2017: 1
  - OSI2022: 0.915
- OSI2022 with:
  - OSI2014: 0.893
  - OSI2015: 0.893
  - OSI2016: 0.916
  - OSI2017: 0.915
  - OSI2022: 1

### Key observations from the table
- High persistence across adjacent years:
  - OSI2014–OSI2015 correlation is 0.997.
  - OSI2016–OSI2017 correlation is 0.997.
- Correlations decline modestly when comparing earlier years (2014–2017) with 2022:
  - Examples: OSI2014–OSI2022 is 0.893; OSI2017–OSI2022 is 0.915.
- Intermediate-year correlations are strong:
  - OSI2016–OSI2022: 0.916; OSI2017–OSI2022: 0.915.
- The matrix indicates overall strong temporal stability of the Operational Strength Index, with somewhat lower correlation between the 2014–2017 period and 2022 than within the 2014–2017 block.

*Source: Authors’ calculations*

### Appendix Table AIII.5. Robustness: Alternative Fixed Effects: Income Group X Year Fixed

### Appendix Table AIII.5. Robustness: Alternative Fixed Effects: Income Group X Year Fixed

### Overview
- This appendix table reports OLS (columns 1 and 2) and IV (columns 3 to 6) panel fixed-effects regressions of expert weighted Operational Strength Index (OSI) on Tax-to-GDP ratio excluding trade taxes and social contributions.
- Columns 1, 3 and 5 use the full sample; columns 2, 4 and 6 restrict to EMDEs. In columns 3 and 4 (columns 5 and 6) the instrument is the intensity of FTE used in tax administration IMF Capacity Development over the past four (five) to two years.
- Country fixed effects and Income Group X Year fixed effects are included in all regressions.

### Key regression findings (coefficients and significance)
- Operational Strength Index [0,1]:
  - Column (1) OLS All: 0.0064 (standard error 0.0156)
  - Column (2) OLS EMDEs: 0.0118 (0.0173)
  - Column (3) IV All [-4 to -2]: 0.0859 ***
    - (0.0059)
  - Column (4) IV EMDEs [-4 to -2]: 0.1236 ***
    - (0.0127)
  - Column (5) IV All [-5 to -2]: 0.0692 ***
    - (0.0050)
  - Column (6) IV EMDEs [-5 to -2]: 0.1026 ***
    - (0.0209)

- Tax Policy Yield, exc. Trade Tax and SSC:
  - Column (1): 0.3105 ***
    - (0.1169)
  - Column (2): 0.3044 **
    - (0.1255)
  - Column (3): 0.3454 ***
    - (0.0135)
  - Column (4): 0.3453 ***
    - (0.0218)
  - Column (5): 0.3380 ***
    - (0.0184)
  - Column (6): 0.3376 ***
    - (0.0152)

- #Tax Staff/LaborForce and squared term (nonlinear effect):
  - #Tax Staff/LaborForce:
    - Column (1): 0.4681 **
      - (0.2114)
    - Column (2): 0.5155 **
      - (0.2035)
    - Column (3): 0.5471 ***
      - (0.0740)
    - Column (4): 0.6119 ***
      - (0.1090)
    - Column (5): 0.5305 ***
      - (0.0783)
    - Column (6): 0.5938 ***
      - (0.1181)
  - Sq(#Tax Staff/Labor Force):
    - Column (1): -2.0507 **
      - (0.8336)
    - Column (2): -2.4301 ***
      - (0.9090)
    - Column (3): -2.2629 ***
      - (0.3249)
    - Column (4): -2.6071 ***
      - (0.5296)
    - Column (5): -2.2182 ***
      - (0.3392)
    - Column (6): -2.5739 ***
      - (0.5488)

- Other covariates with consistent signs and reported coefficients (selected):
  - Real GDP growth, lagged:
    - 0.0006 *** (0.0001); 0.0007 *** (0.0002); 0.0006 *** (0.0000); 0.0008 *** (0.0001); 0.0006 *** (0.0000); 0.0008 *** (0.0001)
  - Log (GDP per capita, USD), lagged:
    - 0.0185 (0.0203); 0.0270 (0.0207); 0.0186 (0.0155); 0.0320 ** (0.0137); 0.0186 (0.0157); 0.0310 ** (0.0139)
  - Sq(Log (GDP per capita, USD)), lagged:
    - -0.0018 (0.0014); -0.0025 * (0.0015); -0.0018 * (0.0010); -0.0029 *** (0.0009); -0.0018 * (0.0010); -0.0028 *** (0.0009)
  - Trade openness (% of GDP), lagged:
    - 0.0178 ** (0.0076); 0.0208 ** (0.0088); 0.0161 *** (0.0036); 0.0172 *** (0.0055); 0.0164 *** (0.0035); 0.0179 *** (0.0052)
  - External debt (% of GDP), lagged:
    - -0.0002 (0.0014); -0.0159 ** (0.0074); -0.0007 (0.0006); -0.0146 *** (0.0019); -0.0006 (0.0006); -0.0148 *** (0.0019)
  - Inflation, lagged:
    - -0.0230 (0.0159); -0.0240 (0.0151); -0.0246 *** (0.0032); -0.0299 *** (0.0028); -0.0243 *** (0.0032); -0.0288 *** (0.0032)
  - Terms of Trade (2000=1), lagged:
    - 0.0177 * (0.0092); 0.0199 ** (0.0089); 0.0139 (0.0098); 0.0157 * (0.0093); 0.0147 (0.0099); 0.0165 * (0.0096)
  - Oil exports (% of GDP), lagged:
    - -0.1072 * (0.0579); -0.1375 *** (0.0462); -0.0915 *** (0.0133); -0.1140 *** (0.0220); -0.0948 *** (0.0115); -0.1184 *** (0.0187)
  - Log (Agri, % of GDP), lagged:
    - -0.0050 (0.0072); -0.0075 (0.0079); -0.0098 ** (0.0045); -0.0126 *** (0.0039); -0.0088 * (0.0047); -0.0117 *** (0.0042)
  - Control Corruption, lagged:
    - 0.0028 (0.0048); 0.0042 (0.0054); 0.0010 (0.0020); 0.0022 (0.0049); 0.0014 (0.0020); 0.0026 (0.0048)

### Robustness and diagnostics (selected statistics)
- Observations: 523, 419, 529, 424, 529, 424 (by column)
- Number of countries: 121, 99, 121, 99, 121, 99 (by column)
- within R-squared: 0.9833 (col 1), 0.9757 (col 2); not reported for IV columns
- Country FE: Yes in all columns
- Income Group X Year FE: Yes in all columns
- AIC (IV columns): -3498.3 (col 3), -2806.3 (col 4), -3511.0 (col 5), -2825.8 (col 6)
- BIC (IV columns): -3481.2 (col 3), -2790.1 (col 4), -3493.9 (col 5), -2809.6 (col 6)
- Ho: OSI is exogenous — reported value 0 in IV columns (cols 3–6)
- Instrument strength and identification:
  - Cragg-Donald Wald F statistic: 36.497 (col 3), 25.743 (col 4), 42.939 (col 5), 30.275 (col 6)
  - Kleibergen-Paap rk Wald F statistic: 10.829 (col 3), 12.848 (col 4), 12.702 (col 5), 15.176 (col 6)

### Notes on estimation and significance
- Significance notation in tables:
  - * p < 0.10
  - ** p < 0.05
  - *** p < 0.010
- Robust standard errors are reported in parentheses.
- The IV specifications (columns 3–6) use as instrument the intensity of FTE used in tax administration IMF Capacity Development over the past four (five) to two years.

*Appendix Table AIII.5 appears in the IMF Working Paper "Enhancing Tax Capacity: Revenue Gains from Strengthening Tax Administration."*

### Appendix Table AIII.11. Robustness: Dropping Small Islands

### Appendix Table AIII.11. Robustness: Dropping Small Islands

### Regression setup and identification
- Estimation method: OLS (columns 1–2) and IV (columns 3–6) panel fixed-effects regressions of expert weighted OSI on Tax-to-GDP ratio excluding trade taxes and social contributions.
- Sample splits:
  - Columns 1, 3, and 5: full sample.
  - Columns 2, 4, and 6: EMDEs only.
- Instruments (columns 3–4 and 5–6): intensity of FTE used in tax administration IMF Capacity Development over the past four (five) to two years.
- Significance markers: * p < 0.10, ** p < 0.05, *** p < 0.010. Robust standard errors in parentheses.

### Key estimated coefficients (columns 1–6)
- Operational Strength Index [0,1]:
  - Column (1) OLS All: 0.0064 (0.0037)
  - Column (2) OLS EMDEs: 0.0094 (0.0065)
  - Column (3) IV All [-4 to -2]: 0.1136*** (0.0048)
  - Column (4) IV EMDEs [-4 to -2]: 0.1445*** (0.0090)
  - Column (5) IV All [-5 to -2]: 0.0950*** (0.0043)
  - Column (6) IV EMDEs [-5 to -2]: 0.1219*** (0.0116)

- Tax Policy Yield, exc. Trade Tax and SSC:
  - (1) 0.1615*** (0.0262)
  - (2) 0.2054*** (0.0402)
  - (3) 0.2751*** (0.0458)
  - (4) 0.3381*** (0.0542)
  - (5) 0.2555*** (0.0322)
  - (6) 0.3159*** (0.0432)

- #Tax Staff/LaborForce:
  - (1) 0.3417*** (0.0631)
  - (2) 0.3786*** (0.0796)
  - (3) 0.4491*** (0.0492)
  - (4) 0.4692*** (0.0687)
  - (5) 0.4305*** (0.0475)
  - (6) 0.4541*** (0.0668)

- Sq(#Tax Staff/Labor Force):
  - (1) -1.6554*** (0.1537)
  - (2) -1.9885*** (0.2425)
  - (3) -1.9210*** (0.1324)
  - (4) -2.0946*** (0.2185)
  - (5) -1.8750*** (0.1256)
  - (6) -2.0769*** (0.2154)

- Active Taxpayer/Labor Force:
  - (1) 0.0000 (0.0000)
  - (2) -0.0000 (0.0001)
  - (3) -0.0000 (0.0000)
  - (4) -0.0001 (0.0001)
  - (5) -0.0000 (0.0000)
  - (6) -0.0000 (0.0001)

- Real GDP growth, lagged:
  - (1) 0.0005*** (0.0001)
  - (2) 0.0006*** (0.0001)
  - (3) 0.0006*** (0.0001)
  - (4) 0.0007*** (0.0001)
  - (5) 0.0006*** (0.0001)
  - (6) 0.0007*** (0.0001)

- Log (GDP per capita, USD), lagged:
  - (1) 0.0213 (0.0114)
  - (2) 0.0136 (0.0080)
  - (3) 0.0173 (0.0132)
  - (4) 0.0183*** (0.0067)
  - (5) 0.0180 (0.0134)
  - (6) 0.0175** (0.0070)

- Sq(Log (GDP per capita, USD)), lagged:
  - (1) -0.0017* (0.0007)
  - (2) -0.0013* (0.0005)
  - (3) -0.0013 (0.0009)
  - (4) -0.0017*** (0.0005)
  - (5) -0.0014 (0.0009)
  - (6) -0.0016*** (0.0005)

- Trade openness (% of GDP), lagged:
  - (1) 0.0171*** (0.0019)
  - (2) 0.0232*** (0.0029)
  - (3) 0.0158*** (0.0035)
  - (4) 0.0204*** (0.0053)
  - (5) 0.0160*** (0.0033)
  - (6) 0.0209*** (0.0050)

- External debt (% of GDP), lagged:
  - (1) 0.0002 (0.0010)
  - (2) -0.0126*** (0.0018)
  - (3) 0.0002 (0.0007)
  - (4) -0.0129*** (0.0036)
  - (5) 0.0002 (0.0007)
  - (6) -0.0129*** (0.0033)

- Inflation, lagged:
  - (1) -0.0011 (0.0016)
  - (2) -0.0083* (0.0033)
  - (3) -0.0001 (0.0018)
  - (4) -0.0118** (0.0055)
  - (5) -0.0003 (0.0016)
  - (6) -0.0112** (0.0054)

- Terms of Trade (2000=1), lagged:
  - (1) 0.0180** (0.0046)
  - (2) 0.0198*** (0.0041)
  - (3) 0.0122* (0.0065)
  - (4) 0.0150** (0.0061)
  - (5) 0.0132** (0.0062)
  - (6) 0.0158*** (0.0060)

- Oil exports (% of GDP), lagged:
  - (1) -0.0296 (0.0219)
  - (2) -0.0722*** (0.0156)
  - (3) 0.0082 (0.0209)
  - (4) -0.0340*** (0.0107)
  - (5) 0.0017 (0.0204)
  - (6) -0.0404*** (0.0107)

- Log (Agri, % of GDP), lagged:
  - (1) -0.0081 (0.0075)
  - (2) -0.0121 (0.0074)
  - (3) -0.0135 (0.0087)
  - (4) -0.0184* (0.0095)
  - (5) -0.0125 (0.0085)
  - (6) -0.0174* (0.0094)

- Control Corruption, lagged:
  - (1) 0.0071** (0.0017)
  - (2) 0.0104** (0.0035)
  - (3) 0.0050** (0.0020)
  - (4) 0.0076** (0.0038)
  - (5) 0.0054*** (0.0020)
  - (6) 0.0080** (0.0038)

### Sample size, fit, and diagnostics
- Observations:
  - Columns (1), (3), (5): 440
  - Columns (2), (4), (6): 349
- Number of countries:
  - Columns (1), (3), (5): 100
  - Columns (2), (4), (6): 81
- within R-squared:
  - Column (1): 0.237
  - Column (2): 0.290
- Fixed effects:
  - Country FE: Yes (all columns)
  - Year FE: Yes (all columns)
- Information criteria:
  - AIC (columns 3–6): -2945.4, -2344.9, -2963.4, -2369.5
  - BIC (columns 3–6): -2929.1, -2329.5, -2947.1, -2354.1
- Exogeneity and instrument strength:
  - Ho: OSI is exogenous – columns 3–6: 0 0 0 0
  - Cragg-Donald Wald F statistic (cols 3–6): 33.057, 20.527, 39.671, 25.468
  - Kleibergen-Paap rk Wald F statistic (cols 3–6): 21.708, 19.136, 21.105, 20.964

*IMF WORKING PAPERS Enhancing Tax Capacity: Revenue Gains from Strengthening Tax Administration — Appendix Table AIII.11. Robustness: Dropping Small Islands*

### Appendix Table AIII.17. Robustness: Dropping Strongly Correlated Variables: GDP per Capita and its Square

### Appendix Table AIII.17–AIII.23. Robustness Checks: Dropping/Adding Correlated Controls and Additional Controls

### Overview
- Series of panel fixed-effects regressions (OLS and IV) of expert weighted Operational Strength Index (OSI) on Tax-to-GDP ratio excluding trade taxes and social contributions.
- Columns: (1) OLS All, (2) OLS EMDEs, (3) IV All [-4 to -2], (4) IV EMDEs [-4 to -2], (5) IV All [-5 to -2], (6) IV EMDEs [-5 to -2].
- IVs: intensity of FTE used in tax administration IMF Capacity Development over the past four (five) to two years (used in columns 3–6).
- Robust standard errors in parentheses. Significance: * p < 0.10, ** p < 0.05, *** p < 0.010.

### Key consistent findings across robustness tables (AIII.17–AIII.23)
- Operational Strength Index (OSI) generally positively and significantly associated with Tax-to-GDP ratio excluding trade taxes and social contributions in IV specifications:
  - Appendix Table AIII.17 (Term, Trade Openness, and Corruption):
    - IV All [-4 to -2]: 0.1028 ***
    - IV EMDEs [-4 to -2]: 0.1365 ***
    - IV All [-5 to -2]: 0.0812 ***
    - IV EMDEs [-5 to -2]: 0.1006 ***
    - OLS All: 0.0082 ***
    - OLS EMDEs: 0.0094
  - Appendix Table AIII.18 (Term, Trade Openness, and Agriculture):
    - IV All [-4 to -2]: 0.0994 ***
    - IV EMDEs [-4 to -2]: 0.1373 ***
    - IV All [-5 to -2]: 0.0789 ***
    - IV EMDEs [-5 to -2]: 0.1014 ***
    - OLS All: 0.0056
    - OLS EMDEs: 0.0067
  - Appendix Table AIII.19 (Dropping Terms of Trade):
    - IV All [-4 to -2]: 0.1057 ***
    - IV EMDEs [-4 to -2]: 0.1431 ***
    - IV All [-5 to -2]: 0.0855 ***
    - IV EMDEs [-5 to -2]: 0.1170 ***
    - OLS All: 0.0144 **
    - OLS EMDEs: 0.0174 ***
  - Appendix Table AIII.20 (Dropping Oil Exports):
    - IV All [-4 to -2]: 0.0862 ***
    - IV EMDEs [-4 to -2]: 0.1199 ***
    - IV All [-5 to -2]: 0.0661 ***
    - IV EMDEs [-5 to -2]: 0.0913 ***
    - OLS All: 0.0139 ***
    - OLS EMDEs: 0.0203 **
  - Appendix Table AIII.21 (Adding Informality):
    - IV All [-4 to -2]: 0.1362 ***
    - IV EMDEs [-4 to -2]: 0.1796 ***
    - IV All [-5 to -2]: 0.1225 ***
    - IV EMDEs [-5 to -2]: 0.1550 ***
    - OLS All: 0.0193 ***
    - OLS EMDEs: 0.0154 ***
  - Appendix Table AIII.22 (Adding Urbanization and Age Dependency):
    - IV All [-4 to -2]: 0.1271 ***
    - IV EMDEs [-4 to -2]: 0.1724 ***
    - IV All [-5 to -2]: 0.1013 ***
    - IV EMDEs [-5 to -2]: 0.1295 ***
    - OLS All: 0.0073 **
    - OLS EMDEs: 0.0046
  - Appendix Table AIII.23 (Adding Education and Health Spending):
    - IV All [-4 to -2]: 0.1493 ***
    - IV EMDEs [-4 to -2]: 0.2178 ***
    - IV All [-5 to -2]: 0.1186 ***
    - IV EMDEs [-5 to -2]: 0.1614 ***
    - OLS All: 0.0118 *
    - OLS EMDEs: 0.0104 **

- Tax Policy Yield, exc. Trade Tax and SSC: consistently positive and statistically significant across all tables and specifications (examples):
  - Values reported across tables typically in the 0.2734 to 0.3579 range with many *** signs (e.g., Table AIII.19: 0.2734 *** (OLS All); Table AIII.21: 0.3263 *** (OLS All)).

- #Tax Staff/LaborForce: positive and significant in most specifications:
  - Examples: 0.4471 ** (Table AIII.17 OLS All), 0.4568 ** (Table AIII.18 OLS All), 0.4893 *** (Table AIII.23 OLS All).
- Sq(#Tax Staff/Labor Force): negative and statistically significant in most specifications, indicating diminishing marginal returns:
  - Examples: -1.9940 *** (Table AIII.17 OLS All), -2.0569 *** (Table AIII.18 OLS All), -1.9179 *** (Table AIII.23 OLS All).

- Active Taxpayer/Labor Force: coefficient reported as 0.0000 or -0.0000 in tables; significance varies (some * or ** in select tables), standard errors shown as (0.0000) or (0.0001).

- Real GDP growth, lagged: positive, small, and significant in most specifications (e.g., 0.0006 *** to 0.0008 *** across tables).

- External debt (% of GDP), lagged:
  - Often small positive in full-sample OLS (e.g., 0.0016 in Table AIII.17 OLS All) but negative and significant for EMDEs in many specifications (e.g., -0.0100 * in Table AIII.17 OLS EMDEs; -0.0160 *** in Table AIII.19 OLS EMDEs).

- Inflation, lagged: sign and significance vary by specification:
  - Positive and significant in some tables (e.g., Table AIII.17 OLS All: 0.0054), negative and significant in others (e.g., Table AIII.19 OLS All: -0.0282 ***).

- Terms of Trade (2000=1), lagged: generally small positive and sometimes significant (e.g., 0.0169 * in Table AIII.17 OLS All; 0.0190 ** in Table AIII.22 OLS All). Note Table AIII.19 is the robustness that drops Terms of Trade (but reports Log GDP and squared).

- Oil exports (% of GDP), lagged: typically negative and often statistically significant (examples):
  - Table AIII.17: -0.0945 ***
  - Table AIII.18: -0.0915 ***
  - Table AIII.22: -0.1131 ***

- Log (GDP per capita, USD), lagged and Sq(Log GDP per capita): included in some robustness checks (Tables AIII.19 and AIII.20), showing:
  - Log (GDP per capita, USD), lagged: positive coefficients (e.g., 0.0206, 0.0247), with some significance in EMDE IV specifications (e.g., 0.0323 **).
  - Sq(Log (GDP per capita, USD)), lagged: negative coefficients around -0.0020 to -0.0035, with */***/ significance in several specifications.

- Log (Agri, % of GDP), lagged: negative and often significant where included (e.g., -0.0120 ** in Table AIII.19 OLS All; -0.0046 in Table AIII.17 OLS All).

### Model fit and diagnostics (examples across tables)
- Observations and number of countries vary by table:
  - Example ranges: Observations often around 529–535 for full samples and 424–430 for EMDEs; Number of countries often around 121–122 for full samples and 99–100 for EMDEs. Some tables with added controls reduce sample (e.g., Table AIII.21 Observations 453 / 348; Number of countries 105 / 83).
- within R-squared: examples reported such as 0.168, 0.200, 0.185, 0.230, 0.176, 0.202, 0.203, 0.230, 0.196, 0.250, 0.244, 0.266 (varies by table and column).
- Country FE: Yes; Year FE: Yes in all specifications.
- AIC and BIC reported for IV specifications (examples):
  - Table AIII.17 IV All [-4 to -2] AIC: -3458.7; BIC: -3441.6.
  - Table AIII.22 IV All [-4 to -2] AIC: -3435.3; BIC: -3418.2.
- Ho: OSI is exogenous – reported as 0 in IV specifications.
- Weak instrument / relevance statistics (examples for IV columns):
  - Cragg-Donald Wald F statistic: e.g., 35.792, 22.313, 42.618, 27.373 (Table AIII.17); 36.519, 23.973, 43.300, 28.679 (Table AIII.19).
  - Kleibergen-Paap rk Wald F statistic: e.g., 12.487, 12.946, 14.270, 16.530 (Table AIII.17); 10.036, 10.934, 11.587, 12.757 (Table AIII.19).

### Implications from robustness checks (as evidenced by coefficients and tests)
- The positive and significant IV estimates on OSI across multiple robustness specifications indicate a robust positive association between OSI and tax-to-GDP ratio excluding trade taxes and social contributions across samples and after dropping/adding correlated controls.
- The positive coefficient on Tax Policy Yield and the positive (but diminishing) relationship with #Tax Staff/LaborForce (and negative Sq term) are robust, implying both tax policy and administrative capacity/staffing matter for revenue outcomes.
- Oil exports and agricultural share (Agri % of GDP) repeatedly show negative associations with the tax-to-GDP outcome in many specifications.
- Diagnostics indicate the instruments used (IMF Capacity Development FTE intensity) provide relevance (Cragg-Donald and Kleibergen-Paap statistics reported) and the null of OSI exogeneity is reported as 0 in IV specs.

*Source: Appendix Table AIII.17–AIII.23, wpiea2025219-source-pdf*

### Appendix Table AIII.24. Robustness: Adding Financial Development

### Appendix Table AIII.24. Robustness: Adding Financial Development

### Regression framework and samples
- Estimation methods and samples:
  - Columns (1) and (2): OLS (All; EMDEs).
  - Columns (3) to (6): IV (All; EMDEs), with instruments based on the intensity of FTE used in tax administration IMF Capacity Development over the past four (columns 3 and 4) or five to two years (columns 5 and 6).
  - Columns 1, 3, and 5: full sample.
  - Columns 2, 4, and 6: EMDEs only.
- Dependent variable: expert-weighted Operational Strength Index (OSI) regressed on Tax-to-GDP ratio excluding trade taxes and social contributions, plus controls and interactions including Financial Development Index.
- Fixed effects: Country FE = Yes; Year FE = Yes for all columns.

### Key coefficient estimates (point estimates and robust standard errors in parentheses)
- Operational Strength Index [0,1]:
  - (1) OLS All: 0.0025 (0.0041)
  - (2) OLS EMDEs: 0.0032 (0.0029)
  - (3) IV All [-4 to -2]: 0.1023*** (0.0061)
  - (4) IV EMDEs [-4 to -2]: 0.1521*** (0.0107)
  - (5) IV All [-5 to -2]: 0.0824*** (0.0081)
  - (6) IV EMDEs [-5 to -2]: 0.1165*** (0.0202)

- Tax Policy Yield, exc. Trade Tax and SSC:
  - (1) 0.3369*** (0.0373)
  - (2) 0.2932*** (0.0374)
  - (3) 0.3527*** (0.0130)
  - (4) 0.3158*** (0.0275)
  - (5) 0.3496*** (0.0127)
  - (6) 0.3104*** (0.0157)

- #Tax Staff/LaborForce:
  - (1) 0.4618** (0.1081)
  - (2) 0.5961** (0.1731)
  - (3) 0.5718*** (0.0950)
  - (4) 0.7366*** (0.1583)
  - (5) 0.5499*** (0.1005)
  - (6) 0.7031*** (0.1696)

- Sq(#Tax Staff/Labor Force):
  - (1) -1.9575*** (0.3909)
  - (2) -2.6749** (0.6285)
  - (3) -2.3093*** (0.3695)
  - (4) -3.0525*** (0.6388)
  - (5) -2.2392*** (0.3908)
  - (6) -2.9623*** (0.6689)

- Active Taxpayer/Labor Force:
  - (1) 0.0000 (0.0000)
  - (2) 0.0000 (0.0000)
  - (3) 0.0000 (0.0000)
  - (4) -0.0000 (0.0001)
  - (5) 0.0000 (0.0000)
  - (6) 0.0000 (0.0001)

- Real GDP growth, lagged:
  - (1) 0.0005*** (0.0000)
  - (2) 0.0006*** (0.0001)
  - (3) 0.0006*** (0.0000)
  - (4) 0.0008*** (0.0001)
  - (5) 0.0006*** (0.0000)
  - (6) 0.0007*** (0.0001)

- Trade openness (% of GDP), lagged:
  - (1) 0.0209*** (0.0030)
  - (2) 0.0244*** (0.0044)
  - (3) 0.0190*** (0.0029)
  - (4) 0.0203*** (0.0046)
  - (5) 0.0194*** (0.0028)
  - (6) 0.0213*** (0.0044)

- External debt (% of GDP), lagged:
  - (1) 0.0013 (0.0008)
  - (2) -0.0093 (0.0046)
  - (3) 0.0003 (0.0011)
  - (4) -0.0076** (0.0034)
  - (5) 0.0005 (0.0010)
  - (6) -0.0080** (0.0037)

- Inflation, lagged:
  - (1) -0.0100* (0.0039)
  - (2) -0.0027 (0.0046)
  - (3) -0.0117** (0.0045)
  - (4) -0.0066* (0.0038)
  - (5) -0.0113** (0.0045)
  - (6) -0.0057 (0.0036)

- Terms of Trade (2000=1), lagged:
  - (1) 0.0199** (0.0061)
  - (2) 0.0206** (0.0066)
  - (3) 0.0167* (0.0087)
  - (4) 0.0168* (0.0088)
  - (5) 0.0173** (0.0088)
  - (6) 0.0177* (0.0091)

- Oil exports (% of GDP), lagged:
  - (1) -0.1152*** (0.0126)
  - (2) -0.1390*** (0.0168)
  - (3) -0.0921*** (0.0163)
  - (4) -0.1075*** (0.0241)
  - (5) -0.0967*** (0.0138)
  - (6) -0.1150*** (0.0197)

- Financial Development Index, lagged:
  - (1) -0.0342*** (0.0057)
  - (2) -0.0148* (0.0054)
  - (3) -0.0253*** (0.0039)
  - (4) 0.0094 (0.0160)
  - (5) -0.0270*** (0.0041)
  - (6) 0.0036 (0.0150)

### Sample sizes and model diagnostics
- Observations:
  - (1) 527
  - (2) 422
  - (3) 527
  - (4) 422
  - (5) 527
  - (6) 422
- Number of countries:
  - (1) 120
  - (2) 98
  - (3) 120
  - (4) 98
  - (5) 120
  - (6) 98
- within R-squared:
  - (1) 0.197
  - (2) 0.231
  - (3) – 
  - (4) – 
  - (5) – 
  - (6) –
- AIC:
  - (3) -3443.0
  - (4) -2721.3
  - (5) -3462.0
  - (6) -2760.6
- BIC:
  - (3) -3425.9
  - (4) -2705.1
  - (5) -3444.9
  - (6) -2744.4

### Endogeneity tests and instrument diagnostics
- Ho: OSI is exogenous: columns (3)–(6) report 0 (reject exogeneity in IV specification).
- Cragg-Donald Wald F statistic:
  - (3) 37.281
  - (4) 23.350
  - (5) 43.561
  - (6) 28.509
- Kleibergen-Paap rk Wald F statistic:
  - (3) 9.5356
  - (4) 10.883
  - (5) 11.549
  - (6) 13.650

### Robustness and substantive implications from the table
- IV estimates (columns 3–6) show substantially larger and statistically significant positive coefficients on Operational Strength Index compared with OLS, implying a robust positive relationship between tax administration strength and Tax-to-GDP ratio excluding trade taxes and social contributions in IV specifications.
- Financial Development Index (lagged) coefficients are negative and statistically significant in several specifications (e.g., (1) -0.0342***; (3) -0.0253***; (5) -0.0270***), but in some IV specifications for EMDEs the Financial Development Index coefficient is small and not significant (e.g., (4) 0.0094; (6) 0.0036), indicating sensitivity to sample and instrument window.
- Tax Policy Yield is consistently positive and highly significant across all columns.
- Nonlinear staffing effects: #Tax Staff/LaborForce positive and Sq(#Tax Staff/Labor Force) negative and significant in all columns, consistent with diminishing marginal returns to tax staff per labor force.

*Source: Appendix Table AIII.24. Robustness: Adding Financial Development, wpiea2025219-source-pdf*

### Appendix Table AIII.30. Sensitivity: By Gov't Effectiveness, Avg. 2014-22

### Appendix Table AIII.30. Sensitivity: By Gov't Effectiveness, Avg. 2014-22

### Main estimated coefficients (columns (1) to (6))
- Column labels: (1) OLS All, (2) OLS EMDEs, (3) IV All [-4 to -2], (4) IV EMDEs [-4 to -2], (5) IV All [-5 to -2], (6) IV EMDEs [-5 to -2]
- Operational Strength Index [0,1]:
  - (1) 0.0078 (0.0071)
  - (2) 0.0185*** (0.0034)
  - (3) 0.1453*** (0.0188)
  - (4) 0.2855*** (0.0398)
  - (5) 0.1294*** (0.0147)
  - (6) 0.2522*** (0.0398)
- OSI X Gov't Effectiveness, Avg. 2014-22:
  - (1) 0.0074 (0.0176)
  - (2) 0.0207** (0.0046)
  - (3) 0.0880*** (0.0139)
  - (4) 0.1715*** (0.0262)
  - (5) 0.0851*** (0.0100)
  - (6) 0.1492*** (0.0186)
- Tax Policy Yield, exc. Trade Tax and SSC:
  - (1) 0.3029*** (0.0434)
  - (2) 0.2904*** (0.0423)
  - (3) 0.3491*** (0.0169)
  - (4) 0.3162*** (0.0539)
  - (5) 0.3440*** (0.0148)
  - (6) 0.3139*** (0.0454)
- #Tax Staff/LaborForce:
  - (1) 0.4426** (0.1143)
  - (2) 0.5079** (0.1293)
  - (3) 0.5076*** (0.1004)
  - (4) 0.6925*** (0.1047)
  - (5) 0.4952*** (0.1050)
  - (6) 0.6695*** (0.1074)
- Sq(#Tax Staff/Labor Force):
  - (1) -2.0011** (0.4467)
  - (2) -2.4420*** (0.4993)
  - (3) -2.0650*** (0.3809)
  - (4) -2.7829*** (0.4209)
  - (5) -2.0353*** (0.4078)
  - (6) -2.7402*** (0.4360)
- Active Taxpayer/Labor Force:
  - (1) 0.0000 (0.0000)
  - (2) 0.0000 (0.0000)
  - (3) -0.0000 (0.0000)
  - (4) -0.0001 (0.0001)
  - (5) -0.0000 (0.0000)
  - (6) -0.0001 (0.0001)
- Real GDP growth, lagged:
  - (1) 0.0006*** (0.0000)
  - (2) 0.0007*** (0.0001)
  - (3) 0.0006*** (0.0000)
  - (4) 0.0008*** (0.0001)
  - (5) 0.0006*** (0.0000)
  - (6) 0.0008*** (0.0001)
- Log (GDP per capita, USD), lagged:
  - (1) 0.0164 (0.0176)
  - (2) 0.0213 (0.0135)
  - (3) 0.0052 (0.0213)
  - (4) 0.0180 (0.0200)
  - (5) 0.0055 (0.0233)
  - (6) 0.0188 (0.0200)
- Sq(Log (GDP per capita, USD)), lagged:
  - (1) -0.0017 (0.0012)
  - (2) -0.0022* (0.0009)
  - (3) -0.0009 (0.0014)
  - (4) -0.0020 (0.0014)
  - (5) -0.0009 (0.0016)
  - (6) -0.0021 (0.0014)
- Trade openness (% of GDP), lagged:
  - (1) 0.0193*** (0.0028)
  - (2) 0.0231*** (0.0044)
  - (3) 0.0186*** (0.0030)
  - (4) 0.0214*** (0.0042)
  - (5) 0.0189*** (0.0027)
  - (6) 0.0215*** (0.0039)
- External debt (% of GDP), lagged:
  - (1) -0.0000 (0.0004)
  - (2) -0.0160*** (0.0026)
  - (3) -0.0011* (0.0006)
  - (4) -0.0113*** (0.0017)
  - (5) -0.0010 (0.0006)
  - (6) -0.0120*** (0.0017)
- Inflation, lagged:
  - (1) -0.0245*** (0.0034)
  - (2) -0.0251*** (0.0040)
  - (3) -0.0232*** (0.0024)
  - (4) -0.0334*** (0.0051)
  - (5) -0.0230*** (0.0022)
  - (6) -0.0325*** (0.0056)
- Terms of Trade (2000=1), lagged:
  - (1) 0.0183* (0.0069)
  - (2) 0.0197** (0.0067)
  - (3) 0.0115 (0.0098)
  - (4) 0.0101 (0.0080)
  - (5) 0.0122 (0.0094)
  - (6) 0.0114 (0.0082)
- Oil exports (% of GDP), lagged:
  - (1) -0.1140*** (0.0098)
  - (2) -0.1437*** (0.0148)
  - (3) -0.1001*** (0.0118)
  - (4) -0.1209*** (0.0199)
  - (5) -0.1027*** (0.0102)
  - (6) -0.1231*** (0.0177)
- Log (Agri, % of GDP), lagged:
  - (1) -0.0075 (0.0049)
  - (2) -0.0102* (0.0040)
  - (3) -0.0133** (0.0058)
  - (4) -0.0165*** (0.0048)
  - (5) -0.0125** (0.0059)
  - (6) -0.0158*** (0.0052)
- Control Corruption, lagged:
  - (1) 0.0024 (0.0014)
  - (2) 0.0036 (0.0038)
  - (3) 0.0002 (0.0012)
  - (4) 0.0025 (0.0042)
  - (5) 0.0005 (0.0012)
  - (6) 0.0026 (0.0041)

### Key statistics, samples, and diagnostics
- Observations:
  - (1) 529
  - (2) 424
  - (3) 529
  - (4) 424
  - (5) 529
  - (6) 424
- Number of countries:
  - (1) 121
  - (2) 99
  - (3) 121
  - (4) 99
  - (5) 121
  - (6) 99
- within R-squared:
  - (1) 0.216
  - (2) 0.275
  - (3) – 
  - (4) – 
  - (5) – 
  - (6) –
- Country FE: Yes (all columns)
- Year FE: Yes (all columns)
- AIC:
  - (3) -3430.1
  - (4) -2647.9
  - (5) -3447.9
  - (6) -2688.3
- BIC:
  - (3) -3413.0
  - (4) -2631.7
  - (5) -3430.8
  - (6) -2672.1
- Ho: OSI is exogenous – columns (3) to (6): 0 0 0 0

### Instrumental variables and strength tests
- Instrument: intensity of FTEs used in tax administration from IMF Capacity Development over the last four (five) years used in IV specifications (columns 3–6) when interacting confounding variable with OSI.
- Cragg-Donald Wald F statistic:
  - (3) 7.0805
  - (4) 6.0548
  - (5) 9.4040
  - (6) 7.1098
- Kleibergen-Paap rk Wald F statistic:
  - (3) 23.965
  - (4) 6.5085
  - (5) 33.452
  - (6) 8.4416

### Notes on estimation and inference
- Estimation method: OLS (columns 1 and 2) and IV (columns 3 to 6) panel fixed-effects regressions of expert-weighted OSI on the Tax-to-GDP ratio, excluding trade taxes and social contributions.
- Sample distinctions: Columns 1, 3, and 5 present regressions based on EMs, while columns 3, 4, and 6 focus on LICs (as described in the table notes).
- Interaction terms: When interacting the confounding variable with OSI, the additive term of the confounding variable is not included because it is absorbed by country fixed effects, given that it is time-invariant.
- Significance codes reported in table:
  - * p < 0.10
  - ** p < 0.05
  - *** p < 0.010
- Robust standard errors are reported in parentheses.

*IMF WORKING PAPERS Enhancing Tax Capacity: Revenue Gains from Strengthening Tax Administration — Appendix Table AIII.30. Sensitivity: By Gov't Effectiveness, Avg. 2014-22*

### Appendix Table AIII.36. Sensitivity: Alternative Dependent Variable: Taxes on Sales and Production to GDP

### Appendix Table AIII.36. Sensitivity: Alternative Dependent Variable: Taxes on Sales and Production to GDP Ratio

### Regression setup and identification
- Estimation methods: OLS (columns 1–2) and IV (columns 3–6) panel fixed-effects regressions.
- Dependent variable: Taxes on Sales and Production to GDP.
- Key explanatory variable: Operational Strength Index [0,1] (expert weighted OSI).
- Instrument for OSI in IV regressions: intensity of FTE used in tax administration IMF Capacity Development over the past four (columns 3–4) or five to two years (columns 5–6).
- Samples: Columns 1, 3, 5 = full sample; Columns 2, 4, 6 = EMDEs only.
- Fixed effects: Country FE = Yes; Year FE = Yes.

### Main coefficient estimates (selected variables across columns (1) to (6))
- Operational Strength Index [0,1]:
  - (1) OLS All: 0.0249***
    - (0.0016)
  - (2) OLS EMDEs: 0.0288***
    - (0.0054)
  - (3) IV All [-4 to -2]: 0.1059***
    - (0.0094)
  - (4) IV EMDEs [-4 to -2]: 0.1341***
    - (0.0112)
  - (5) IV All [-5 to -2]: 0.0765***
    - (0.0072)
  - (6) IV EMDEs [-5 to -2]: 0.0843***
    - (0.0201)

- Tax Policy Yield, exc. Trade Tax and SSC:
  - (1) 0.2002**
    - (0.0480)
  - (2) 0.2579***
    - (0.0341)
  - (3) 0.2206***
    - (0.0421)
  - (4) 0.2598***
    - (0.0264)
  - (5) 0.2132***
    - (0.0517)
  - (6) 0.2589***
    - (0.0310)

- #Tax Staff/LaborForce:
  - (1) 0.0127
    - (0.0711)
  - (2) 0.0691
    - (0.1059)
  - (3) 0.0711
    - (0.0534)
  - (4) 0.1419*
    - (0.0743)
  - (5) 0.0499
    - (0.0524)
  - (6) 0.1075
    - (0.0798)

- Sq(#Tax Staff/Labor Force):
  - (1) 0.0090
    - (0.1682)
  - (2) -0.4820
    - (0.2666)
  - (3) -0.1384
    - (0.1464)
  - (4) -0.5636***
    - (0.1825)
  - (5) -0.0850
    - (0.1403)
  - (6) -0.5250***
    - (0.2037)

- Active Taxpayer/Labor Force:
  - (1) -0.0000
    - (0.0000)
  - (2) 0.0000
    - (0.0000)
  - (3) -0.0000**
    - (0.0000)
  - (4) -0.0000
    - (0.0000)
  - (5) -0.0000**
    - (0.0000)
  - (6) -0.0000
    - (0.0000)

- Real GDP growth, lagged:
  - (1) 0.0004***
    - (0.0000)
  - (2) 0.0004***
    - (0.0000)
  - (3) 0.0004***
    - (0.0000)
  - (4) 0.0005***
    - (0.0000)
  - (5) 0.0004***
    - (0.0000)
  - (6) 0.0005***
    - (0.0001)

- Trade openness (% of GDP), lagged:
  - (1) 0.0048***
    - (0.0008)
  - (2) 0.0033**
    - (0.0012)
  - (3) 0.0047***
    - (0.0015)
  - (4) 0.0020
    - (0.0012)
  - (5) 0.0048***
    - (0.0013)
  - (6) 0.0026**
    - (0.0011)

- External debt (% of GDP), lagged:
  - (1) 0.0009***
    - (0.0002)
  - (2) -0.0146***
    - (0.0020)
  - (3) 0.0003
    - (0.0002)
  - (4) -0.0139***
    - (0.0030)
  - (5) 0.0005***
    - (0.0002)
  - (6) -0.0142***
    - (0.0022)

- Inflation, lagged:
  - (1) -0.0254***
    - (0.0028)
  - (2) -0.0192***
    - (0.0018)
  - (3) -0.0275***
    - (0.0021)
  - (4) -0.0283***
    - (0.0065)
  - (5) -0.0268***
    - (0.0019)
  - (6) -0.0240***
    - (0.0056)

- Oil exports (% of GDP), lagged:
  - (1) -0.0511***
    - (0.0099)
  - (2) -0.0374***
    - (0.0050)
  - (3) -0.0191*
    - (0.0109)
  - (4) -0.0018
    - (0.0037)
  - (5) -0.0307***
    - (0.0115)
  - (6) -0.0186***
    - (0.0057)

### Sample, fit, and diagnostic statistics
- Observations:
  - (1) 508
  - (2) 395
  - (3) 508
  - (4) 395
  - (5) 508
  - (6) 395
- Number of countries:
  - (1) 123
  - (2) 99
  - (3) 123
  - (4) 99
  - (5) 123
  - (6) 99
- within R-squared:
  - (1) 0.190
  - (2) 0.230
  - (3) – 
  - (4) – 
  - (5) – 
  - (6) –
- AIC:
  - (3) -3894.7
  - (4) -2972.2
  - (5) -3950.0
  - (6) -3056.7
- BIC:
  - (3) -3877.8
  - (4) -2956.3
  - (5) -3933.1
  - (6) -3040.8

### IV diagnostics and endogeneity tests
- Ho: OSI is exogenous:
  - (3) 0
  - (4) 0
  - (5) 0
  - (6) 0
- Cragg-Donald Wald F statistic:
  - (3) 27.032
  - (4) 14.298
  - (5) 33.034
  - (6) 17.772
- Kleibergen-Paap rk Wald F statistic:
  - (3) 9.8954
  - (4) 6.5730
  - (5) 11.981
  - (6) 8.9397

### Statistical significance notation
- * p < 0.10
- ** p < 0.05
- *** p < 0.010
- Robust standard errors reported in parentheses.

*Source: Appendix Table AIII.36 from IMF Working Paper "Enhancing Tax Capacity: Revenue Gains from Strengthening Tax Administration" (Working Paper No. WP/2025/219).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025219-source-pdf.pdf_