## _wp07183

## Source details

**Canonical URL:** [_wp07183](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2007/_wp07183.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2007/_wp07183.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2007/_wp07183.pdf.json)

---

### Introduction and research questions
- VAT now:
  - raises about 20 percent of the world’s tax revenue, and
  - affects about 4 billion people.
- Key empirical questions:
  - Has the VAT enhanced the efficiency of the overall tax system and aided revenue mobilization?
  - What has driven the global spread of the VAT?
- Empirical approach:
  - Unbalanced panel dataset for 143 countries over 26 years.
  - Indirect test: presence of a VAT should be associated with a higher tax ratio (tax revenues/GDP) if VAT reduces marginal deadweight loss.
  - Model endogeneity by estimating a VAT take-up (adoption) equation alongside revenue regressions.

### Background on the VAT: mechanisms and potential effects
- Essence:
  - Levied on all transactions but credited against tax due on traders’ sales so that, with unbroken crediting and correct border treatment, it ultimately bears only on final consumption.
- Potential efficiency gains:
  - Replaces turnover taxes and single-stage sales taxes.
  - Facilitates compensation for tariff-reduction packages.
  - Often central to modernizing tax administration and generalizing self-assessment.
- Potential weaknesses and risks:
  - Broken crediting chains can create production inefficiencies.
  - Large-scale fraud (example: around 1.5–2.5 percent of net revenue, or more, in the United Kingdom).
  - Imperfections in refund systems and excessive exemptions can make VAT function as a tax on exports and intermediates.
  - Interaction with informality: VAT may be inferior to tariffs in presence of a large informal sector (Emran and Stiglitz (2005); Keen (2006)).

### The revenue-impact framework and estimable equations
- Theoretical setup:
  - Representative consumer U(C,G), revenue R, gross income Y = Y(R,V) where V is a VAT presence dummy.
  - VAT increases efficiency iff it reduces marginal deadweight loss: RVY > 0.
  - Tax ratio r ≡ R / Y; first-order approximation (tax-effort form augmented for instrument efficiency):
    - VYYVr
      3210
      ββββ+++≈.
- Key proposition:
  - The sign of ∂r/∂V (conditional on observed Y) equals the sign of RVY. Thus, a positive association between VAT presence and the tax ratio implies VAT reduces marginal deadweight loss.
- Empirical implication:
  - Include VAT dummy and VAT interactions in tax-effort regressions; explicitly model VAT take-up to control for endogeneity.

### The decision to adopt a VAT (theory and conditions)
- Adoption modeled with an implementation cost K; VAT adopted iff inequality (5) in the source holds.
- Sufficient condition for adoption (6): adoption reduces deadweight loss by more than implementation costs.
- Theoretical results apply under both benevolent and non-benevolent (policy-maker–benefit L) objective formulations.
- Practical caveats: federal structures, administrative capacity, and distributional concerns can affect adoption and design.

### Empirical specification and estimation approach
- Revenue equation (equation (7) in source):
  - ittiititVititit uXVXVr+++ ′ + ′ +=λμββα
  - it V — VAT dummy; it X — other regressors; i μ — country fixed effect; t λ — time effect.
  - Error it u assumed spatially and temporally uncorrelated in baseline.
- Adoption equation (dynamic probit, equation (8) in source):
  - itV = 1 if it Z′δ + it ε ≥ 0, with it ε normally distributed and assumed spatially and temporally uncorrelated.
  - Lagged VAT dummy 1,−ti V included to capture state dependence.
- Logical-consistency and identification:
  - Recursive mixed system; current revenue excluded from Z to ensure logical consistency.
  - If it u and it ε independent: estimate revenue eqn by OLS/GMM and adoption by dynamic probit ML.
  - If correlated errors: two-step correction using auxiliary Mills-type variable M constructed from estimated probit (equation (9)) and included in revenue regression (equation (10)).

### Data and sample
- Sample period: 1975 to 2000.
- Final sample: unbalanced panel of 143 countries over 1975-2000 (Former Soviet Union and Central and Eastern Europe excluded).
- Revenue data: WEO, general government and grants; each country has at least ten years of data; mean number of years available is 23.4 out of a maximum of 26.
- VAT measurement: binary VAT dummies it V (presence/absence); design features not captured.
- Key explanatory variables (labels preserved): Ln(YPC), AGR, OPEN, POP, DEPYOUNG, DEPOLD, IMFCR, IMFNCR, FED, regional dummies (AP, AS, EU+, NME, SI, AFR), NEIGHBOUR.

### Main findings — Adoption equation (dynamic probit)
- Lagged VAT dummy (V-1):
  - Highly significant, evidence of strong state dependence.
  - Reported marginal effects approximately: 0.963, 0.951, 0.953, 0.951, 0.951 in columns 1–5; probit-scale values 4.644, 4.574 in columns 6–7.
- Openness (OPEN):
  - Generally negative coefficient; significant at 10 percent or better in several specifications.
  - Instrumenting OPEN by population (column 7) leaves coefficient negative, increases magnitude, remains significant at 10 percent.
- AGR (share of agriculture):
  - Negative and significant in several specifications (e.g., -0.654, -0.816 reported).
- NEIGHBOUR (regional VAT prevalence):
  - Robustly positive; examples of marginal effects: 0.4, 0.542, 0.4 with z-statistics indicating significance.
- IMF program participation:
  - IMFNCR (non-crisis programs): positive and significant (example marginal effect 0.236 with z-statistic (3.86)**).
  - IMFCR (crisis programs): less robust, smaller marginal effects.
- Lagged tax revenue (r-1):
  - Significantly negative (e.g., -1.011, -1.03) — higher existing revenue reduces likelihood of adopting VAT.
- FED (federal dummy):
  - Negative and significant in some specifications (federal countries less likely to adopt VAT), but effect less robust across specifications.
- Demographics (DEPOLD, DEPYOUNG):
  - No consistent significant effects.
- Interpretation:
  - Regional spillovers (NEIGHBOUR) and IMF non-crisis program participation raise adoption probability; greater openness lowers probability despite increasing conditional revenue gains.

### Main findings — Revenue equation (ln(r) regressions with country fixed effects)
- Fixed effects included in all regressions; Hausman test rejects random effects (example statistic: 92.33 rejecting at 1 percent).
- Baseline (Table 2, Column 1):
  - VAT dummy V = 0.034 (reported significant (2.70)**).
  - Interpretation: adoption increases the revenue ratio by 3.4 percent (about one half of a percentage point of GDP at an initial revenue ratio of 15 percent).
- Dynamics and persistence (Column 2 example):
  - Lagged dependent variable ln(r-1) = 0.626 (reported (13.65)**).
  - Short-run effect of VAT in Column 2: 1.7 percent (V = 0.017).
  - Long-run effect in Column 2: 4.54 percent = 1.7/(1-0.626) (exact expression preserved).
- Interaction effects:
  - ln(YPC)*V: robustly positive — higher income countries gain more from VAT adoption.
  - OPEN*V: generally positive — more open economies gain more from VAT adoption.
  - AGR*V, FED*V: included but less robust/significant.
- Other controls:
  - DEPOLD positively associated with the tax ratio (examples: DEPOLD coefficients ~1.101, 1.193 in some specifications).
  - IMF participation associated with higher revenue ratios (IMFCR and IMFNCR positive and significant in some columns).
- Selection correction (Mills term M):
  - Column 7: M = 0.0003 (z-statistic (0.04)); small and insignificant — no evidence of correlated errors between probit and revenue equations.

### Joint, robustness, and instrumentation
- Joint significance:
  - VAT and its interactions jointly significant in many specifications (example: F(5,2194)=3.56, p-value (0.003)).
- Region dummies interacted with VAT (Column 5):
  - Uniformly insignificant; primary heterogeneity captured by ln(YPC)*V and OPEN*V.
- Serial correlation and instruments:
  - Columns 2–7 include lagged dependent variable instrumented (third and fourth lags) to remove serial correlation.
  - Sargan tests in Columns 2–7 show instrument validity (examples: χ2(1)=0.069 (0.793), χ2(1)=0.084 (0.773)).

### Predicting revenue gains (prediction exercise using Column 4)
- Impact and long-run calculation:
  - Short-run predicted change in r associated with presence of VAT calculated from estimated coefficients and interactions (notation preserved in source).
  - Long-run predicted effect obtained by multiplying impact effect by 1/(1-λ), where λ is coefficient on lagged dependent variable.
  - Example: regression 4, λ = 0.586, so long-run effect = 2.41 times the impact effect (exact factor preserved as "2.41 times the impact effect").
- Illustrative regional sample predictions (from Table 3, based on Column (4) estimates):
  - For countries without a VAT by 2000 (mean impact effects, averaged 1990-2000; entries preserved as in source):
    - Americas: Average (rΔ)/r = 6.24
    - Asia-Pacific: 4.73
    - North Africa and Middle East: 0.02
    - Small Islands: 2.87
    - Sub–Saharan Africa: 0.91
  - For countries with a VAT (cumulated effects up to 2000; upper part of Table 3 examples):
    - Asia-Pacific: average (rΔ)/r = 2.10 (19 countries with Δr>0, 3 with Δr<0)
    - Americas: average = 0.51 (14 with Δr>0, 9 with Δr<0)
    - EU+: average = 4.15 (17 with Δr>0, 0 with Δr<0)
    - Sub–Saharan Africa: average = -0.81 (11 with Δr>0, 14 with Δr<0)
- Critical income levels by region (Table 3 exact figures):
  - Asia-Pacific: 1,546
  - Americas: 3,027
  - EU+: 4,462
  - North Africa and Middle East: 3,083
  - Small Islands: 1,785
  - Sub–Saharan Africa: 1,124
- Interpretation:
  - Predictions concern revenue/efficiency gains once VAT is in place; they do not account for implementation or transition costs.
  - Mills term insignificance implies no need for sample-selection correction in these predictions.
  - Broad tendency: most countries—both adopters and non-adopters by 2000—predicted to gain in revenue from VAT; sub-Saharan Africa is the least clear region.

### Policy-relevant implications (authors’ presentation)
- Adoption determinants:
  - Countries with low lagged revenue more likely to adopt VAT.
  - Participation in IMF non-crisis programs raises likelihood of adoption.
  - Regional spillovers (NEIGHBOUR) and federal structure (FED) matter.
  - Greater openness reduces probability of adoption but increases conditional revenue benefit.
- Revenue effects and heterogeneity:
  - Presence of a VAT associated with higher revenue ratios; short-run and long-run effects differ (example short-run 1.7 percent, long-run 4.54 percent in Column 2).
  - Gains heterogeneous and depend on country characteristics (ln(YPC)*V and OPEN*V positive).
  - Higher income and more open countries tend to realize larger revenue gains from VAT.
- Implementation considerations:
  - Adoption may impose substantial administrative and compliance costs, especially where tariff-based systems dominate and inland administration is weak.
  - Policy choice should weigh predicted revenue/efficiency gains against implementation and transition costs specific to country circumstances.

### Summary and conclusions (concise)
- Study: panel of 143 countries over 25/26 years; recursive two-equation system for VAT adoption and revenue impact.
- Determinants of adoption:
  - Income per capita is not a necessary determinant; agriculture share reduces likelihood; lower past revenue raises likelihood; IMFNCR increases likelihood; NEIGHBOUR raises likelihood; openness reduces likelihood.
- Impact of VAT on revenue:
  - Simple specification: VAT adoption associated with a long-run increase in overall revenue-to-GDP ratio of about 4.5 percent.
  - Allowing interactions: shift effect becomes negative but countervailing gains greater in higher income and more open economies.
  - Predicted country-level effects: most adopters and most non-adopters predicted to gain in revenue; sub-Saharan Africa shows mixed evidence and possible offsetting effects from reduced tariff revenue.

*Source: _wp07183*

### 1. Estimates of the Adoption Equation ..................................................................................

### 1. Estimates of the Adoption Equation

### Introduction: scope and questions
- The value-added tax (VAT) now:
  - raises about 20 percent of the world’s tax revenue, and
  - affects about 4 billion people.
- Key empirical questions posed:
  - Has the VAT lived up to claims that it enhances the efficiency of the overall tax system and aids revenue mobilization?
  - What has driven the global spread of the VAT?
- Motivations for the study:
  - The VAT has become central to tax reform in many developing countries.
  - Little prior theoretical or empirical work directly addresses (i) whether VAT adoption increases total tax revenue (the “money machine” idea) and (ii) determinants of VAT take-up.
- Empirical approach overview:
  - Use an unbalanced panel dataset for 143 countries over 26 years.
  - Adopt an indirect test: show that access to a more efficient tax instrument should lead an optimizing government to increase the tax ratio (tax revenues/GDP), then test whether presence of a VAT is associated with higher tax ratios.
  - Model endogeneity by explicitly modeling the VAT take-up decision and estimating a take-up (adoption) equation alongside revenue regressions.

### Background on the VAT: mechanisms and potential effects
- Essence of the VAT:
  - Levied on all transactions but credited against tax due on traders’ sales so that, when the crediting chain is unbroken and tax is properly charged on imports and remitted on exports, it ultimately bears only on final consumption.
- Potential sources of efficiency gain from VAT adoption:
  - Replaces turnover taxes (which bear on intermediate transactions and induce production inefficiencies).
  - Replaces single-stage sales taxes (vulnerable to evasion/avoidance and often resulting in high rates on narrow bases).
  - Facilitates compensation for tariff-reduction packages (preserves production efficiency while offsetting revenue losses).
  - Often central to modernizing tax administration and generalizing self-assessment, which may ease administration and compliance across other taxes.
- Potential weaknesses and risks:
  - Broken crediting chains can create production inefficiencies.
  - Large-scale fraud such as “carousel fraud”—example: amounted to around 1.5–2.5 percent of net revenue, or more, in the United Kingdom (Keen and Smith, 2006).
  - Imperfections in refund systems and excessive statutory exemptions can make VAT function as a tax on exports and intermediates, potentially reducing exports and national output.
  - Interactions with informality: expanding consumption tax base can increase inefficient informal production; VAT may be inferior to tariffs in presence of informal sector (Emran and Stiglitz (2005)), with important qualifications about VAT incidence on imports in developing countries (Keen (2006)).

### The revenue impact of a tax innovation: framework and estimable equations
- Setup:
  - Representative consumer with preferences U(C,G) over private consumption C and public expenditure G.
  - Public expenditure financed by taxation generating revenue R; gross income Y = Y(R,V) depends on level of taxation R and tax instruments parameter V (V is a dummy in empirics: V = 1 if VAT present, 0 otherwise).
  - Pre-tax income decreases with R: 0 > Y_R; marginal deadweight loss increases with revenue so 0 < Y_RR.
  - VAT increases efficiency iff it reduces marginal deadweight loss: RVY > 0.
- Benevolent government first-order condition (modified Samuelson rule), equation (1):
  - 11 > − = (U_C / U_G) R
- Implicit definition and linear approximation of tax ratio:
  - Define tax ratio r ≡ R / Y.
  - Implicit condition rewritten as equation (2):
    - 0],)1[()1),(](,)1[(),,(
    =−+−−≡rYYrUVrYYrYYrUYVrH
    GRC
  - First-order approximation (standard tax effort form augmented for tax instrument efficiency), equation (3):
    - VYYVr
      3210
      ββββ+++≈.
- Key theoretical proposition:
  - PROPOSITION: ),( YVr V has the same sign as RVY.
  - Interpretation: The tax ratio will be positively associated with the presence of a VAT if and only if deployment of a VAT reduces the marginal deadweight loss of the tax system. As it becomes less costly to raise revenue at the margin, more will optimally be raised; conditional on observed income Y, the tax ratio rises.
- Empirical implications:
  - Add a VAT presence dummy and interaction terms to tax-effort style regressions to test the “money machine” hypothesis (whether VAT presence is associated with higher tax ratios).
  - Model the VAT take-up decision explicitly to control for endogeneity in estimating VAT impacts.

### Literature context and empirical gaps noted
- Sparse literature on VAT despite widespread adoption.
- Prior empirical studies:
  - Some model VAT revenue and compliance; a few estimate revenue-maximizing rates.
  - Desai and Hines (2005): openness and export performance negatively related to both presence of VAT and revenue reliance on it (particularly outside high-income countries).
  - Ebrill and others (2001): similar questions but using single cross-section and treating VAT presence as exogenous.
  - Keen and Lockwood (2006): explore “money machine” theoretically and empirically for OECD panel where VAT revenue data are available; term the positive association between VAT presence and total revenue the ‘weak’ money machine hypothesis.
- Novelty of current approach:
  - Uses panel data across 143 countries and 26 years.
  - Explicitly models VAT adoption (take-up) to address endogeneity and to investigate determinants of adoption (including roles of regional emulation and international financial institutions).

*Source: _wp07183 - 1. Estimates of the Adoption Equation*

### introduction of a VAT on the grounds that it may lead to an undesirable increase in the size

### _wp07183 - introduction of a VAT on the grounds that it may lead to an undesirable increase in the size

### Theoretical framing
- The presence of a VAT is related to total tax revenue, not simply to revenue from sales taxation; the VAT must be evaluated as part of a wider tax system.
- Empirical finding that the presence of a VAT is associated with a higher tax ratio bears on the efficiency of the VAT, not on its welfare impact.
- The framework accommodates non-benevolent policy makers by allowing the policy maximand to include a public expenditure item L that benefits only the policy-maker. The policy problem is written in the text in the form:
  - "R max { L max [] ),),((,LRRRVYUL−−Ψ ≡ R max ],),([ * RRRVYU− ,  (4)" (as presented in the source).
- Proposition 1 continues to apply regardless of the preference structure; the derivative of the tax ratio with respect to V, conditional on observed income, has the same sign as RVY (as stated in the source).
- Distributional concerns may lead governments to alter VAT rate and exemption structure, but this is expected to mitigate rather than overturn the efficiency gains motivating VAT adoption.

### The Decision to Adopt a VAT
- Adoption is a major policy decision that often involves changes in tax administration, information campaigns, and a shift to systematic self-assessment; adoption is costly and far from exogenous.
- Implementation of the VAT is modeled as involving an additional real resource loss of fixed amount K, reducing revenue available to finance the public good G.
- The VAT is adopted if and only if the following inequality (presented as (5) in the source) holds:
  - "G max],),1([GKGKGYU−−+ > G max ],),0([GGGYU−            ,                                    (5)            where 1=V refers to the state in which the VAT is adopted and 0=V to that in which it is not."
- A sufficient condition for adoption, presented as (6) in the source, is:
  - "),,0(),1( GYKGYK−+< G∀                                                (6)"
  - Interpretation: adoption unambiguously reduces deadweight loss from the tax system by more than it increases implementation costs, provided (6) holds.
- The same conclusion holds under the non-benevolent policy-making formulation discussed earlier.
- Practical considerations beyond deadweight and implementation costs can influence adoption (e.g., federal states allocating sales tax powers to subnational jurisdictions raise particular implementation difficulties). Footnote reference: 12.
- Distributional policies may be handled better by other tax and spending instruments; examples include developing-country spending measures for poverty alleviation (footnote 13) and nuanced assessments of VAT regressivity (footnote 14).

### Limitations (as stated in the source)
- Interpreting a positive association between total revenues and the presence of a VAT as a supply-side gain assumes that additional variables in both revenue and adoption equations adequately control for changes in general demand for public expenditure.
- Direct tests distinguishing demand- and supply-based explanations require data on revenue from the VAT itself, which the authors state is not available for a panel as large as theirs, especially for lower income countries.
- The analysis focuses on the VAT's impact in the revenue equation (an ex post gain once a VAT is in place) but does not estimate adoption costs K due to the absence of an explicit functional form for the policy maximand.

### Empirical specification

- A. The Revenue Equation
  - The revenue ratio r is modeled as presented in the source (equation (7)):
    - "ittiititVititit uXVXVr+++ ′ + ′ +=λμββα    (7)"
  - Notation and assumptions preserved verbatim from the source:
    - i and t are country and time indicators respectively.
    - it V is a dummy variable taking the value unity if country i has a VAT in year t and zero otherwise.
    - it X captures other variables affecting the tax ratio.
    - i μ is a country-specific effect, t λ is a time effect.
    - The idiosyncratic error it u is assumed to be both spatially and temporally uncorrelated: "0][= jtit uuE for ,ij≠ and 0][= isit uuE for .ts≠"
  - The VAT-augmented revenue equation allows the effect of the VAT to be more complex than a simple intercept shift, entering in interaction with other variables.

- B. The Adoption Equation
  - Adoption is modeled by supposing the welfare difference is given by , ' ititit ZWεδ−=Δ for explanatory variables Z and errors ε, leading to the threshold model (equation (8)) presented in the source:
    - "⎩ ⎨ ⎧ < ≥ = itit itit it Z Z V εδ εδ ',0 ',1                                     (8)"
  - it ε is assumed normally distributed and both spatially and temporally uncorrelated; adoption is estimated as a probit with:
    - " ]Pr[]1Pr[ ititit ZVδε ′ ≤==, where (.)Φ denotes the standard normal distribution (and (.) φ will denote its density)."
  - The lagged VAT dummy 1,−ti V is included among regressors in it Z to allow for observed persistence/sluggishness in VAT adoption; the source notes only five countries have ever removed a VAT (two in their sample: Belize and Malta), so the coefficient on 1,−ti V will be dominated by persistence.

- C. Logical Consistency and Cross Equation Restrictions
  - The source highlights the need for attention to logical consistency in two-equation systems with one binary variable, referencing Maddala (1983), and indicates this is addressed in the continuation of the text.

*Source: _wp07183 - introduction of a VAT on the grounds that it may lead to an undesirable increase in the size*

### Appendix II to have two important implications. The first is that current revenue

### _wp07183 - Appendix II to have two important implications. The first is that current revenue

### Model and Estimation Approach
- System: revenue equation (7) and adoption equation (8) form a recursive mixed system of continuous (revenue) and discrete (adoption) dependent variables.
- Identification: both equations are always identified; logical consistency implies current revenue cannot be an element of Z.
- Estimation when shocks independent:
  - Revenue equation (7) estimated by OLS or GMM (with lagged dependent variable as instrument).
  - Adoption equation (8) estimated as a dynamic probit (“probit with state dependence”) by maximum likelihood; initial VAT dummies 0iV taken as exogenous.
- Two-step correction if correlation between revenue error (it u) and adoption error (it ε):
  - Dynamic probit estimated first; estimate δ̂ used to construct auxiliary variable M (equation (9)).
  - Revenue equation then estimated with M included (equation (10)); estimated coefficient on M is an estimate of covariance between it u and it ε.
- Fixed effects cannot be included in the dynamic probit (8).

### Data and Sample
- Sample period: 1975 to 2000.
- Final sample: unbalanced panel of 143 countries over 1975-2000 (Former Soviet Union and Central and Eastern Europe excluded).
- Revenue data: from the World Economic Outlook (WEO), relate to general government and grants.
  - Each country has at least ten years of data on revenue; mean number of years available is 23.4 out of a maximum of 26.
- VAT measurement: binary VAT dummies itV describing whether country i had a VAT in place at time t (yes/no distinction; design features not captured due to data limitations).
- Key explanatory variables used (exact labels preserved):
  - Ln(YPC), AGR (share of agriculture in GDP), OPEN (sum of GDP shares of imports and exports), POP, DEPYOUNG, DEPOLD, IMFCR, IMFNCR, FED, regional dummies (AP, AS, EU+, NME, SI, AFR), NEIGHBOUR (proportion of countries in same region with a VAT).

### Adoption of the VAT — Main Findings
- Estimation: dynamic probit; coefficient estimates reported as marginal effects in first five columns of Table 1 (read as impact on probability of VAT adoption evaluated at mean of regressors).
- Lagged VAT dummy (V-1) highly significant and positive in all specifications (evidence of state dependence).
  - V-1 reported marginal effects approximately: 0.963, 0.951, 0.953, 0.951, 0.951 in columns 1–5; and larger point values in columns 6–7 (4.644, 4.574) when shown in probit form.
- Openness (OPEN):
  - Generally enters with a negative coefficient; significant at 10 percent or better in several specifications.
  - Instrumenting OPEN by population (column 7) leaves coefficient negative, increases magnitude, and remains significant at 10 percent.
- AGR (agriculture share): negative and significant in several specifications (e.g., -0.654, -0.816 in Table 1 columns).
- NEIGHBOR: robustly positive; examples of marginal effects: 0.4, 0.542, 0.4, 0.4 with z-statistics indicating significance (3.28)** etc.
- IMF program participation:
  - IMFNCR (non-crisis development/poverty relief programs): positive and significant (e.g., marginal effect 0.236 with z-statistic (3.86)**).
  - IMFCR (crisis programs) less robust, smaller marginal effects.
- Lagged tax revenue (r-1): significantly negative (e.g., -1.011, -1.03) — higher existing revenue reduces likelihood of adopting VAT.
- FED dummy: negative and significant in some specifications (federal countries less likely to adopt a VAT), but effect less robust across specifications.
- Demographic variables (DEPOLD, DEPYOUNG): no consistent significant effect on adoption.
- Regional clustering: significant “regional bursts” in VAT adoption captured by NEIGHBOUR; effect not explained away by common deeper causes.

### Revenue Impact of the VAT — Main Findings
- Dependent variable: ln(r) (log revenue ratio). Country fixed effects included in all regressions (Hausman test rejects random effects; example statistic: 92.33, rejecting at 1 percent for preferred specification).
- Baseline (Table 2, Column 1):
  - VAT dummy V = 0.034 (reported significant (2.70)**).
  - Interpretation: adoption increases the revenue ratio by 3.4 percent (about one half of a percentage point of GDP at an initial revenue ratio of 15 percent).
- Dynamics included (Columns 2–7):
  - Lagged dependent variable ln(r-1) ~ 0.626 in Column 2 (reported (13.65)**).
  - Short-run effect of VAT in Column 2: 1.7 percent (V = 0.017).
  - Long-run effect in Column 2: 4.54 percent = 1.7/(1-0.626) (exact expression preserved as reported).
- Interaction effects (Column 3 onward):
  - ln(YPC)*V: robustly positive (example reported t-statistics, and noted as significant), indicating higher income countries gain more from VAT adoption.
  - OPEN*V: generally positive (examples show significance at 5 percent or 10 percent in specifications), implying more open economies gain more from VAT adoption.
  - AGR*V, FED*V: interaction terms included but less robust/significant.
- Other robust controls:
  - DEPOLD positively associated with the tax ratio (examples in Table 2: DEPOLD coefficients ~1.101, 1.193 with significance in some specifications).
  - IMF participation associated with higher revenue ratios (IMFCR and IMFNCR positive and significant in some columns).
- Mills ratio M (correcting for potential covariance between adoption and revenue errors):
  - In Column 7, M reported as 0.0003 (with z-statistic (0.04)); small and insignificant — no evidence of correlated errors between probit and revenue equations.

### Joint and Regional Results
- Joint significance: F-tests indicate joint significance of VAT and its interactions in many specifications (e.g., F(5,2194)=3.56, p-value (0.003) in Table 2).
- Region dummies interacted with VAT (Column 5) were uniformly insignificant; interactions with VAT mainly captured via ln(YPC)*V and OPEN*V.
- Serial correlation and instrument validity:
  - Column 1 exhibited serial correlation; Columns 2–7 include lagged dependent variable instrumented (third and fourth lags) removing serial correlation.
  - Sargan tests in Columns 2–7 indicate instrument validity (examples: χ2(1)=0.069 (0.793), χ2(1)=0.084 (0.773), etc., p-values in brackets).

### Predicting Revenue Gains — Approach and Illustrative Results
- Regression used for prediction: Column 4 of Table 2 selected for prediction exercise.
- Short-run predicted change in r associated with presence of VAT (equation (11)):
  - itV FED itV AGR itV OPEN itV YPC itFEDAGROPENYPC r
  - Δˆ = βˆ * (interactions)  (preserved notation as in source: itVFEDitVAGRitVOPENitVYPCit ****ˆˆˆˆlnˆr=Δ with VX*ˆβ the estimated coefficient on the interaction between X and the VAT dummy).
- Long-run predicted effect obtained by multiplying impact effect by 1/(1-λ), where λ is coefficient on lagged dependent variable.
  - Example: from regression 4, λ = 0.586, so long-run effect = 2.41 times the impact effect (exact factor preserved as "2.41 times the impact effect").
- Sample predictions (from Table 3, calculated from point estimates in Column (4) of Table 2; regions defined as in source):
  - For countries without a VAT by 2000: mean impact effects (averaged 1990-2000) by region — examples reported:
    - Americas: Average (rΔ)/r = 6.24 (for non-VAT group: "3 0 6.24" entry).
    - Asia-Pacific: 4.73 (entry "6 1 4.73").
    - North Africa and Middle East: 0.02 (entry "9 0 0.02").
    - Small Islands: 2.87 (entry "13 1 2.87").
    - Sub–Saharan Africa: 0.91 (entry "8 6 0.91").
  - For countries with a VAT (cumulated effects up to 2000): regional mean examples (upper part of Table 3):
    - Asia-Pacific: average (rΔ)/r = 2.10 (19 countries with Δr>0, 3 with Δr<0).
    - Americas: average = 0.51 (14 with Δr>0, 9 with Δr<0).
    - EU+: average = 4.15 (17 with Δr>0, 0 with Δr<0).
    - Sub–Saharan Africa: average = -0.81 (11 with Δr>0, 14 with Δr<0).
  - Critical income levels by region (Table 3, exact figures preserved):
    - Asia-Pacific: 1,546
    - Americas: 3,027
    - EU+: 4,462
    - North Africa and Middle East: 3,083
    - Small Islands: 1,785
    - Sub–Saharan Africa: 1,124
- Interpretation of predictive exercise:
  - Predictions relate only to revenue and efficiency gains from VAT once adopted (do not account for implementation or transition costs).
  - Mills term insignificance implies no need to correct for sample selection in these predictions.
  - Broad tendency: most countries—both those with and without a VAT by 2000—are predicted to gain in revenue from implementing a VAT; the picture is least clear for sub-Saharan Africa where gains and losses are more balanced.

### Policy-relevant Implications (as presented in source)
- Adoption determinants:
  - Countries with low lagged revenue are more likely to adopt a VAT; participation in IMF non-crisis programs raises likelihood of adoption.
  - Regional spillovers (NEIGHBOR) and federal structure (FED) matter for adoption decisions.
  - Greater openness reduces the probability of adoption but increases the revenue benefit conditional on adoption.
- Revenue effects:
  - Presence of a VAT associated with higher revenue ratios in many settings; short-run and long-run effects differ (example short-run 1.7 percent, long-run 4.54 percent in Column 2).
  - Gains from VAT are heterogeneous and depend on country characteristics (income per capita, openness, AGR, FED).
  - Higher income countries tend to realize larger revenue gains from VAT (ln(YPC)*V positive).
  - More open economies tend to derive larger revenue gains from VAT (OPEN*V positive), suggesting border-based VAT collection advantages.
- Implementation considerations:
  - Although VAT can raise revenue and improve tax system effectiveness, adoption may impose substantial administrative and compliance costs (particularly where tariff-based revenue systems have not developed inland tax administration).
  - Policy choice should weigh predicted revenue/efficiency gains against implementation and administrative transition costs specific to country circumstances.

*Source: Authors’ calculations and analysis from the IMF working paper content provided.*

### Appendix 1 of Ebrill et al (2001). For co

### Appendix 1 of Ebrill et al (2001). For co

### VI. SUMMARY AND CONCLUSIONS
- Study: panel of 143 countries observed over 25 years; estimates a recursive two-equation system for (i) decision to adopt a VAT and (ii) impact of VAT on revenue once adopted.
- Determinants of VAT adoption:
  - Income per capita is not a determinant: "the VAT is no longer the preserve of high income countries."
  - Larger agricultural sector reduces likelihood of adoption.
  - Lower past revenue ratio increases likelihood of adoption (greater need to enhance revenues).
  - Participation in a non-crisis IMF program significantly increases likelihood of adopting a VAT.
  - Yardstick competition: higher proportion of regional adopters increases likelihood of adoption.
  - More open economies are less likely to adopt a VAT (striking result).
- Impact of VAT on revenue:
  - Simple specification: VAT adoption associated with a long run increase in overall revenue-to-GDP ratio of about 4.5 percent.
  - Allowing interaction with country circumstances:
    - Shift effect becomes negative.
    - Countervailing gains tend to be greater in higher income and more open economies.
      - Higher income: plausibly greater ability to administer and comply.
      - More open economies: easier to collect VAT on imports; lower likelihood of adoption may reflect ability to rely on tariff revenue instead of modernizing tax administration.
  - Predicted revenue effects calculated for each country indicate:
    - Adoption has increased revenue and improved effectiveness in the large majority of countries that implemented a VAT, and would do so for most non-adopters.
    - Sub-Saharan Africa: evidence is mixed.
      - Possible interpretation: VAT may have improved tax system efficiency but offset by reduced use of other taxes (notably tariffs).
      - Suggests significant work remains to improve VAT functioning in much of the region.

### APPENDIX I — PROOF OF THE PROPOSITION (selected expressions and conclusions)
- From necessary condition (2): ,/ rVV HHr−= with RVCV YUH=  (equation (A.1)).
- Show  .0< r H  by deriving:
  - (A.2): )}1({})1({ −+++−−= RCGGGCGRCCRRCr YUUUYUYYUH
  - (A.3): rearranged expression for H involving C, G, U, Y.
  - (A.4): GCG RRRC r YYU Y H ααα++−= 
    - αG defined in (A.5); αC defined analogously.
  - Quasi-concavity implies  ,0<+ CG αα and C is normal iff 0< G α.
  - With maintained assumptions that both RY and RRY are strictly negative, and C normal, (A.4) implies .0< r H.

### APPENDIX II — LOGICAL CONSISTENCY (key implications)
- Adoption decision depends on it r and other variables it Z. From (8):
  - 1 = it V iff ititit Zrεδγ≥+'.
  - 0 = it V iff ititit Zrεδγ<+'.
- Logical consistency requires probabilities summing to unity; leads to (A.6) and (A.7).
- Either (i) 0 = γ, or (ii) 01== VitVit rr.
  - From (7), latter generally requires 0 == V βα which rules out any effect of V on r.
- Authors assume 0 = γ (no feedback from it r to it V).
  - Logical consistency then dictates interaction only through presence of it V in revenue equation and correlation of errors it u and it ε.
  - This does not preclude lagged revenue in adoption equation or overlap between it Z and exogenous variables it X.
- With 0 = γ, conditions for identification are very weak: no restrictions required on it u or it ε (see Maddala (1983), model 5).

### APPENDIX III — ESTIMATION OF THE REVENUE EQUATION WITH CORRELATED ERRORS (key expressions)
- From (7):
  - (A.8): ][]1[ ititittiitVititit ZuEXXVrEεδλμββα≥+++ ′ + ′ + == 
  - (A.9): ][]0[ ititittiititit ZuEXVrEεδλμβ<+++ ′ == 
  - Standard formulae yield conditional expectations in (A.10) involving Φ and δε, σεε, δ, Z.
- Combining gives (A.11): expression for E[it V X X V r] with auxiliary variable Mit defined as the term in square brackets: the combination of Φ and δ terms.

### APPENDIX IV — PROOF THAT 1−Ω = Er Y (key steps)
- From (2): CGRCCrY UYU Y r HH +−+ ⎟⎠⎞ ⎜⎝⎛ =)1(  (A.11)
- Using (1) and definition in (A.5) yields (A.12): C rY Y r HH α− ⎟⎠⎞ ⎜⎝⎛ =
- Recalling (A.3) and rearranging gives (A.13): expression linking r, Y, H and α terms.
- Define Ω as in (A.14): ]1,0[⊂ ++− + ≡ Ω CGG RRRC CG YYU ααα αα .
- Recalling Hicks (1939) that derivative of demand for G with respect to lump sum income equals ),/( CGC ααα+ the result 1−Ω = Er Y follows.

### Table 4 — Data appendix: Variable definitions, sources, and summary statistics (selected entries)
- Sample: 143 countries over 25 years; country list provided in source.
- RY: ratio of general government tax revenue plus grants to GDP; Source: World Economic Outlook, IMF; Obs 3431; Mean 0.295; Min 0.0008048; Max 1.663265.
- V: Dummy (=1 if country has VAT, =0 otherwise); Source: Table 1.3, Ebrill, Keen, Bodin and Summers (2001), updated by authors; Obs 3744; Mean 0.337; Min 0; Max 1.
- NEIGHBOR: Proportion of other countries in the region that have already adopted a VAT; Obs 3744; Mean 0.353; Min 0; Max 1.
- OPEN: Value of imports plus exports as fraction of GDP; Source: World Bank Development Indicators; Obs 3302; Mean 0.780; Min 0.011; Max 4.390.
- YPC: GDP per capita at PPP (units: 1000 constant 2000 US dollars); Obs 3308; Mean 7.911; Min 0.466; Max 57.792.
- AG: Share of agriculture in value-added; Obs 3201; Mean 0.194; Min 0.001; Max 0.743.
- POP: Total population, millions; Obs 3714; Mean 30.345; Min 0.041; Max 1262.46.
- DEPYOUNG: Population 14 or younger, fraction of total; Obs 3505; Mean 0.368; Min 0.143; Max 0.552.
- DEPOLD: Population 65 or over, fraction of total; Obs 3505; Mean 0.056; Min 0.011; Max 0.181.
- FED: Dummy (1 for federal countries, 0 for unitary); Source: Triesman(2002); Obs 3888; Mean 0.146; Min 0; Max 1.
- IMFNCR: Dummy =1 if in a non-crisis IMF program for 6 months+ in calendar year; Source: IMF; Obs 3744; Mean 0.087; Min 0; Max 1.
- IMFCR: Dummy =1 if in a crisis IMF program for 6 months+ in calendar year; Obs 3744; Mean 0.133; Min 0; Max 1.
- Regional dummies: EU, AS, AF, CBRO, NMED, AP, SI defined per Table A1.1, Ebrill, Keen, Bodin and Summers (2001).

*Source: Appendix 1 of Ebrill et al (2001), as provided in the supplied content.*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2007/_wp07183.pdf_