## 1. Finite Mixture Model Estimations

## Source details

**Canonical URL:** [1. Finite Mixture Model Estimations](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019286-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2019/wpiea2019286-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019286-print-pdf.pdf.json)

---

### Introduction and key empirical findings
- Dataset and scope:
  - Bank of International Settlements (BIS) restricted locational banking statistics recording foreign-owned deposits in 39 countries held by residents of over 200 counterparty jurisdictions on a quarterly basis since 1995.
  - Sample of over 230,000 observations.
- Main estimated effects of exchange of information (EOI) frameworks on deposits in offshore jurisdictions:
  - Bilateral agreements related to EOI upon request (EOIR) are associated with deposits held in offshore jurisdictions decreasing by between 8 and 12 percent following conclusion of an EOIR agreement (relationship has weak statistical significance).
  - Automatic exchange of information (AEOI) agreements are estimated to reduce deposits in offshore jurisdictions by around 25 percent (statistically significant).
- Round-tripping and the US FATCA experience:
  - No evidence that the US automatic exchange initiative, FATCA, has reduced round-tripping to the US.
  - Deposits from offshore centers into the US increased following the introduction of FATCA Intergovernmental Agreements (IGAs); possible contributing factors include the removal of a withholding tax threat, a lack of information exchange reciprocity by the US, and the non-imposition of tax on certain US sourced portfolio interest received by non-residents.

### Contribution to literature and methodological advances
- Coverage and extensions:
  - Analyze effects of all existing bilateral EOI agreements over the last 20 years including over 3,000 EOIR agreements, 85 FATCA agreements, the European Union Savings Directive and its related arrangements, and the CRS.
- Identification improvements:
  - Implement a triple-differenced estimator to estimate effects on offshore and non-offshore jurisdictions separately and use the difference as the estimator of EOI’s impact on tax evasion.
- Sensitivity to offshore classification:
  - Use a finite mixture model to estimate offshore-status and EOI impact simultaneously, addressing subjectivity in offshore lists.

### Finite mixture model purpose and classification results
- Method and algorithm:
  - Finite mixture model (EM-algorithm) jointly estimates probabilistic classification of countries into offshore centers and the response of deposits to EOI.
  - Iterative three-step EM procedure: (1) weighted estimation of deposit-response equations, (2) compute posterior probabilities of offshore status, (3) regress posteriors on predictors to update priors.
- Main classification findings:
  - Offshore centers are typically associated with:
    - low income tax rates (personal and corporate),
    - high financial secrecy,
    - English as an official language,
    - large FDI stocks,
    - small trade flows (relative to GDP).
  - UN membership is a significant predictor (in the finite mixture results).
  - Governance indicators are not statistically significant predictors of offshore status in this analysis.
- Consistency:
  - Finite mixture model estimates are consistent with baseline regressions: AEOI, but not EOIR, impacts cross-border deposits in offshore jurisdictions.

### Background on EOI frameworks (chronology and features)
- EOIR:
  - EOIR historically relied on bilateral DTTs or TIEAs; OECD amended Model Tax Convention in 2005 by adding paragraph 4 to Article 26.
- EU Savings Directive (EUSD):
  - Agreement reached June 2003; came into effect on 1 July 2005.
  - Repealed and superseded in part by Council Directive 2014/107/EU implementing broader AEOI consistent with the CRS as of 1 January 2016.
- FATCA:
  - Enacted in 2010 as part of the HIRE Act; went into effect on July 1, 2014.
  - To date in the paper: 113 jurisdictions have reached agreement with the US to undertake AEOI under FATCA IGAs (of which 88 IGAs are in force, 13 are signed but not yet in force, and 12 are agreed in substance).
  - Projected revenue and costs cited:
    - Projected to generate $8.7 billion in revenue between fiscal years 2010-2020, a yearly average of $792 million.
    - IRS has spent approximately $380 million to implement and administer the FATCA program.
- CRS and multilateral AEOI:
  - CRS final version released July 2014 by OECD.
  - As of November 2018, 101 jurisdictions committed to start AEOI under the CRS by 2017 or 2018, and an additional seven countries committed to doing so in 2019 or 2020.
  - CRS scope includes custodial institutions, depository institutions, investment entities, specified insurance companies; requires looking through passive entities to identify controlling persons.
  - CRS allows certain thresholds (e.g., option to exclude preexisting entity accounts below US$ 250,000).

### Skepticism and limitations around AEOI effectiveness
- Concerns and empirical limitations:
  - EUSD narrow scope could be circumvented via trusts, foundations, entity use, or asset relocation.
  - FATCA may have generated less revenue than anticipated; public evidence limited:
    - Since 2009 Offshore Voluntary Disclosure Program: more than 56,000 taxpayers paid a total of $11.1 billion in back taxes, interest and penalties.
    - TIGTA July 5, 2018 Final Report concludes IRS still not prepared to enforce compliance with FATCA.
  - CRS operational gaps: thresholds, customer self-certification, beneficial-ownership concealment techniques, and “citizenship/residence by investments” schemes.
- Need for systematic analysis motivates empirical strategy.

### Empirical strategy and estimation details
- Baseline regression framework:
  - Prior specification: log(foreign-owned deposits) = EOI_indicator * beta + time fixed effects + pair fixed effects + error.
  - Triple-differenced estimator: estimates effects separately for offshore and non-offshore jurisdictions and uses the difference as the estimator of EOI’s impact on tax evasion.
  - Generalized regression estimated on deposits held in all banks by savers located in non-offshore jurisdictions (subsets A and C).
  - Include four indicator variables for specific EOI agreements (EOI4, EUSD, FATCA, CRS) and an indicator AEOI equal to one if any automatic exchange framework exists.
  - Include two sets of time-specific fixed effects (one for offshore banks and one for non-offshore banks) and pair-specific fixed effects.
  - Robust standard errors (Arellano, 1987) clustered at the bank-saver level and allowing for arbitrary heteroscedasticity.
- Data partitioning logic:
  - BIS locational data partitioned into four subsets by bank (offshore/non-offshore) and saver (offshore/non-offshore).
  - Use of subsets A and C together controls for factors unrelated to tax evasion (e.g., DTT-related FDI effects).
  - Triple-differenced coefficient on interaction with offshore bank indicator captures relative offshore effect.

### Data description and key statistics
- Primary source: BIS restricted locational banking statistics, bilateral deposit stocks held by non-resident non-banks.
- Sample coverage:
  - Quarterly data for 39 reporting countries.
  - 19 reporting countries classified as offshore jurisdictions according to JZ2014.
  - 19 reporting countries have data as far back as Q1 1995 for a subset.
  - Latest quarter of available data at time of writing is Q2 2018.
- Aggregate patterns:
  - On average, aggregate cross-border deposits represent around 10 percent of GDP in any given quarter.
  - Starting in 2003, aggregate foreign-owned deposits increased and reached 15 percent of GDP in 2008.
  - Cross-border deposits-to-GDP ratio of offshore centers historically peaked at around 60 percent.

### Descriptive statistics (selected values from Table 2)
- Sample size: 311,778 observations for most variables.
- Non‐bank (nb) deposits: Observations 311,778; Min 0.001; Mean 1,066.31; Max 749,655
- Log of nb deposits: Observations 311,778; Min -6.908; Mean 2.970; Max 13.52737
- EOI4: Observations 311,778; Min 0; Mean 0.114; Max 1
- EUSD: Observations 311,778; Min 0; Mean 0.087; Max 1
- FAT: Observations 311,778; Min 0; Mean 0.001; Max 1
- CRS: Observations 311,778; Min 0; Mean 0.0374; Max 1
- AEOI: Observations 311,778; Min 0; Mean 0.109; Max 1
- Log of Bank GDP: Observations 311,778; Min 0.574; Mean 24.315; Max 30.661
- Log of Saver GDP: Observations 311,778; Min 1.009; Mean 24.539; Max 30.601
- Bank Voice: Observations 309,782; Min -1.387; Mean 1.110; Max 1.801
- Saver Voice: Observations 309,782; Min -2.313; Mean 0.124; Max 1.801
- Bank FX: Observations 311,778; Min 0.376; Mean 240093.5; Max 6973905

### A. Effect on deposits in offshore jurisdictions (baseline results)
- AEOI agreements:
  - Deposits in offshore jurisdictions decreased by between 35 percent (for the CRS) to 50 percent (for the EUSD) following AEOI introduction in some specifications.
  - Combined AEOI*Offshore: -0.535*** in one specification; combined AEOI measured effect around -53 percent and significant at the 1 percent level (column 6).
- EOIR:
  - Reduced deposits in offshore jurisdictions by around 12 percent in some estimates, though often not statistically significant.
- Selected coefficient examples (Table 3):
  - EOI4*Offshore coefficients include -0.431***, -0.127, -0.129, -0.114 (standard errors in table).
  - EUSD*Offshore: -0.514***.
  - CRS*Offshore: -0.351***.
  - FAT*Offshore: -0.457 (not consistently significant).
- Preferred specification (bank-specific time trends + two sets of time FE):
  - Measured AEOI effect reduced to 25 percent.
  - 95 percent confidence interval implies bank deposits held in offshore jurisdictions by non-offshore residents declined by between 13 and 37 percent.

### B. Robustness
- Sample restricted to 2002–2018: loses ~18 percent of baseline observations.
- Robust findings:
  - EOIR exerts a small and statistically non-significant impact on offshore deposits across robustness checks.
  - AEOI frameworks retain a strong negative impact; when constrained to uniform effects, measured AEOI effect ~-46 percent in restricted sample.
  - Including Log(Saver Gdp), Log(Bank Gdp), Log(FX rate), governance indicators: measured effect remains robust.

### C. Effect on countering round-tripping (subset B; Table 5)
- Sample: deposits held in non-offshore jurisdictions by savers from offshore jurisdictions.
- Main findings:
  - Presence of an EOI relationship reduces deposits held in non-offshore locations by residents of offshore centers (evidence most significant for EOIR).
  - EOI4*Offshore coefficients (Table 5): -0.206 *; -0.202 *; -0.261***; -0.265***; -0.277***.
  - NON_FATCA*Offshore (EUSD and CRS combined): -0.190 * (Model 2) and similar negatives across specifications.
  - FATCA*Offshore (US): positive coefficients around 0.22–0.247 with some specifications significant at 10 percent level; interpreted as deposits in the US increasing by roughly 20–23 percent following FATCA agreements with offshore jurisdictions.
- Interpretation:
  - EOIR, EUSD, and CRS reduce round-tripping into non-offshore banks by offshore residents (estimates between -14 and -20 percent when considered separately).
  - FATCA associated with increased deposits into the US by offshore savers.

### D. Finite mixture model estimation results (Table 6)
- AEOI coefficients (selected): -0.254***, -0.249***, -0.249***, -0.163**, -0.330***, -0.240*** (standard errors in table).
- EOI4 coefficients vary (examples): 0.019, 0.021, -0.057, 0.008, -0.013, 0.08.
- Predictors of posterior probability of being an offshore center (selected):
  - log(Gdp): positive and significant in several specifications (e.g., 0.282***, 0.272***, 0.879***, 1.412***).
  - log(Population): negative and significant in some specifications (e.g., -0.323***, -0.320***).
  - Secrecy: significant in some specs (e.g., 4.660**).
  - CIT rate: negative predictor in some specs (e.g., -0.014 *, -0.146***, -0.253***).
  - English (official language): positive and significant (e.g., 0.790***, 0.765***).
  - UN member: negative and significant (e.g., -0.516***, -0.500***).
  - PIT rate: negative and significant in some specifications (e.g., -0.021***, -0.022***).
  - FDI: positive and significant (e.g., 0.084***, 0.082***).
  - Trade: negative predictor in some specifications (e.g., -0.024***, -0.080***).
- Key finite-mixture findings:
  - Cross-border deposits decrease by an excess of between 16 and 33 percent in offshore centers (as identified by the EM algorithm) following AEOI introduction; effects statistically significant at the 1 percent level.
  - Personal income tax rates often better predictors of offshore status than corporate tax rates; increasing PIT by 10 percentage points reduces odds of being an offshore center by almost 20 percent in discussed specification.
  - Using country-specific fixed effects (time-constant posterior probabilities), final specification identifies 20 countries as offshore centers; measured AEOI effects remain comparable.

### Conclusions and policy-relevant findings
- Broad conclusions:
  - EOIR does not show a consistent reduction in offshore jurisdiction deposits; effects are variable and often not statistically significant.
  - Strong evidence that AEOI (EUSD, CRS, FATCA combined) reduced deposits in offshore jurisdictions; average estimated reduction around 25 percent in summary statements, with preferred specifications indicating reductions of 13–37 percent (95 percent CI) or an excess reduction of 16–33 percent in EM-identified offshore centers.
- Comparative effectiveness:
  - EUSD and CRS: appear most effective with strong statistical relationships and reductions in offshore deposits.
  - FATCA: measured impact often not statistically significant on offshore deposits and associated with increased deposits into the US from offshore savers (roughly 20–23 percent increase).
- Caveats:
  - Deposit reductions in one offshore jurisdiction may imply relocation to other jurisdictions; global tax-evasion implications are ambiguous.
  - BIS locational data cover cross-border bank deposits only and exclude other assets (bonds, securities, real estate).
  - Classification of offshore jurisdictions materially affects measured impacts; results sensitive to lists and predictor choices.
- Policy implication:
  - AEOI frameworks, particularly multilateral and automatic exchange arrangements like EUSD and CRS, are associated with substantial declines in bank deposits in offshore jurisdictions, consistent with reduced use of those jurisdictions to conceal income; policymakers should consider complementarities, potential relocation effects, and design features (e.g., reciprocity) that influence where concealed assets migrate.

*Source: wpiea2019286-print-pdf*

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

### References

### Tables
- 1. Subsets of Locational Data and Measured Effects ........................................................................................ 13
- 2. Descriptive Statistics ................................................................................................................................................. 18
- 3. Baseline Results .......................................................................................................................................................... 19
- 4. Robustness Checks .................................................................................................................................................... 21
- 5. Round-Tripping Results........................................................................................................................................... 23
- 6. Finite Mixture Model Results ................................................................................................................................ 25

### Figures
- 1. EOIR under DTTs and TIEAs ...................................................................................................................................... 8
- 2. Evolution of Cross-Border Bank Deposits (in % of group-specific GDP) ............................................. 17

### Other sections
- Annex

*Source: wpiea2019286-print-pdf - References*

### 1. Finite Mixture Model Estimations ....................................................................................

### 1. Finite Mixture Model Estimations

### Introduction and key empirical findings
- Dataset and scope:
  - Bank of International Settlements (BIS) restricted locational banking statistics recording foreign-owned deposits in 39 countries held by residents of over 200 counterparty jurisdictions on a quarterly basis since 1995.
  - Sample of over 230,000 observations.
- Main estimated effects of exchange of information (EOI) frameworks on deposits in offshore jurisdictions:
  - Bilateral agreements related to EOI upon request (EOIR) are associated with deposits held in offshore jurisdictions decreasing by between 8 and 12 percent following conclusion of an EOIR agreement (relationship has weak statistical significance).
  - Automatic exchange of information (AEOI) agreements are estimated to reduce deposits in offshore jurisdictions by around 25 percent (statistically significant).
- Round-tripping and the US FATCA experience:
  - No evidence that the US automatic exchange initiative, FATCA, has reduced round-tripping to the US.
  - Deposits from offshore centers into the US increased following the introduction of FATCA Intergovernmental Agreements (IGAs); possible contributing factors noted include the removal of a withholding tax threat, a lack of information exchange reciprocity by the US, and the non-imposition of tax on certain US sourced portfolio interest received by non-residents.

### Contribution to literature and methodological advances
- Extensions relative to prior work:
  - Analyze the effects of all existing bilateral EOI agreements over the last 20 years rather than a subset, including over 3,000 EOIR agreements, 85 FATCA agreements, the European Union Savings Directive and its related arrangements, and the CRS.
- Identification improvement:
  - Employ a triple-differenced estimator to better identify causal impact of EOI agreements and avoid omitted variable bias inherent in prior DD-only approaches that focused narrowly on offshore jurisdictions.
  - Triple-differenced approach estimates effects on offshore and non-offshore jurisdictions separately and uses the difference of these effects as an estimator of EOI’s impact on tax evasion.
- Sensitivity to offshore classification:
  - Recognize subjectivity in classifying countries as offshore; test sensitivity to choice of offshore list using a finite mixture model to estimate offshore-status and EOI impact simultaneously.

### Finite mixture model and classification results
- Methodological purpose:
  - Use a finite mixture model to estimate offshore-status of over 200 jurisdictions and the impact of EOI agreements simultaneously.
  - Iterative algorithm: obtain probabilistic classification of countries into offshore centers and then re-estimate the response to information exchange.
- Main classification findings:
  - Offshore centers are typically (though not exclusively) characterized by:
    - low income tax rates (both personal and corporate),
    - high financial secrecy,
    - English as an official language,
    - large FDI stocks,
    - small trade flows (both relative to GDP).
  - Contrary to Dharmapala and Hines (2009), governance indicators are not found to be statistically significant predictors of offshore status in this analysis.
  - UN membership is a significant predictor of a country’s offshore status in the finite mixture model results.
- Consistency with baseline regressions:
  - Finite mixture model estimates are consistent with baseline regressions: AEOI, but not EOIR, impacts cross-border deposits in offshore jurisdictions.

### Background on EOI frameworks (chronology and features)
- Traditional EOIR:
  - EOIR required a requesting jurisdiction to present a sufficiently detailed request identifying a specific taxpayer and demonstrating foreseeably relevance to domestic tax administration; historically relied on bilateral DTTs or TIEAs.
  - OECD amended Model Tax Convention in 2005 by adding paragraph 4 to Article 26 to address the absence of obligation to provide requested information when the requested jurisdiction had no domestic interest in it.
- EU Savings Directive (EUSD):
  - Agreement reached June 2003; came into effect on 1 July 2005.
  - Required banks and “paying agents” to disclose information on interest income earned by non-resident individuals to domestic tax authorities for automatic exchange.
  - EUSD later repealed and superseded in part by Council Directive 2014/107/EU implementing broader AEOI consistent with the CRS as of 1 January 2016.
- FATCA:
  - Enacted in 2010 as part of the HIRE Act; FATCA targets non-compliance by US taxpayers using undeclared foreign accounts.
  - FATCA went into effect on July 1, 2014.
  - To date in the paper: 113 jurisdictions have reached agreement with the US to undertake AEOI under FATCA IGAs (of which 88 IGAs are in force, 13 are signed but not yet in force, and 12 are agreed in substance).
  - FATCA projected revenue and costs cited:
    - Projected to generate $8.7 billion in revenue between fiscal years 2010-2020 (Joint Committee on Tax), a yearly average of $792 million.
    - IRS has spent approximately $380 million to implement and administer the FATCA program.
  - FATCA IGAs: three types described (Model 1A/1B, Model 2) and distinctions in reciprocity and reporting obligations noted.
- Common Reporting Standard (CRS) and multilateral AEOI:
  - CRS final version released July 2014 by OECD; designed for multilateral implementation with full reciprocity between participating jurisdictions.
  - As of November 2018, 101 jurisdictions committed to start AEOI under the CRS framework by 2017 or 2018, and an additional seven countries committed to doing so in 2019 or 2020.
  - CRS scope: custodial institutions, depository institutions, investment entities, specified insurance companies; requires looking through passive entities to identify controlling persons (beneficial owners).
  - CRS allows certain thresholds and due diligence procedures (e.g., option to exclude preexisting entity accounts below US$ 250,000) that could present opportunities for circumvention if exploited.

### Skepticism and limitations around AEOI effectiveness
- Concerns raised about EUSD, FATCA, and CRS effectiveness include:
  - EUSD narrow scope (initial focus on individuals) could be circumvented via trusts, foundations, or other entities; relocation of assets outside covered jurisdictions; relocation to investments outside EUSD scope.
  - FATCA may have generated less revenue than initially anticipated; public evidence limited:
    - Since 2009 Offshore Voluntary Disclosure Program: more than 56,000 taxpayers paid a total of $11.1 billion in back taxes, interest and penalties (bulk believed from anti-money laundering penalties related to FBAR non-filing rather than tax non-compliance).
    - TIGTA July 5, 2018 Final Report concludes IRS still not prepared to enforce compliance with FATCA.
  - CRS operational gaps:
    - Jurisdictions may allow thresholds (e.g., preexisting entity accounts below US$ 250,000) and allow customer self-certification for new individual accounts.
    - Techniques to conceal beneficial ownership include diluting ownership below the 25 percent threshold and misuse of “citizenship/residence by investments” schemes to claim residence in jurisdictions that enable circumvention.
- Many critiques are based on anecdotal evidence and a priori analysis of due diligence and reporting obligations; paper turns to systematic empirical analysis to assess impact.

### Empirical strategy and estimation details
- Baseline regression framework (difference-in-differences and extensions):
  - Prior empirical specification estimated variants of:
    - log(foreign-owned deposits) = EOI_indicator * beta + time fixed effects + pair fixed effects + error.
  - Issues with prior DD approaches:
    - Focusing only on offshore jurisdictions as both treatment and control can confound positive effects from factors such as double tax treaties (which may attract FDI and increase deposits) with negative effects from increased detection risk.
  - Triple-differenced estimator implemented:
    - Estimate effects separately for offshore and non-offshore jurisdictions and use the difference of these effects as the estimator of EOI’s impact on tax evasion.
    - Generalized regression (equation (2) in the paper) estimated on the sample of deposits held in all banks (offshore and non-offshore) by savers located in non-offshore jurisdictions (subsets A and C).
    - Include four indicator variables for the existence of specific EOI agreements (EOI4, EUSD, FATCA, and CRS) and an indicator AEOI which takes the value of one if any automatic exchange of information framework exists.
    - Include two sets of time-specific fixed effects (one for offshore banks and one for non-offshore banks) and pair-specific fixed effects.
    - Use robust standard errors (Arellano, 1987) clustering at the bank-saver level and allowing for arbitrary heteroscedasticity.
- Data subdivision and identification logic:
  - BIS locational data can be partitioned into four subsets by bank (offshore / non-offshore) and saver (offshore / non-offshore).
  - Previous studies used various subsets:
    - JZ2014 used subset C (offshore banks; savers non-offshore).
    - Menkhoff and Miethe (2017) used subset B (non-offshore banks; savers offshore) to study round-tripping.
    - CSS2018 combined subsets A and C and included an interaction with an offshore-bank indicator.
  - Authors argue using subsets A and C together allows control for factors unrelated to tax evasion but correlated with EOI (e.g., increased legal certainty from DTTs attracting FDI and raising deposits).
  - Triple-differenced estimator (coefficient on interaction with offshore bank indicator) captures the relative effect of EOI on offshore banks that exceeds the effect on non-offshore banks.

*Source: wpiea2019286-print-pdf - 1. Finite Mixture Model Estimations*

### Section V.D then tests the sensitivity of our findings to the choice of offshore list.

### Section V.D then tests the sensitivity of our findings to the choice of offshore list.

### Identification strategy and hypotheses
- Effective EOI agreements are expected to reduce deposits in offshore jurisdictions where income is hidden; the coefficient on treatment is expected to take a negative value for effective EOI agreements.
- Two estimation logics:
  - Difference-in-differences variant with residence (saver) country time-specific fixed effects to capture differences in pre-treatment trends between offshore and non-offshore jurisdictions.
  - A re-estimation of equation (2) on deposits held by all savers in non-offshore banks (subsets A and B), where the saver indicator takes the value of one if saver j is an offshore jurisdiction; the coefficient is again expected to be negative for effective EOI agreements.

### Importance and challenge of classifying offshore banks
- The classification of countries into offshore centers and non-offshore centers is critical for measuring EOI’s impact.
- Misclassification concerns:
  - An excessively broad definition of offshore centers will result in a downward bias of EOI’s measured impact.
  - A too narrow a definition may result in either a downward or upward bias (for instance, upward bias arises where deposits are relocated to another offshore jurisdiction in the control group that was mistakenly classified as not being offshore).

### Stochastic, data-driven classification: finite mixture model
- Approach rationale:
  - Country-specific estimates on the sensitivity of cross-border deposits to EOI provide information on a country’s likely offshore status.
  - Implemented using a finite mixture model, jointly estimating the impact of EOI and the probabilistic classification of countries in an iterative EM procedure.
- Model specification (preserved notation and structure):
  - Functional form for cross-border deposits (log scale) presented in equation (3), allowing different responses ߚ
௖௕,௖௦
 depending on whether the bank and saver are classified as offshore or not, and allowing time fixed effects to differ by classification.
  - Probability that country i is an offshore center, ݌
௜
, modeled as a logistic function of predictor vector ݖ
௜௧
 (equation (4)): log(݌
௜௧
/(1݌െ
௜௧
)) = ᇱ ݖ
௜௧
.
  - Predictors include both time-varying and time-constant variables: WB governance indicators, measures of a country’s size (GDP, population), English language use, UN membership, trade-openness (sum of imports and exports as a share of GDP), FDI stocks (as a share of GDP), tax rates (corporate and personal), a financial secrecy index (available for one year), and dummy variables indicating inclusion in an existing list of offshore centers.
  - Joint density of bilateral log deposit stocks is a mixture of four normal distributions corresponding to the four possible bank/saver offshore/non-offshore combinations (equation (5)). Parameters ߠ and probability vector ݌ are obtained by maximum likelihood estimation using an EM-algorithm.

### EM-algorithm iterative procedure (intuitive three steps)
- Step 1: Estimate equation (3) with weighted least squares for each of the four subsets; weights reflect prior probability of group membership. To test whether the response to AEOI differs, estimate these four equations in a SUR framework.
- Step 2: Compute posterior probabilities that a country is an offshore center, based on last round’s parameter estimates; this step identifies outliers by contrasting observed reactions with prototypical reactions. Posterior probabilities are restricted to be constant across time.
- Step 3: Use computed posterior probabilities to estimate equation (4), refining measured posterior probabilities by explaining them with control variables related to the likelihood of being an offshore center. Predicted probabilities are then used as priors in the next iteration.
- Convergence: Iteration ends when posteriors do not change from one round to the next.
- Alternative specification: In one specification the third step is effectively skipped by including a vector of country-specific fixed effects in estimating equation (4), making predicted and computed posteriors equivalent.

### Data description and key statistics
- Primary data source: BIS’ restricted locational banking statistics, analyzing bilateral deposit stocks held by non-resident non-banks.
- Sample coverage:
  - Quarterly data for 39 reporting countries.
  - 19 of these are classified as offshore jurisdictions according to JZ2014.
  - 19 reporting countries have data as far back as Q1 1995 for a subset.
  - The latest quarter of available data at the time of writing is Q2 2018.
- Aggregate patterns:
  - On average, aggregate cross-border deposits represent around 10 percent of GDP in any given quarter.
  - Starting in 2003, aggregate foreign-owned deposits increased and reached 15 percent of GDP in 2008.
  - Cross-border deposits-to-GDP ratio of offshore centers historically at much higher levels; offshore centers reached a peak at around 60 percent.

*Source: wpiea2019286-print-pdf - Section V.D then tests the sensitivity of our findings to the choice of offshore list.*

### 2008.  In comparison, the foreign deposits-to-GDP ratio in non-offshore jurisdictions has been

### wpiea2019286-print-pdf - 2008.  In comparison, the foreign deposits-to-GDP ratio in non-offshore jurisdictions has been

### Data sources and indicators
- BIS locational statistics: cross-border bank deposits (used to investigate AEOI initiatives' effectiveness).
- EOIR dataset: extended JZ2014 dataset using OECD’s Exchange of Tax Information Portal; covers over 6,500 bilateral relationships and records date when EOI upon request became effective, legal basis (DTT or TIEA), and presence of paragraph 4 of Article 26 of the OECD Model Tax Convention.
- Savings Directive (EUSD): collected from European Commission; covers over 1,300 bilateral relationships.
- FATCA: agreements downloaded from US Treasury; at time of writing, 88 FATCA relationships are in effect.
- CRS: activated bilateral relationships from OECD’s Automatic Exchange Portal; based on CRS, more than 4,600 country pairs exchange information currently.
- Control variables: logs of GDP and foreign exchange rates (IMF WEO), World Bank Governance indicators.

### Descriptive statistics (Table 2)
- Sample size: 311,778 observations for most variables.
- Non‐bank (nb) deposits: Observations 311,778; Min 0.001; Mean 1,066.31; Max 749,655
- Log of nb deposits: Observations 311,778; Min -6.908; Mean 2.970; Max 13.52737
- EOI4: Observations 311,778; Min 0; Mean 0.114; Max 1
- EUSD: Observations 311,778; Min 0; Mean 0.087; Max 1
- FAT: Observations 311,778; Min 0; Mean 0.001; Max 1
- CRS: Observations 311,778; Min 0; Mean 0.0374; Max 1
- AEOI: Observations 311,778; Min 0; Mean 0.109; Max 1
- Log of Bank GDP: Observations 311,778; Min 0.574; Mean 24.315; Max 30.661
- Log of Saver GDP: Observations 311,778; Min 1.009; Mean 24.539; Max 30.601
- Bank Voice: Observations 309,782; Min -1.387; Mean 1.110; Max 1.801
- Saver Voice: Observations 309,782; Min -2.313; Mean 0.124; Max 1.801
- Bank FX: Observations 311,778; Min 0.376; Mean 240093.5; Max 6973905

### Empirical approach
- Panel estimations with pair-specific fixed effects and time-specific fixed effects; from specification 3 onward a flexible time-fixed effects structure allowing different trends for offshore and non-offshore banks.
- Triple-differenced estimator used to identify treatment effect of EOI agreements on tax evasion.
- Subsample analyses: subset C (offshore banks & non-offshore savers), subset B (non-offshore banks & savers from offshore jurisdictions).
- Finite mixture (EM) model: simultaneous estimation of probability of being an offshore jurisdiction and sensitivity of cross-border deposits to EOI; some specifications allow posterior probabilities to vary annually.

### A. Effect on deposits in offshore jurisdictions (baseline results, Table 3)
- AEOI agreements: deposits in offshore jurisdictions decreased by between 35 percent (for the CRS) to 50 percent (for the EUSD) following AEOI introduction.
- EOIR: reduced deposits in offshore jurisdictions by around 12 percent, though this effect is not statistically significant in most specifications.
- Specific estimates from Table 3 (selected coefficients, with significance where indicated):
  - EOI4*Offshore interactions (various specs): coefficients include -0.431***, -0.127, -0.129, -0.114 (standard errors shown in table).
  - EUSD*Offshore: -0.514*** (standard error in table).
  - CRS*Offshore: -0.351*** (standard error in table).
  - FAT*Offshore: -0.457 (not consistently significant).
  - Combined AEOI*Offshore: -0.535*** in one specification; combined AEOI measured effect around -53 percent and significant at the 1 percent level (column 6).
- Interpretation:
  - AEOI frameworks show strong and statistically significant reductions in cross-border deposits to offshore jurisdictions.
  - EOIR evidence is mixed: some specifications suggest negative effects (around -12 percent) but often lack statistical significance.

### B. Robustness (Table 4)
- Sample restricted to 2002–2018: loses ~18 percent of baseline observations.
- Robust findings:
  - EOIR exerts a small and statistically non-significant impact on offshore deposits across robustness checks.
  - AEOI frameworks retain a strong negative impact; when constrained to uniform effects, measured AEOI effect ~-46 percent in restricted sample.
- Preferred specification (bank-specific time trends + two sets of time FE):
  - Measured AEOI effect reduced to 25 percent.
  - 95 percent confidence interval implies bank deposits held in offshore jurisdictions by non-offshore residents declined by between 13 and 37 percent.
  - Including Log(Saver Gdp), Log(Bank Gdp), Log(FX rate), governance indicators: measured effect remains robust (Table 4, Models 3–5).
- Noted comparisons:
  - JZ2014 reported EOIR reduces deposits by ~12 percent; current analysis confirms magnitude but with greater variability making effect potentially null.
  - CSS2018 found CRS impact of 14 percent; other studies show varying results depending on offshore classification and timeframe.

### C. Effect on countering round-tripping (Table 5)
- Sample: deposits held in non-offshore jurisdictions by savers from offshore jurisdictions (subset B).
- Main findings:
  - Presence of an EOI relationship reduces deposits held in non-offshore locations by residents of offshore centers (evidence most statistically significant for EOIR).
  - EOI4*Offshore coefficients (Table 5): -0.206 *; -0.202 *; -0.261***; -0.265***; -0.277*** (standard errors in table).
  - NON_FATCA*Offshore (EUSD and CRS combined): -0.190 * (Model 2) and similar negative coefficients across specifications (standard errors in table).
  - FATCA*Offshore (US): positive coefficients around 0.22–0.247 with some specifications significant at 10 percent level; interpreted as deposits in the US increasing by roughly 20–23 percent following FATCA agreements with offshore jurisdictions.
- Interpretation:
  - EOIR, EUSD, and CRS reduce round-tripping into non-offshore banks by offshore residents (estimates between -14 and -20 percent when considered separately).
  - FATCA is an outlier: associated with increased deposits into the US by offshore savers, plausibly due to FATCA’s non-reciprocal features and US tax/withholding rules.

### D. Finite mixture model (Table 6)
- Parameter estimation:
  - EOI4 coefficients across columns: examples include 0.019, 0.021, -0.057, 0.008, -0.013, 0.08 (standard errors in table).
  - AEOI coefficients: -0.254***, -0.249***, -0.249***, -0.163**, -0.330***, -0.240*** (standard errors in table).
- Posterior probability (probability estimation) predictors (selected):
  - log(Gdp): positive and significant in several specifications (e.g., 0.282***, 0.272***, 0.879***, 1.412*** with standard errors in table).
  - log(Population): negative and significant in some specifications (e.g., -0.323***, -0.320***).
  - Secrecy: coefficients vary; become significant in some specifications (e.g., 4.660** with standard error in table).
  - CIT rate (in percent): negative predictor in some specs (e.g., -0.014 *), stronger negative in others (e.g., -0.146***, -0.253***).
  - English (official language): positive and significant (e.g., 0.790***, 0.765***; larger magnitudes in time-constant specs).
  - UN member: negative and significant (e.g., -0.516***, -0.500***; larger negative in time-constant specs).
  - PIT rate (in percent): negative and significant in some specifications (e.g., -0.021***, -0.022***).
  - FDI: positive and significant (e.g., 0.084***, 0.082***; larger in time-constant specs).
  - Trade: negative predictor in some specifications (e.g., -0.024***, -0.080***).
- Key finite-mixture results:
  - Cross-border deposits decrease by an excess of between 16 and 33 percent in offshore centers (as identified by the EM algorithm) following AEOI introduction; effects statistically significant at the 1 percent level.
  - Non-sovereign, English-speaking countries with low tax rates, large FDI stocks, small trade flows, and high financial secrecy are more likely to be classified as offshore centers by the EM algorithm.
  - Personal income tax rates often better predictors of offshore status than corporate tax rates; increasing PIT by 10 percentage points reduces odds of being an offshore center by almost 20 percent in discussed specification.
  - Using country-specific fixed effects (time-constant posterior probabilities), final specification identifies 20 countries as offshore centers; measured AEOI effects remain comparable despite narrower classification.

### Conclusions and policy-relevant findings
- Broad conclusion:
  - No consistent reduction in offshore jurisdiction deposits following EOIR introduction; EOIR effects are variable and often not statistically significant.
  - Strong evidence that AEOI (EUSD, CRS, FATCA combined) reduced deposits in offshore jurisdictions; average estimated reduction around 25 percent after any AEOI agreement in baseline/summary statements, with preferred specifications indicating reductions of 13–37 percent (95 percent CI) or an excess reduction of 16–33 percent in EM-identified offshore centers.
- Comparative effectiveness:
  - EUSD and CRS: appear most effective with strong statistical relationships and reductions in deposits in offshore jurisdictions.
  - FATCA: measured impact often not statistically significant on offshore deposits and, uniquely, associated with increased deposits into the US from offshore savers (roughly 20–23 percent increase), suggesting potential relocation rather than deterrence.
- Caveats and interpretation:
  - A reduction in deposits in one offshore jurisdiction could imply relocation of deposits to other jurisdictions; global tax evasion implications are ambiguous.
  - Not all cross-border deposits are for tax-evasion purposes; dataset limited to cross-border bank deposits and excludes other assets (bonds, securities, real estate).
  - Classification of offshore jurisdictions materially affects measured impacts; results sensitive to lists and predictor choices.
- Policy implication:
  - AEOI frameworks (particularly multilateral and automatic exchange arrangements like EUSD and CRS) are associated with substantial declines in bank deposits in offshore jurisdictions, consistent with reduced use of those jurisdictions to conceal income; policymakers should consider complementarities, potential relocation effects, and the design features (e.g., reciprocity) that influence where concealed assets migrate.

*Italic source: wpiea2019286-print-pdf - 2008.  In comparison, the foreign deposits-to-GDP ratio in non-offshore jurisdictions has been*

### Annex 1. Finite Mixture Model Estimations

### Annex 1. Finite Mixture Model Estimations

### Estimation objective and first-order condition
- The parameter vector ߠ
௖௕,௖௦ maximizing the log-likelihood must satisfy the first-order condition (equation (6)):
  - ߱෍
௜௝௧
௖௕,௖௦	
௜௝௧
߲logቀ݂
൫
ݕ
௜௝௧
ห
ߠ
௖௕,௖௦
,ݔ
௜௝௧
൯
ቁ
ߠ߲
௖௕,௖௦
ൌ0
- The posterior probability that bank i is classified as cb and saver j as cs is defined in (7):
  - ߱
௜௝௧
௖௕,௖௦	
ൌ
ݍ
௜,௝
௖௕,௖௦	
݂
൫
ݕ
௜௝௧
ห
ߠ
௖௕,௖௦
,ݔ
௜௝௧
൯
݄
൫
ݕ
௜௝௧
ห
ߠ,ݔ
௜௝௧
൯

### EM-style iterative estimation procedure (three steps)
- Overview:
  - The algorithm estimates the probabilities ݌ = (݌
ଵ , ... ݌
ே ) and the parameters of the four density distributions via an iterative 3-step procedure.
  - Iteration continues until posterior probabilities do not change from one round to the next.
  - All estimations control for pair-specific fixed effects and take the JZ list to assign prior probabilities of country classification.

- Step 1 — Estimation of ߠ:
  - Conditional on a vector of posterior probabilities, an explicit solution to equation (5) is given by a weighted least squares regression of ݕ
௜௝௧ on ݔ
௜௝௧ where the vector of posterior probabilities is used as a weight.
  - Initialization:
    - The JZ2014 list is used to assign country-specific posteriors.
    - The posterior is set to 0.85 for countries classified as offshore centers and to 0.15 for countries not classified as offshore centers.
    - Footnote: In the initial estimation round, countries included in the JZ list are given a posterior probability of 75 percent of being an offshore jurisdiction.

- Step 2 — Updating of country-specific probabilities ݌:
  - Each country is assigned a unique probability of being an offshore center.
  - Conditional on parameter estimates and last round’s prior ݌
௜, the likelihood of country i being an offshore center is proportional to (equation (8)):
    - ߥ
௜
ை௙௙
݌ൌ
௜
ෑෑ
ൣ
݂
൫
ݕ
௜௝௧
ห
ߠ
ை௙௙,௢௡
,ݔ
௜௝௧
൯
൅
݂
൫
ݕ
௜௝௧
ห
ߠ
ை௙௙,ை௙௙
,ݔ
௜௝௧
൯൧
௧௝
  - The likelihood of country i not being an offshore center is proportional to (equation (9)):
    - ߥ
௜
௢௡
ൌሺ1݌െ
௜
ሻෑা়
ൣ
݂
൫
ݕ
௜௝௧
ห
ߠ
௢௡,௢௡
,ݔ
௜௝௧
൯
൅
݂
൫
ݕ
௜௝௧
ห
ߠ
௢௡,௢௙௙
,ݔ
௜௝௧
൯൧
௧௝
  - Posterior probability update (equation (10)):
    - ݌
௜
௧ାଵ
ൌ
ߥ
௜
௢௙௙
ߥ
௜
௢௙௙
ߥ൅
௜
௢௡

- Step 3 — Predictor step:
  - Estimate expected value of country-specific probabilities conditional on explanatory variables z using regression (equation (11)):
    - ݌
௜
௧ାଵ
ߛሺ݇ൌ
ᇱ
ݖ
௜
ሻ
    - k is a logistic function.
  - Combine the vector of country-specific posterior probabilities ݌ to obtain joint probabilities ݍ
௖௕,௖௦ of observing any country-pair combination, which are then used as priors in Step 1.

### Sample construction and classification details
- To increase classification accuracy, savers from the restricted dataset with fewer than 1400 observations are dropped.
- This results in a sample of 85 countries classified with the EM algorithm.
- All estimations control for pair-specific fixed effects and use the JZ list to assign prior probabilities of country classification.

*Source: wpiea2019286-print-pdf - Annex 1. Finite Mixture Model Estimations*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019286-print-pdf.pdf_