## 3.1    Methodology

## Source details

**Canonical URL:** [3.1    Methodology](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021153-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021153-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021153-print-pdf.pdf.json)

---

### Overview and research contribution
- Paper examines what happens to a country’s capital flows after it is identified in the Financial Action Task Force (FATF)’s list of countries or jurisdictions with strategic deficiencies in their framework for Anti-Money Laundering/Combating the Financing of Terrorism (AML/CFT) (the gray list).
- Contributions claimed:
  - First study to look at the impact of gray-listing on all components of capital flows: FDI, portfolio flows, and banking and other flows.
  - Uses more recent data on gray-listing and includes a larger sample of gray-listed countries.
  - First study (to the authors’ knowledge) to study the problem using machine learning (lasso).

### Data and sample
- Sample: 89 emerging and developing countries (EMDCs) in 2000–2017.
- Period: 2000q1–2017q4.
- Gray-listing coverage: During this period, 78 countries have been gray-listed at least once (not all remain in estimation sample due to missing values).
- Capital flows data: IMF Financial Flows Analytics (FFA) database — quarterly data on FDI, portfolio flows (debt and equity), and other investments (private, official, bank, and non-bank) available on a gross basis (inflows and outflows) and a net basis; used as share of (quarterly) GDP for analysis.
- Gray-listing dates: FATF public statements (issued three times a year: February, June, and October) and FATF annual reports.
- Sample construction notes:
  - List of EMDCs taken from the World Economic Outlook (WEO) in 2000, with additions: Afghanistan, Brunei Darussalam, Iraq, Korea, Kosovo, Liberia, Montenegro, Serbia, and Timor-Leste.
  - Six advanced economies included: Greece, Hong Kong, Iceland, Ireland, Israel, and Singapore.
  - China excluded because of its unique characteristics including its size.
  - Quarterly GDP for 8 more gray-listed countries supplemented using IFS and Haver and estimated quarterly GDP based on quarterly industrial production data via the Chow-Lin method (Chow and Lin, 1971).
  - Extreme values in capital flows excluded before running models; detection based on fixed-effect (within) transformation and generalized box plot adjusting for asymmetry and skewness (Bruffaerts et al., 2014).

### Event-history (event study) methodology
- Event history analysis framework estimated via fixed-effect panel regression:
  - y_{i,t} = α_i + γ λ_t + Σ_s β_s δ_{s,t} + ε_{i,t}
  - α_i: country fixed effect measuring average capital flow for country i outside the GFC and the gray-listing event.
  - λ_t: dummy = 1 during the global financial crisis (GFC); γ measures average deviation during the GFC.
  - δ_{s,t}: dummy = 1 when a country is s periods away from gray-listing; β_s captures evolution within an s-quarter window surrounding gray-listing.
- Event window specification:
  - s set to nine quarters (four quarters before, four quarters after) in baseline.
  - Alternative half-window k allowed to vary between 1 and 3 quarters when defining event window [t0 − k, t0 + k]; literature examples cited where k varies between 1 and 3 years or 1 to 4 quarters.

### Econometric identification and adjustments
- Goal: isolate and estimate the treatment effect of gray-listing while controlling for confounders.
- Exclusion of extreme values to reduce influence of outliers (see detection method above).
- Acknowledged uncertainty in timing of FATF announcements due to case-by-case evaluation and extensions; some capital flow types may respond faster to announcements than others.

### Machine learning approach
- Uses lasso and the double-selection lasso extension for causal inference in panel data to address model selection with many potential confounders and to improve estimation efficiency.
- Rationale:
  - Lasso performs variable selection via penalized regression (penalty controlled by λ).
  - Plain lasso may exclude variables needed for unbiased causal inference; double-selection lasso (Belloni, Chernozhukov, and Hansen, 2014a) addresses this by selecting controls from both outcome and treatment equations and then estimating treatment effect by OLS on the union of selected controls.
  - Panel-data extension by Belloni et al. (2014b) accommodates within-group correlation and heteroskedasticity.

### Double-selection lasso procedure and formalization
- Three-step double-selection lasso for inference on β_0:
  1. Lasso of y on x's (without d); selected controls = A.
  2. Lasso of d on x's; selected controls = B.
  3. OLS of y on d and w where w = A ∪ B.
- Panel “within model” notation:
  - ÿ_{i,t} = ẍ′_{i,t} β + ε̈_{i,t}, where ẍ_{i,t} = x_{i,t} − (1/T) ∑_{t=1}^T x_{i,t}.
  - Lasso estimator: β̂_lasso(λ) = arg min_β (1/(nT)) ∑_{i=1}^n ∑_{t=1}^T (ÿ_{i,t} − ẍ′_{i,t} β)^2 + (λ/(nT)) ∑_{j=1}^p ψ_j |β_j|.
  - Penalty loadings ψ_j account for predictor-specific variances and clustered errors when relevant.

### Data used in empirical implementation (controls and preprocessing)
- Dependent variables (all as a share of GDP): Gross capital inflow; FDI inflow; portfolio inflow; other inflows.
- Preprocessing:
  - Exclude extreme values before within-model estimation.
  - Drop post-gray-listing observations for four quarters to avoid post-event dynamics biasing estimation of β_0.
- Treatment variable:
  - d_{i,t} (= 1 for gray-listing), defined over a window [t_0 − k, t_0 + k], baseline k = 3; robustness checks include k = 1, 2, 4.
- Controls:
  - 15 domestic primary variables (with four-quarter lags): six exchange rate variables (exchange rate vis-a-vis the SDR, nominal effective exchange rate, real effective exchange rate — each for period average and end of period); two official reserves variables (in U.S. dollars and as a share of GDP); two current account balance variables (in U.S. dollars and as a share of GDP); one-year ahead forecast of real GDP growth rate; measure of capital account openness; measure of exchange rate flexibility; credit ratings.
  - Missing values in primary domestic variables reduce sample to 81 countries, of which 28 have been gray-listed.
  - 22 international variables: VIX, the U.S. short-term rate, the U.S. 10-year rate, the U.S. Federal Funds rate; two commodity price indexes (one including and one excluding gold); real GDP growth rate of G4 countries (total and individually); money supply of G4 countries as a share of GDP (total and individually); growth rate of the money supply of G4 countries (total and individually); dummy for the GFC.
  - Several permutations (levels and differences), interactions with exchange rate flexibility, credit ratings, and GFC dummy, and a full set of year fixed effects.
  - Total number of variables: 2,842.

### Key empirical findings (double-selection lasso + OLS estimates)
- Effect of Gray-Listing on Capital Inflows (in percent of GDP)
  - Total coefficient: −7.550 ∗∗∗
    - Standard Errors: (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - FDI coefficient: −3.034 ∗∗∗
    - Standard Errors: (1.016)
    - Conf. Intervals: [-5.02, -1.04]
  - Portfolio coefficient: −2.926 ∗∗∗
    - Standard Errors: (0.981)
    - Conf. Intervals: [-4.85, -1.00]
  - Other coefficient: −3.551 ∗∗∗
    - Standard Errors: (0.685)
    - Conf. Intervals: [-4.89, -2.21]
- Decomposition of “other” inflows (in percent of GDP)
  - Official: Coefficient −0.925 ∗∗
    - Standard Errors: (0.393)
    - Conf. Intervals: [-1.70, 0.15]
  - Non-Official: Coefficient −3.796 ∗∗∗
    - Standard Errors: (0.755)
    - Conf. Intervals: [-5.28, -2.32]
  - Banks: Coefficient −1.964 ∗∗∗
    - Standard Errors: (0.667)
    - Conf. Intervals: [-3.27, -0.66]
  - Other Private: Coefficient −2.372 ∗∗∗
    - Standard Errors: (0.443)
    - Conf. Intervals: [-3.24, -1.50]
- Interpretation:
  - Point estimate for total capital inflow effect is −7.6 percent of GDP (statistically significant at the 1 percent level using robust standard errors). The 95 percent confidence interval is between −10.6 percent of GDP and −4.6 percent of GDP.
  - FDI inflow effect: −3.0 percent of GDP (significant at the 1 percent level).
  - Portfolio flow effect: −2.9 percent of GDP (significant at the 1 percent level).
  - Largest estimated impact within “other” inflows is for non-official sector flows: −3.8 percent of GDP.

### Event-study (fixed-effect panel) descriptive estimates (summary)
- Gross capital inflow is on average over 6 percentage points of GDP lower compared to historical norms at time of gray-listing.
- Component-specific average declines at time of gray-listing (event-study estimates):
  - FDI inflow declines by 3.2 percent of GDP.
  - Portfolio inflow declines by 3.3 percent of GDP.
  - Other inflow declines by 3.1 percent of GDP.
- Gross outflow declines on average by around 3 percentage points at the time of gray-listing, partially offsetting the drop in gross inflows.
- Observed anticipatory behavior: surge in capital outflow a few quarters ahead of the announcement.
- Other-investment components — average inflows decline by 2 to 3 percent of GDP:
  - Private sector inflows: −3.3 percent of GDP.
  - Official sector inflows: −0.5 percent of GDP.
  - Bank flows: −2.1 percent of GDP.
  - Nonbank flows: −2.4 percent of GDP.
- Net errors and omissions show a slight increase at time of gray-listing (see Appendix C), insufficient to fully explain declines in capital inflow.

### Robustness: sensitivity to event window
- Baseline event window: k = 3 (window [t_0 − k, t_0 + k]).
- Effect of Gray-Listing using Different Event Windows (coefficient on total capital inflow, percent of GDP)
  - k = 4: Coefficient −8.203 ∗∗∗
    - (Standard Errors) (1.524)
    - Conf. Intervals: [-11.19, -5.22]
  - k = 3: Coefficient −7.550 ∗∗∗
    - (Standard Errors) (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - k = 2: Coefficient −5.789 ∗∗∗
    - (Standard Errors) (1.649)
    - Conf. Intervals: [-9.02, -2.56]
  - k = 1: Coefficient −4.836 ∗∗∗
    - (Standard Errors) (1.683)
    - Conf. Intervals: [-8.14, -1.54]
- Interpretation:
  - Absolute size larger for wider event windows (k = 4).
  - Coefficients remain negative and significant for narrower windows, though smaller in absolute magnitude.

### Small number of gray-listing observations (Section 4.2) — imbalance and resampling
- Gray-listing observations constitute about 0.7 percent of the sample.
- Resampling methods applied to increase gray-listing share to about 2 percent:
  - SMOTE (Synthetic Minority Oversampling Technique).
  - Random undersampling (RU).
  - Combination of SMOTE and RU.
- Estimation results (effect on total capital inflow, percent of GDP)
  - Baseline
    - Coefficient: −7.550 ***
    - Standard Errors: (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - SMOTE
    - Coefficient: −6.316 ***
    - Standard Errors: (1.320)
    - Conf. Intervals: [-8.90, -3.73]
  - RU
    - Coefficient: −7.245 ***
    - Standard Errors: (1.535)
    - Conf. Intervals: [-10.25, -4.24]
  - SMOTE & RU
    - Coefficient: −7.504 ***
    - Standard Errors: (1.599)
    - Conf. Intervals: [-10.64, -4.37]
- Findings:
  - Estimated coefficients remain negative and similar in size across resampling methods.
  - SMOTE reduces standard errors and narrows 95 percent confidence intervals (noted as “ranging from −9 to −4 percent of GDP” under SMOTE); RU and combined approaches produce estimates similar to baseline.

### Interpretation, mechanisms, and policy relevance
- Possible mechanisms for decline in capital flows:
  - De-risking: banks exit relationships with customers based in high-risk countries to reduce compliance costs.
  - Market enforcement: investors use gray-listing as a heuristic to evaluate risk and reallocate resources away from affected countries.
- Anticipatory outflows explained by information asymmetry (domestic investors acting ahead of foreign investors).
- V-shaped recovery interpretation:
  - Initial decline may represent overshooting followed by volatility as foreign investors update beliefs (theoretical support cited).
- Policy relevance:
  - Robust negative effect estimates inform policymakers about potential economic costs of gray-listing and the incentive to address AML/CFT gaps.
  - Results relevant for investors’ expectations and the Fund’s surveillance, capacity development, and program forecasting (e.g., impacts on external reserves, Balance of Payments vulnerability, and capacity to repay).

_Italic: Source: wpiea2021153-print-pdf — 3.1 Methodology (excerpts) and 4.2 Small Number of Gray-Listing Observations._

### 3.1    Methodology   .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .    11

### 3.1    Methodology

### Overview and research contribution
- Paper examines what happens to a country’s capital flows after it is identified in the Financial Action Task Force (FATF)’s list of countries or jurisdictions with strategic deficiencies in their framework for Anti-Money Laundering/Combating the Financing of Terrorism (AML/CFT) (the gray list).
- Contributions claimed:
  - First study to look at the impact of gray-listing on all components of capital flows: FDI, portfolio flows, and banking and other flows.
  - Uses more recent data on gray-listing and includes a larger sample of gray-listed countries.
  - First study (to the authors’ knowledge) to study the problem using machine learning (lasso).

### Data and sample
- Sample: 89 emerging and developing countries (EMDCs) in 2000–2017.
- Period: 2000q1–2017q4.
- Gray-listing coverage: During this period, 78 countries have been gray-listed at least once (not all remain in estimation sample due to missing values).
- Capital flows data: IMF Financial Flows Analytics (FFA) database — quarterly data on FDI, portfolio flows (debt and equity), and other investments (private, official, bank, and non-bank) available on a gross basis (inflows and outflows) and a net basis; used as share of (quarterly) GDP for analysis.
- Gray-listing dates: FATF public statements (issued three times a year: February, June, and October) and FATF annual reports.
- Sample construction notes:
  - List of EMDCs taken from the World Economic Outlook (WEO) in 2000, with additions: Afghanistan, Brunei Darussalam, Iraq, Korea, Kosovo, Liberia, Montenegro, Serbia, and Timor-Leste.
  - Six advanced economies included to maintain sample size: Greece, Hong Kong, Iceland, Ireland, Israel, and Singapore.
  - China excluded because of its unique characteristics including its size.
  - Unavailability of quarterly GDP constrains sample size; quarterly GDP for 8 more gray-listed countries supplemented using IFS and Haver and estimated quarterly GDP based on quarterly industrial production data via the Chow-Lin method (Chow and Lin, 1971).
  - Extreme values in capital flows excluded before running models; detection based on fixed-effect (within) transformation and generalized box plot adjusting for asymmetry and skewness (Bruffaerts et al., 2014).

### Event-history (event study) methodology
- Event history analysis framework used to explore evolution of capital flows around gray-listing.
- Estimated fixed-effect panel regression (equation (1)):
  - y_{i,t} = α_i + γ λ_t + Σ_s β_s δ_{s,t} + ε_{i,t}
  - y_{i,t} is capital flows for country i in time t.
  - α_i is a country fixed effect measuring average capital flow for country i outside the GFC and the gray-listing event (“tranquil times”).
  - λ_t is a dummy that equals 1 in each quarter during the global financial crisis (GFC) and 0 otherwise; γ measures average deviation during the GFC.
  - δ_{s,t} is a dummy that equals 1 when a country is s periods away from gray-listing in time t; β_s captures evolution within an s-quarter window surrounding gray-listing.
  - ε_{i,t} is the residual error.
- Event window specification:
  - s set to nine quarters (four quarters before, four quarters after).
  - Alternative approach allows k (half-window) to vary between 1 and 3 quarters when defining event window [t0 − k, t0 + k]; literature examples cited where k varies between 1 and 3 years or 1 to 4 quarters.

### Econometric identification and adjustments
- Purpose of regression analysis: isolate and estimate the treatment effect of gray-listing while controlling for confounding factors.
- Exclusion of extreme values to reduce influence of outliers (see detection method above).
- Timing of FATF announcements not automatic; gray-listing timing arises from case-by-case evaluation and extensions — event timing uncertainty explicitly acknowledged; some capital flow types may respond faster to announcements than others.

### Machine learning approach
- For empirical execution, the paper uses a machine learning technique known as lasso to assist in estimation (introduced in Section 3).

### Key empirical findings (summary statements from the paper)
- Main lasso-regression result reported in the introduction:
  - Capital inflows decline on average by 7.6 percent of GDP when the country is gray-listed.
  - FDI inflows decline on average by −3.0 percent of GDP.
  - Portfolio inflows decline on average by −2.9 percent of GDP.
  - Other investment inflows decline on average by −3.6 percent of GDP.
  - The estimated impacts are all statistically significant.
- Event-study (fixed-effect panel regression) descriptive estimates reported in Section 2 (figure captions and text):
  - Gross capital inflow is on average over 6 percentage points of GDP lower compared to historical norms at time of gray-listing.
  - Component-specific average declines at time of gray-listing (event-study estimates):
    - FDI inflow declines by 3.2 percent of GDP.
    - Portfolio inflow declines by 3.3 percent of GDP.
    - Other inflow declines by 3.1 percent of GDP.
  - Gross outflow declines on average by around 3 percentage points at the time of gray-listing, partially offsetting the drop in gross inflows.
  - Observed anticipatory behavior: surge in capital outflow a few quarters ahead of the announcement (consistent with information asymmetry where domestic investors have more information than foreign investors).
  - Other-investment components (interbank loans, suppliers’ credits, trade credits, and other items) — average inflows decline by 2 to 3 percent of GDP:
    - Private sector inflows: −3.3 percent of GDP.
    - Official sector inflows: −0.5 percent of GDP.
    - Bank flows: −2.1 percent of GDP.
    - Nonbank flows: −2.4 percent of GDP.
  - Net errors and omissions show a slight increase at time of gray-listing (see Appendix C), but not enough to fully explain declines in capital inflow — suggesting some diversion to shadow flows but much of the decline is a genuine reduction in inflows.

### Interpretation and mechanisms discussed
- Possible mechanisms for decline in capital flows:
  - De-risking: banks exit relationships with customers based in high-risk countries to reduce compliance costs.
  - Market enforcement: investors use gray-listing as a heuristic to evaluate risk and reallocate resources away from affected countries.
- V-shaped recovery interpretation:
  - Initial decline may represent overshooting followed by volatility as foreign investors update beliefs; theoretical support cited (Bacchetta and van Wincoop, 2000).
- Anticipatory outflows explained by information asymmetry literature (domestic investors acting ahead of foreign investors).

*Source: wpiea2021153-print-pdf — 3.1 Methodology (excerpts from the paper).*

### 3.1  Methodology

### 3.1  Methodology

### Overview of empirical strategy
- Estimate a simple model of capital flow:
  - y_{i,t} = α d_{i,t} + x′_{i,t} β + μ_i + ε_{i,t},  i = 1,...,n; t = 1,...,T
  - d_{i,t} is the causal variable of interest or treatment effect (gray-listing indicator); x_{i,t} is the set of controls; μ_i is a country fixed effect; ε_{i,t} is the error.
- Two broad categories of confounders are considered:
  - Global factors: VIX, changes in money supply of major advanced economies, growth rates of major advanced economies, the Federal Funds Rate.
  - Domestic factors: level of per capita income, real GDP growth, trade openness, financial openness, exchange rate regime, exchange rates, foreign reserves, current account balances, bank credit, sovereign credit ratings.
- Challenges addressed:
  - Model selection when many potential confounders exist.
  - Estimation inefficiency and multicollinearity if too many variables are included.
  - High dimensionality and possible nonlinearities and interactions that make exhaustive permutations infeasible.
- Chosen approach: lasso and the double-selection lasso extension for causal inference in panel data.

### Rationale for lasso and double-selection lasso
- Lasso properties:
  - Minimizes sum of squared deviations plus a penalty on the sum of absolute values of coefficients to force many coefficients to zero (performs model selection).
  - Tuning parameter λ controls penalty level; larger λ shrinks more coefficients to zero.
- Limitation of plain lasso:
  - Generally does not produce unbiased estimates of coefficients or valid standard errors for causal inference because it selects covariates based on predictive ability, possibly excluding variables in the true DGP and causing omitted variable bias.
- Double-selection lasso (Belloni, Chernozhukov, and Hansen, 2014a):
  - Modified estimation procedure that generates unbiased coefficients and standard errors for a subset of variables (suitable for causal interpretation of the coefficient on the variable of interest).
  - Panel data extension by Belloni et al. (2014b) accommodates within-group correlation and heteroskedasticity.

### Double-selection lasso procedure (for inference on β_0)
- Consider y_i = β_0 d_i + β_1 x_{i,1} + ... + β_p x_{i,p} + ε_i; goal: valid inference on β_0.
- Three steps:
  1. Use lasso to estimate y_i on x's without d_i. Denote the set of lasso-selected controls by A.
  2. Use lasso to estimate d_i on x's. Denote the set of lasso-selected controls by B.
  3. Estimate by OLS: y_i = β_0 d_i + w′_i β + ε_i, where w_i = A ∪ B (the union of selected controls).
- Panel-data “within model”:
  - ÿ_{i,t} = ẍ′_{i,t} β + ε̈_{i,t}, where ẍ_{i,t} = x_{i,t} − (1/T) ∑_{t=1}^T x_{i,t}.
  - Lasso estimator solves: β̂_lasso(λ) = arg min_β (1/(nT)) ∑_{i=1}^n ∑_{t=1}^T (ÿ_{i,t} − ẍ′_{i,t} β)^2 + (λ/(nT)) ∑_{j=1}^p ψ_j |β_j|.
  - Penalty loadings ψ_j account for predictor-specific variances; under homoskedasticity ψ_j = sqrt((1/n) ∑_{i=1}^n x_{i,j}^2).
  - For clustered errors, ψ_j = sqrt((1/(nT)) ∑_{i} u_{ij}^2) with u_{ij} = ∑_{t} x_{ijt} ε_{it} (super-observation aggregation).

### Model formalization
- Lasso minimization (cross-sectional notation):
  - Minimize (1/n) ∑_{i=1}^n (y_i − x′_i β)^2 + λ ∑_{j=1}^p |β_j|.
- Penalized estimator with predictor-specific loadings:
  - β̂_lasso(λ) = arg min_β (1/n) ∑_{i=1}^n (y_i − x′_i β)^2 + (λ/n) ∑_{j=1}^p ψ_j |β_j|.

### Data used in empirical implementation
- Dependent variables (all as a share of GDP):
  - Gross capital inflow; FDI inflow; portfolio inflow; other inflows. (Data on capital flows from the FFA database.)
- Quarterly GDP supplemented from IFS and Haver where possible.
- Preprocessing:
  - Exclude extreme values before running the within-model estimation.
  - Drop post-gray-listing observations for four quarters to avoid post-event dynamics biasing estimation of β_0.
- Treatment variable:
  - Indicator variable d_{i,t} (= 1 for gray-listing), defined over a window [t_0 − k, t_0 + k], where t_0 is the date of the FATF announcement and k varied between 1 and 3 quarters (baseline k = 3; robustness checks include k = 1, 2, 4).
- Controls:
  - 15 domestic primary variables (included with four-quarter lags):
    - Six exchange rate variables: exchange rate vis-a-vis the SDR, nominal effective exchange rate, real effective exchange rate (each for period average and end of period).
    - Two official reserves variables (in U.S. dollars and as a share of GDP).
    - Two current account balance variables (in U.S. dollars and as a share of GDP).
    - One-year ahead forecast of real GDP growth rate.
    - A measure of capital account openness.
    - A measure of exchange rate flexibility.
    - Credit ratings.
  - Missing values in primary domestic variables reduce the sample size to 81 countries, of which 28 have been gray-listed.
  - 22 international variables:
    - VIX, the U.S. short-term rate, the U.S. 10-year rate, the U.S. Federal Funds rate.
    - Two commodity price indexes (one including and one excluding gold).
    - Real GDP growth rate of G4 countries (the U.S., the Euro Area, the U.K., and Japan) in total and individually.
    - Money supply of G4 countries as a share of GDP in total and individually.
    - Growth rate of the money supply of G4 countries in total and individually.
    - A dummy variable for the GFC.
  - Several permutations of each variable used (levels and differences), interaction effects of each variable with three categorical variables (exchange rate flexibility, credit ratings, and GFC dummy), and a full set of year fixed effects.
  - Total number of variables: 2,842.

### Results: effect of gray-listing on capital inflows (double-selection lasso + OLS)
- Table 1: Effect of Gray-Listing on Capital Inflows (in percent of GDP)
  - Total coefficient: −7.550 ∗∗∗
    - Standard Errors: (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - FDI coefficient: −3.034 ∗∗∗
    - Standard Errors: (1.016)
    - Conf. Intervals: [-5.02, -1.04]
  - Portfolio coefficient: −2.926 ∗∗∗
    - Standard Errors: (0.981)
    - Conf. Intervals: [-4.85, -1.00]
  - Other coefficient: −3.551 ∗∗∗
    - Standard Errors: (0.685)
    - Conf. Intervals: [-4.89, -2.21]
- Interpretation and statistical significance:
  - Point estimate for the effect of gray-listing on total capital inflow is −7.6 percent of GDP (statistically significant at the 1 percent level using robust standard errors). The 95 percent confidence interval is between −10.6 percent of GDP and −4.6 percent of GDP.
  - Effect on FDI inflow: −3.0 percent of GDP (significant at the 1 percent level).
  - Effect on portfolio flow: −2.9 percent of GDP (significant at the 1 percent level).
  - Relative magnitudes for FDI and portfolio inflows are similar.
- Decomposition of “other” inflows (Table 2; in percent of GDP)
  - Official: Coefficient −0.925 ∗∗
    - Standard Errors: (0.393)
    - Conf. Intervals: [-1.70, 0.15]
  - Non-Official: Coefficient −3.796 ∗∗∗
    - Standard Errors: (0.755)
    - Conf. Intervals: [-5.28, -2.32]
  - Banks: Coefficient −1.964 ∗∗∗
    - Standard Errors: (0.667)
    - Conf. Intervals: [-3.27, -0.66]
  - Other Private: Coefficient −2.372 ∗∗∗
    - Standard Errors: (0.443)
    - Conf. Intervals: [-3.24, -1.50]
  - Findings:
    - Largest estimated impact for non-official sector flows: −3.8 percent of GDP.
    - Banking sector flows: −2.0 percent of GDP.
    - Other (non-bank) private flows: −2.4 percent of GDP.
    - Official sector flows: −0.9 percent of GDP.
    - All estimates significant at the 5 percent level or better.
- Magnitude context:
  - The impact of −7.6 percent of GDP is described as large but not unprecedented: corresponds to about the 25th percentile of the distribution of quarterly declines in gross capital inflow for EMDCs during the GFC, and about the 45th percentile for countries affected by the Asian Crisis in 1997–98.

### Robustness: sensitivity to event window
- Baseline event window: k = 3 (window [t_0 − k, t_0 + k]).
- Robustness checks for k = 1, 2, and 4 produce consistently negative and statistically significant effects on total capital inflows.
- Table 3: Effect of Gray-Listing using Different Event Windows (coefficient on total capital inflow, percent of GDP)
  - k = 4: Coefficient −8.203 ∗∗∗
    - (Standard Errors) (1.524)
    - Conf. Intervals: [-11.19, -5.22]
  - k = 3: Coefficient −7.550 ∗∗∗
    - (Standard Errors) (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - k = 2: Coefficient −5.789 ∗∗∗
    - (Standard Errors) (1.649)
    - Conf. Intervals: [-9.02, -2.56]
  - k = 1: Coefficient −4.836 ∗∗∗
    - (Standard Errors) (1.683)
    - Conf. Intervals: [-8.14, -1.54]
- Interpretation:
  - Absolute size of the estimated effect is larger for wider event windows (k = 4).
  - Coefficients remain negative and significant for narrower windows (k = 2 and k = 1), though smaller in absolute magnitude.
  - Possible interpretations: wider windows may capture additional confounding influences or longer realization of market responses; narrower windows may miss adjustment that occurs earlier or later relative to the announcement.

_Italic: Source: wpiea2021153-print-pdf - 3.1 Methodology (excerpts)._

### 4.2  Small Number of Gray-Listing Observations

### 4.2  Small Number of Gray-Listing Observations

### Data imbalance and problem statement
- Gray-listing observations constitute about 0.7 percent of the sample.
- The small share of gray-listing observations reduces the model’s accuracy for estimating the impact of gray-listing on capital flows.
- To assess robustness, the authors re-estimate the model using different resampling techniques to increase the share of gray-listing observations up to about 2 percent of the data.

### Resampling methods used
- Oversampling using SMOTE (Synthetic Minority Oversampling Technique):
  - SMOTE creates synthetic copies similar to, but not identical to, existing minority-class observations.
  - SMOTE is applied here for resampling the minority class based on the predictor (independent) variable (i.e., gray-listing).
- Random undersampling (RU):
  - Randomly selects and deletes majority-class observations until gray-listing observations reach 2 percent of the data.
- Combination of SMOTE and random undersampling:
  - Combined approach often yields better out-of-sample performance according to machine learning literature.

### Estimation results (Table 4)
- The estimated coefficients of gray-listing remain negative and similar in size to the baseline across resampling methods.
- Exact estimates reported in Table 4 (effect of gray-listing on total capital inflow, in percent of GDP):
  - Baseline
    - Coefficient: −7.550 ***
    - Standard Errors: (1.522)
    - Conf. Intervals: [-10.53, -4.57]
  - SMOTE
    - Coefficient: −6.316 ***
    - Standard Errors: (1.320)
    - Conf. Intervals: [-8.90, -3.73]
  - RU
    - Coefficient: −7.245 ***
    - Standard Errors: (1.535)
    - Conf. Intervals: [-10.25, -4.24]
  - SMOTE & RU
    - Coefficient: −7.504 ***
    - Standard Errors: (1.599)
    - Conf. Intervals: [-10.64, -4.37]
- Where the number of minority-class observations is substantially increased (i.e., under SMOTE), standard errors become smaller as expected.
- The 95 percent confidence intervals are smaller under SMOTE, “ranging from−9 to−4 percent of GDP,” but remain similar to the baseline in other resampling approaches (−11 to−4 percent of GDP) as noted in the text.

### Robustness and interpretation
- Negative effect of gray-listing on capital inflows is robust to alternative resampling methods designed to mitigate class imbalance.
- The magnitude of the negative effect is large in the paper’s broader findings: on average −7.6 percent of GDP (reported elsewhere in the chapter).
- Resampling via SMOTE narrows standard errors and confidence intervals relative to the baseline, improving precision when synthetic minority observations are added.
- Random undersampling and the combined approach do not uniformly reduce standard errors relative to the baseline.

### Policy relevance and implications (as connected to robustness)
- Robust negative effect estimates inform policymakers about potential economic costs of gray-listing and the incentive to address AML/CFT gaps.
- Results are relevant for investors’ expectations of market responses to gray-listing and for the Fund’s surveillance, capacity development, and program forecasting (e.g., impacts on external reserves, Balance of Payments vulnerability, and capacity to repay).

*Source: wpiea2021153-print-pdf - 4.2  Small Number of Gray-Listing Observations*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021153-print-pdf.pdf_