## 3. Estimates of Parameters of Equation (5), Various Scenarios

## Source details

**Canonical URL:** [3. Estimates of Parameters of Equation (5), Various Scenarios](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06182.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06182.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06182.pdf.json)

---

### I. Introduction — objective and model
- Objective: estimating fiscal reaction functions from panel data where the primary balance (p) responds to lagged debt (d) and other determinants.
- Generalized specification (equation (5)):
  - p it = ρ d i,t-1 + Σ k γ k y i,k,t + Σ j β j x i,j,t + α i + ε i,t
  - Unobserved fixed effect variance: 2 0 η σ≥
  - Disturbance variance: 2 ,0 ε σ>
- Key empirical questions:
  - (i) Does fiscal policy satisfy the intertemporal budget constraint?
  - (ii) Is fiscal policy countercyclical?
- Methodological issue: LSDV bias arises because d i,t-1 is correlated with past idiosyncratic shocks ε i,t-s (s>1), violating strict exogeneity in short panels (small T).

### II. Analytical derivation of biases (OLS and LSDV)
- Model assumptions and moments follow equation (5) setup.
- Case summaries and bias directions:
  - Case 1 (γ = 0, β = 0, 2 0 b σ =):
    - OLS: negatively biased (downward) for ρ in presence of fixed effects.
    - LSDV: positively biased for ρ. Asymptotic LSDV bias (plim N→∞ ρ̂ − ρ) given in equation (13) is positive and decreasing in 2 η σ.
  - Case 2 (γ = 0, β = 0, 2 0 b σ > 0):
    - Bias expression (14): denominator gains term 2 b QR σ+, so larger 2 b σ tends to reduce LSDV bias (if Q+R>0).
  - Case 3 (exogenous regressors included, 2 0 b σ = 0):
    - Inclusion of exogenous regressors increases expected magnitude of LSDV bias because M < I reduces the denominator in bias expression.
  - Case 4 (endogenous regressors included, 2 0 b σ = 0):
    - Effect of endogenous regressors on direction/size of LSDV bias is ambiguous; with instrumentation, country dummies still expected to exert a positive bias on ρ.
- Overall analytical conclusion:
  - Country-specific dummy variables (LSDV) are expected to exert a positive bias on estimated ρ in short panels; OLS is downward-biased.
  - The true ρ is expected to lie between OLS (downward) and LSDV (upward) estimates.

### III. Monte Carlo experiments — design and signal-to-noise
- Purpose: quantify size/direction of OLS and LSDV biases in equation (5) and compare with AR(1) model (equation (3)).
- Data-generating and calibration choices:
  - x i,t = λ x i,t-1 + ξ i,t with ξ ~ N(0, 2 ξ σ) and λ = 0.5 (equation (15)).
  - Signal-to-noise ratios considered: 2 s σ = 2 and 8 (for comparison with Kiviet (1995)).
  - AR(1) experiments: ρ alternates between 0.1 and 0.4; β set to 1 − ρ; (1) η ε σ ρ σ = − in AR(1) setup.
  - Fiscal reaction function experiments: f i,t set to f (sometimes drawn from Uniform[0.95,1.10]); ρ values: ρ = 0, 0.03, 0.06, 0.09.
- Monte Carlo sample sizes and repetitions:
  - First set (comparison with AR(1)): N = 100, 1000 simulations, T = 5, 10, 20, 30.
  - Second set (emerging-market calibration): N = 30, T = 15, 1000 simulations.
- Signal-to-noise definitions:
  - AR(1) model: signal-to-noise ratio defined as in equation (16).
  - Fiscal reaction function: signal-to-noise ratio defined in equation (17), with θ = f ρ −.

### IV. Monte Carlo scenarios and parameter values (data-consistent calibration)
- Calibration inputs (emerging market panel 1990–2004):
  - λ = 0.5; 2 ξ σ = 3.
  - Estimated standard deviation of country fixed effects: 2.8 = 2 η σ.
  - Estimated standard deviation of disturbances: 2.5 = 2 ε σ.
  - β (output gap coefficient): calibrated to 0.35 when output gap instrumented, 0.20 when not instrumented.
  - f drawn from Uniform[0.95,1.10]; midpoint f = 1.025 used for example.
  - Debt-shock standard deviations: b σ ≈ 11 in full sample; excluding outliers yields b σ ≈ 4.
  - Empirical range for ρ in emerging markets: between 0.02 and 0.06.
- Three Monte Carlo scenarios (Table 2 parameterization preserved exactly):
  - Scenario I:
    - ρ = 0.0, 0.03, 0.06, 0.09
    - β = 0.35
    - ε σ = 2.5
    - η σ = 2.8
    - f ~ Uniform[0.95,1.1]
    - λ = 0.5
    - ξ σ = 3
    - b σ = 0
    - py σ = 0
  - Scenario II:
    - Same as Scenario I except b σ = 5
  - Scenario III:
    - Same as Scenario II except py σ = −0.01 and output gap generated with contemporaneous endogeneity per equation (18)
- Estimators applied:
  - OLS
  - LSDV (country dummies)
  - Arellano and Bond (1991) one-step difference GMM (GMM)
  - Blundell and Bond (1998) one-step system GMM (SGMM)
  - SGMMR: SGMM using only second and third lags of d i,t-1 as instruments
  - For Scenario III: LSDVIV = two-stage least squares with country fixed effects instrumenting y i,t with x i,t; additional variants reported in Table 3.

### V. Key Monte Carlo findings (summary of results and tables)
- Comparison with AR(1) model (first set of simulations; Table 1):
  - AR(1) model (equation (3)):
    - OLS is upward-biased for ρ.
    - LSDV is downward-biased for ρ (Nickell-type bias).
  - Fiscal reaction function model (equation (5)):
    - OLS leads to negative bias for ρ.
    - LSDV leads to positive bias for ρ (opposite signs relative to AR(1)).
  - Bias magnitudes:
    - LSDV biases decline with increasing T and with higher signal-to-noise ratio (2 s σ).
    - For T ≥ 10 in equation (5), LSDV bias typically smaller than 10 percent of true ρ and often affects only the second decimal.
    - LSDV bias more sensitive to signal-to-noise: bias for 2 s σ = 8 typically less than one tenth of bias for 2 s σ = 2.
- Emerging-market calibrated scenarios (Table 3; N = 30, T = 15, 1000 draws):
  - True ρ lies between average OLS and average LSDV (or LSDVIV in Scenario III) estimates across simulations.
  - Scenario I (b σ = 0, py σ = 0):
    - For ρ = 0.03:
      - OLS mean 0.010
      - LSDV mean 0.050
      - GMM mean 0.041
      - SGMM mean 0.011
      - SGMMR mean 0.012
    - For ρ = 0.06:
      - OLS mean 0.021
      - LSDV mean 0.138
      - GMM mean 0.147
      - SGMM mean 0.023
      - SGMMR mean 0.024
    - RMSE examples:
      - Scenario I, ρ = 0.03: LSDV [RMSE] = [0.023]; OLS [RMSE] = [0.020].
  - Scenario II (b σ = 5, py σ = 0):
    - Biases generally reduced compared with Scenario I.
    - For ρ = 0.03: OLS 0.011; LSDV 0.040; GMM 0.037; SGMM 0.012; SGMMR 0.013.
  - Scenario III (b σ = 5, py σ = −0.01, endogenous output gap — LSDVIV used):
    - For ρ = 0.03: OLS 0.011; LSDVIV 0.039; GMM 0.037; SGMM 0.011; SGMMR 0.012.
    - LSDVIV estimates of β (output-gap coefficient) can display very large S.E. and RMSE (e.g., LSDVIV reported S.E. (1.768) and [RMSE] [1.767] for some β estimates), signaling weak-instruments/overfitting when many instruments including country dummies are used.
- Estimator performance takeaways:
  - OLS and SGMM can be severely negatively biased for ρ in many settings; SGMM often produces large negative bias for ρ in some simulations.
  - LSDV and LSDVIV provide the lowest bias and lowest RMSE for ρ across simulated scenarios; GMM is a close runner-up.
  - When regressors (e.g., output gap) are endogenous, GMM and SGMM perform better for coefficients on those endogenous variables (β), yielding lower RMSEs than LSDV/LSDVIV for β.
  - All estimators produce estimates near zero when true ρ = 0, so false indication that ρ > 0 is unlikely in these simulations.
- Numerical design and reported calibration values preserved exactly:
  - λ = 0.5; ξ σ = 3; η σ = 2.8; ε σ = 2.5.
  - Scenario ρ grid: 0.0, 0.03, 0.06, 0.09.
  - Scenario II: b σ = 5.
  - Scenario III: py σ = −0.01.
  - Simulation dimensions: N = 30, T = 15 for emerging-market calibration; first set used N = 100, T = 5, 10, 20, 30; 1000 simulations per configuration.
- Example exact entries from Tables preserved:
  - AR(1), OLS: N = 100, T = 5, ρ = 0.1, 2 s σ = 2: ρ̂ = 0.4236 (Bias 0.3236), S.E. = 0.0402.
  - AR(1), LSDV: N = 100, T = 5, ρ = 0.1, 2 s σ = 2: ρ̂ = −0.0327 (Bias −0.1327), S.E. = 0.0339.
  - Scenario I, ρ = 0.03: OLS = 0.010 (S.E. (0.001)) [RMSE 0.020]; LSDV = 0.050 (S.E. (0.011)) [RMSE 0.023]; GMM = 0.041 (S.E. (0.016)) [RMSE 0.020]; SGMM = 0.011 (S.E. (0.001)) [RMSE 0.019]; SGMMR = 0.012 (S.E. (0.002)) [RMSE 0.018].
  - Scenario II, ρ = 0.06: OLS = 0.022 (S.E. (0.002)) [RMSE 0.038]; LSDV = 0.080 (S.E. (0.017)) [RMSE 0.026]; GMM = 0.092 (S.E. (0.036)) [RMSE 0.048].
  - Scenario III: LSDVIV can report S.E. (1.768) and [RMSE] [1.767] for some β estimates.

### VI. Practical implications and recommendations
- Estimator-choice guidance:
  - For estimating ρ (fiscal reaction to debt / solvency tests): prefer LSDV or LSDVIV methods — they produce precise estimates of ρ with modest biases in realistic short-panel settings (bias often < 10% of true ρ for T ≥ 10 under calibrated scenarios).
  - For testing fiscal policy countercyclicality (coefficients on potentially endogenous variables like output gap): prefer GMM methods (Arellano and Bond (1991) difference GMM and Blundell and Bond (1998) system GMM) using exogenous instruments for the output gap.
- Specific cautions:
  - Exogenous shocks to debt (b i,t with 2 b σ > 0) reduce LSDV bias on ρ.
  - Inclusion of exogenous regressors increases expected magnitude of LSDV bias on ρ.
  - Endogenous regressors have ambiguous analytic effect on LSDV bias; instrumenting such regressors is necessary, but instrumentation combined with country dummies can increase bias for ρ while reducing RMSE for coefficients on endogenous variables.
  - Use of many instruments (including country dummies) in first-stage IV (LSDVIV) can inflate RMSE and standard errors for coefficients like β (weak instruments problem).
- Bottom-line guidance:
  - Use LSDV (or LSDVIV when instrumenting y is required) for intertemporal solvency tests (coefficient ρ).
  - Use GMM (difference or system) when testing fiscal policy countercyclicality (coefficients on potentially endogenous variables like output gap).

*Source: _wp06182 - References..............................................................................................................*

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

### _wp06182 - References..............................................................................................................

### References
- References................................................................................................................................26

### Tables
- 1. Estimates of Equation (3).....................................................................................................14
- 2. Parameter Values Used in Simulations................................................................................17

*Source: _wp06182 - References..............................................................................................................*

### 3. Estimates of Parameters of Equation (5), Various Scenarios ..............................................19

### 3. Estimates of Parameters of Equation (5), Various Scenarios

### I. Introduction — objective and model
- Focus: estimating fiscal reaction functions from panel data where the primary balance (p) responds to lagged debt (d) and other determinants.
- Commonly estimated specification (generalized form shown as equation (5)):
  - Primary balance variable: p it
  - Lagged debt: d i,t-1
  - Endogenous regressors: y i,k,t with coefficients γ k
  - Exogenous regressors: x i,j,t with coefficients β j
  - Unobserved fixed effect: η i with variance 2 0 η σ≥
  - Disturbance: ε i,t with variance 2 ,0 ε σ> 
- Key empirical questions: (i) Does fiscal policy satisfy the intertemporal budget constraint? (ii) Is fiscal policy countercyclical?
- Methodological issue highlighted: LSDV (least-squares-with-dummy-variables) bias arises because lagged debt, d i,t-1, is correlated with past idiosyncratic shocks ε i,t-s (s>1), violating strict exogeneity in short panels (small T).

### II. Analytical derivation of biases (OLS and LSDV)
- General model considered (equation (5)):
  - p it = ρ d i,t-1 + Σ k γ k y i,k,t + Σ j β j x i,j,t + α i + ε i,t, with assumptions on moments and orthogonality as stated.
- Key analytical results (Case analysis):
  - Case 1 (γ = 0, β = 0, 2 0 b σ =): benchmark with only lagged debt and no independent shocks to debt.
    - OLS: analytically shown to be negatively biased (downward bias) for ρ in presence of fixed effects due to correlation between η i and d i,t-1.
    - LSDV: analytically shown to be positively biased for ρ. Expression for asymptotic LSDV bias (plim N→∞ ρ̂ − ρ) is given in equation (13) and is positive and decreasing in 2 η σ.
  - Case 2 (γ = 0, β = 0, 2 0 b σ > 0): adding exogenous shocks to debt (b i,t)
    - Bias expression (14): denominator gains the term 2 b QR σ+, so a larger 2 b σ tends to reduce the LSDV bias (if Q+R>0), because past primary-surplus shocks account for a smaller share of variance of d i,t-1.
  - Case 3 (exogenous regressors included, 2 0 b σ = 0):
    - Inclusion of exogenous regressors increases expected magnitude of LSDV bias because M < I reduces the denominator in bias expression.
  - Case 4 (endogenous regressors included, 2 0 b σ = 0):
    - Effect of endogenous regressors on direction/size of LSDV bias is ambiguous analytically; if instrumented, country dummies still expected to exert a positive bias on ρ.
- Overall analytical conclusion:
  - Use of country-specific dummy variables (LSDV) is expected to exert a positive bias on estimated ρ in short panels; OLS is downward-biased. The true ρ expected to lie between OLS (downward) and LSDV (upward) estimates.

### III. Monte Carlo experiments — design and signal-to-noise
- Purpose: quantify the size/direction of OLS and LSDV biases in equation (5) and compare with canonical AR(1) model (equation (3)).
- Data-generating and calibration choices:
  - Explanatory variable x i,t generated as: x i,t = λ x i,t-1 + ξ i,t with ξ ~ N(0, 2 ξ σ) and λ = 0.5 (equation (15)).
  - Signal-to-noise ratios considered for comparison with Kiviet (1995): 2  s σ = 2 and 8.
  - For AR(1) experiments: ρ alternates between 0.1 and 0.4; β set to 1 − ρ; (1) η ε σ ρ σ = − in AR(1) setup.
  - For fiscal reaction function experiments: f i,t set to f (sometimes drawn from Uniform[0.95,1.10]); ρ values considered in later scenarios: ρ = 0, 0.03, 0.06, 0.09.
- Monte Carlo sample sizes and repetitions:
  - First set (comparison with AR(1)): N = 100, 1000 simulations, T = 5, 10, 20, 30.
  - Second set (data-consistent calibration for emerging markets): N = 30, T = 15, 1000 simulations.
- Signal-to-noise definitions:
  - AR(1) model: signal-to-noise ratio defined as in equation (16).
  - Fiscal reaction function: signal-to-noise ratio defined in equation (17), with θ = f ρ −.

### IV. Monte Carlo scenarios and parameter values (data-consistent calibration)
- Parameter estimates and calibration inputs (emerging market panel 1990–2004 used for calibration):
  - λ = 0.5; 2 ξ σ = 3; λ for x process set to 0.5.
  - Estimated standard deviation of country fixed effects: 2.8 = 2 η σ
  - Estimated standard deviation of disturbances: 2.5 = 2 ε σ
  - β (output gap coefficient) calibrated to 0.35 when output gap instrumented, 0.20 when not instrumented.
  - Practical range for f drawn from Uniform[0.95,1.10]; midpoint f = 1.025 used for example.
  - Debt-shock standard deviations: b σ ≈ 11 in the full sample; excluding outliers yields b σ ≈ 4.
  - Empirical range for ρ in emerging markets: between 0.02 and 0.06 per cited evidence.
- Three Monte Carlo scenarios (Table 2 parameterization preserved exactly):
  - Scenario I:
    - ρ = 0.0, 0.03, 0.06, 0.09
    - β = 0.35
    - ε σ = 2.5
    - η σ = 2.8
    - f ~ Uniform[0.95,1.1]
    - λ = 0.5
    - ξ σ = 3
    - b σ = 0
    - py σ = 0
  - Scenario II:
    - same as Scenario I except b σ = 5 (normal shocks to debt)
  - Scenario III:
    - same as Scenario II except py σ = −0.01 and output gap generated with contemporaneous endogeneity per equation (18)
- Estimators applied in experiments:
  - OLS
  - LSDV (country dummies)
  - Arellano and Bond (1991) one-step difference GMM (GMM)
  - Blundell and Bond (1998) one-step system GMM (SGMM)
  - SGMMR: SGMM using only second and third lags of d i,t-1 as instruments
  - For Scenario III (endogenous output gap): LSDVIV = two-stage least squares with country fixed effects instrumenting y i,t with x i,t; also additional reported variants in Table 3.

### V. Key Monte Carlo findings (summary of results and tables)
- Comparison with AR(1) model (first set of simulations; Table 1):
  - In AR(1) model (equation (3)):
    - OLS estimator is upward-biased for ρ.
    - LSDV estimator is downward-biased for ρ (Nickell-type bias).
  - In fiscal reaction function model (equation (5)):
    - OLS leads to negative bias for ρ.
    - LSDV leads to positive bias for ρ (opposite signs relative to AR(1)).
  - Bias magnitudes:
    - LSDV biases decline with increasing T and with higher signal-to-noise ratio (2 s σ).
    - For T ≥ 10 in equation (5), LSDV bias typically smaller than 10 percent of true ρ and often affects only the second decimal.
    - LSDV bias in equation (5) is more sensitive to signal-to-noise: bias for 2 s σ = 8 typically less than one tenth of bias for 2 s σ = 2.
- Estimates under emerging-market calibrated scenarios (Table 3; N = 30, T = 15, 1000 draws):
  - The true ρ lies between average OLS and average LSDV (or LSDVIV in Scenario III) estimates across simulations.
  - Scenario I (b σ = 0, py σ = 0):
    - For ρ = 0.03: OLS mean 0.010; LSDV mean 0.050; GMM mean 0.041; SGMM mean 0.011; SGMMR mean 0.012.
    - For ρ = 0.06: OLS mean 0.021; LSDV mean 0.138; GMM mean 0.147; SGMM mean 0.023; SGMMR mean 0.024.
    - RMSEs (selected): LSDV [RMSE] for ρ = 0.03 is [0.023]; OLS [RMSE] is [0.020].
  - Scenario II (b σ = 5, py σ = 0):
    - Biases generally reduced compared with Scenario I (presence of b i,t reduces LSDV bias).
    - For ρ = 0.03: OLS 0.011; LSDV 0.040; GMM 0.037; SGMM 0.012; SGMMR 0.013.
  - Scenario III (b σ = 5, py σ = −0.01, endogenous output gap — LSDVIV used):
    - For ρ = 0.03: OLS 0.011; LSDVIV 0.039; GMM 0.037; SGMM 0.011; SGMMR 0.012.
    - LSDVIV estimates of β (output-gap coefficient) display very large standard errors and RMSEs in some specifications (diagnosed as driven by many instruments including country dummies in first stage).
  - Estimator performance takeaways from Table 3:
    - OLS and SGMM are severely negatively biased for ρ in many settings; SGMM often produces large negative bias for ρ in some settings of the simulations.
    - LSDV and LSDVIV provide the lowest bias and lowest RMSE for ρ across the simulated scenarios; GMM is a close runner-up.
    - When regressors (such as the output gap) are endogenous, GMM and SGMM perform better for coefficients on those endogenous variables (β), yielding lower RMSEs than LSDV/LSDVIV for β.
    - All estimators produce estimates near zero when true ρ = 0, so false indication that ρ > 0 is unlikely in these simulations.
- Numerical highlights preserved exactly (selected):
  - Monte Carlo design values and reported calibrated parameters:
    - λ = 0.5; ξ σ = 3; η σ = 2.8; ε σ = 2.5.
    - Scenario ρ grid: 0.0, 0.03, 0.06, 0.09.
    - Scenario II: b σ = 5.
    - Scenario III: py σ = −0.01.
    - Simulation dimensions: N = 30, T = 15 for emerging-market calibration; first set used N = 100, T = 5, 10, 20, 30; 1000 simulations per configuration.
  - From Table 1 (AR(1) and comparisons): example entries (preserved exactly) illustrate bias directions:
    - For AR(1), OLS with N = 100, T = 5, ρ = 0.1, 2 s σ = 2: ρ̂ = 0.4236 (Bias 0.3236), S.E. = 0.0402.
    - For AR(1), LSDV with N = 100, T = 5, ρ = 0.1, 2 s σ = 2: ρ̂ = −0.0327 (Bias −0.1327), S.E. = 0.0339.
  - From Table 3 (Scenario I examples, preserved exactly):
    - Scenario I, ρ = 0.03: OLS = 0.010 (S.E. (0.001)) [RMSE 0.020]; LSDV = 0.050 (S.E. (0.011)) [RMSE 0.023]; GMM = 0.041 (S.E. (0.016)) [RMSE 0.020]; SGMM = 0.011 (S.E. (0.001)) [RMSE 0.019]; SGMMR = 0.012 (S.E. (0.002)) [RMSE 0.018].
    - Scenario II, ρ = 0.06: OLS = 0.022 (S.E. (0.002)) [RMSE 0.038]; LSDV = 0.080 (S.E. (0.017)) [RMSE 0.026]; GMM = 0.092 (S.E. (0.036)) [RMSE 0.048].
    - Scenario III illustrates that LSDVIV estimates for β can have extremely large S.E. and RMSE (e.g., LSDVIV reported S.E. (1.768) and [RMSE] [1.767] for some β estimates), indicating weak-instruments/overfitting concerns when many instruments including dummies are used.

### VI. Practical implications and recommendations (from paper conclusions)
- Estimator choice recommendations:
  - For estimating ρ (fiscal reaction to debt / solvency tests): LSDV or LSDVIV methods produce precise estimates of ρ with modest biases in realistic short-panel settings (bias often < 10% of true ρ for T ≥ 10 under calibrated scenarios).
  - When explanatory variables (e.g., output gap) are potentially endogenous and coefficients on those variables are of primary interest (countercyclicality tests), GMM methods (Arellano and Bond (1991) difference GMM and Blundell and Bond (1998) system GMM) are preferred using exogenous instruments for the output gap.
- Specific cautions:
  - Presence of exogenous shocks to debt (b i,t with 2 b σ > 0) reduces LSDV bias on ρ.
  - Inclusion of exogenous regressors increases expected magnitude of LSDV bias on ρ.
  - Endogenous regressors have ambiguous analytic effect on LSDV bias; instrumenting such regressors is necessary, but instrumentation combined with country dummies can increase bias for ρ while reducing RMSE for coefficients on endogenous variables.
  - Use of many instruments (including country dummies) in first-stage IV (LSDVIV) can inflate RMSE and standard errors for coefficients like β (weak instruments problem).
- Bottom-line guidance:
  - Use LSDV (country dummies) or GMM methods depending on parameter of interest:
    - Prefer LSDV (or LSDVIV when instrumentation for y is required) for intertemporal solvency tests (coefficient ρ).
    - Prefer GMM (difference or system) when testing fiscal policy countercyclicality (coefficients on potentially endogenous variables like output gap).

*Italicized source attribution preserved as provided in the original content unit.*

### References

### References

### Full list of references

- Abiad, Abdul, and Jonathan D. Ostry, 2005, “Primary Surpluses and Sustainable Debt Levelsin Emerging Market Countries,” IMF Policy Discussion Paper 05/6 (Washington: International Monetary Fund).
- Abiad, Abdul, and Taimur Baig, 2005, "Underlying Factors Driving Fiscal Effort in Emerging Market Economies," IMF Working Paper 05/106 (Washington: International Monetary Fund).
- Arellano, Manuel and Stephen R. Bond, 1991, “Some Specification Tests for Panel Data: Monte Carlo Evidence and an Application to Employment Equations,” Review of Economic Studies, Vol. 58, pp. 277–98.
- Blundell, Richard, and Stephen R. Bond, 1998, “Initial Conditions and Moment Restrictions in Dynamic Panel Data Models,” Journal of Econometrics, Vol. 87, pp. 115–43.
- Bohn, Henning, 1998, “The Behavior of U.S. Public Debt and Deficits”, Quarterly Journal of Economics, Vol. 113, pp. 949–63.
- Bond, Steven, 2002, “Dynamic Panel Data Models: A Guide to Microdata Methods and Practice,” Cemmap Working Paper Series No. CWP09/02 (London: Institute of Fiscal Studies).
- Celasun, O., X. Debrun, and J.D. Ostry, 2006, “Primary Surplus Behavior and Risks to Fiscal Sustainability in Emerging Market Countries: A ‘Fan-Chart’ Approach,” IMF Working Paper 06/67 (Washington: International Monetary Fund).
- Favero, Carlo, 2002, “How Do European Monetary and Fiscal Authorities Behave?” CEPR Discussion Paper No 3426 (London: Center for Economic Policy Research).
- Galí, Jordi, and Roberto Perotti, 2003, “Fiscal Policy and Monetary Integration in Europe,” Economic Policy, Vol. 37 (October), pp. 535–72.
- International Monetary Fund, (IMF) 2003, “Public Debt in Emerging Markets: Is It Too High?,” Chap. III in World Economic Outlook, September (Washington: International Monetary Fund).
- ———, 2004, “Has Fiscal Behavior Changed Under the European Economic and Monetary Union?” in Chap. II of World Economic Outlook, September (Washington: International Monetary Fund).
- Judson, Ruth A., and Ann L. Owen, 1999, “Estimating Panel Data Models: A Guide for Macroeconomists,” Economics Letters, Vol. 65, pp. 9–15.
- - 27 -
- Kiviet, Jan F., 1995, “On Bias, Inconsistency, and Efficiency of Various Estimators in Dynamic Panel Data Models,” Journal of Econometrics, Vol. 68, pp. 53–78.
- Mélitz, Jacques, 1997, “Some Cross-country Evidence About Debts, Deficits and the Behavior of Monetary and Fiscal Authorities," CEPR Discussion Papers, No. 1653 (London: Center for Economic and Policy Research).
- Nickell, Steven, 1981, “Biases in Dynamic Models with Fixed Effects,” Econometrica, Vol. 49, pp. 1417–26.
- Stock, James H., and Moto Yogo, 2005, “Testing for Weak Instruments in Linear IV Regressions,” Chap. 5 in Identification and Inference for Econometric Models: A Festschrift in Honor of Thomas Rothenberg, ed. by J.H. Stock and D.W.K Andrews, (Cambridge, United Kingdom: Cambridge University Press).
- Wooldridge, Jeffrey M., 2001, Econometric Analysis of Cross Section and Panel Data (Cambridge, Massachusetts: MIT Press).
- Wyplosz, Charles, 2005, “Institutions for Debt Sustainability in Brazil,” Chap. 6 in Francesco,Giavazzi, Ilan Goldfajn, and Santiago Herrera, Inflation Targeting, Debt, and the Brazilian Experience, 1999 to 2003 (Cambridge, Massachusetts: MIT Press).

*Source: _wp06182 - References*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06182.pdf_