2011年-IMF国际货币组织全球_Limited_Information_Bayesian_Model_Averaging_for_Dynamic_Panels_with_An_Application_to_a_Trade_Gravity_Model_46页_1mb

Summary of "Limited Information Bayesian Model Averaging for Dynamic Panels with an Application to a Trade Gravity Model"

Core Content

This paper introduces a Limited Information Bayesian Model Averaging (LIBMA) methodology to address model uncertainty in dynamic panel data models that include endogenous regressors. The approach is designed to work with short time periods and is applied to the estimation of a dynamic trade gravity model.

Main Points

1. Model Uncertainty in Bayesian Context

• Model uncertainty arises due to the lack of clear theoretical guidance and multiple plausible model specifications.
• Bayesian Model Averaging (BMA) is a method that incorporates model uncertainty by averaging over all possible models using posterior model probabilities as weights.
• Bayes factors are used to compare models and update prior odds ratios based on observed data.
• The paper emphasizes that BMA provides a more robust inference than selecting a single model, especially when model space is not concentrated on one model.

2. Bayesian Model Averaging (BMA)

• BMA computes the posterior distribution of a parameter of interest by averaging across all models.
• The posterior mean of a parameter is a weighted average of the model-specific posterior means.
• Inclusion probabilities are computed for each regressor to assess its relevance in the model.
• Challenges in implementing BMA include computing the marginal likelihood, handling a large number of models, and specifying prior model probabilities and parameter priors.

3. Choice of Priors

• The paper uses a unit information prior for the parameters, which is a multivariate normal distribution with mean equal to the maximum likelihood estimate and variance equal to the inverse of the Fisher information matrix.
• This prior simplifies the Bayesian Information Criterion (BIC) and reduces computational complexity.
• The Normal-Gamma prior is also mentioned for Gaussian models, but the unit information prior is preferred for its simplicity and effectiveness in this context.

LIBMA Methodology

4. Dynamic Panel Data Model with Endogenous Regressors

• The model includes lagged dependent variables, exogenous variables, and endogenous variables.
• The model is given by:

$$
y_{it} = (y_{i,t-1}, x_{it}, w_{it})M \cdot (\alpha, \theta_x, \theta_w)' + u{it}
$$

• $u_{it}$ is decomposed into individual effects ($\eta_i$) and idiosyncratic error ($v_{it}$).
• The moment conditions are derived for both the level equation and the first difference equation.

5. Moment Conditions

• For the first difference equation, the following moment conditions are used:

$$
E(y_{i,t-s} \Delta v_{it}) = 0, \quad s = 2, \dots, t
$$
$$
E(x_{it}^l \Delta v_{it}) = 0, \quad l = 1, \dots, m
$$
$$
E(w_{i,t-s}^l \Delta v_{it}) = 0, \quad s = 2, \dots, t-1
$$

• For the level equation, the following moment conditions are used:

$$
E(\Delta y_{i,t-1} u_{it}) = 0
$$
$$
E(\Delta w_{i,t-1}^l u_{it}) = 0, \quad l = 1, \dots, q
$$
$$
E(\Delta x_{it}^l u_{it}) = 0, \quad l = 1, \dots, m
$$

• These moment conditions are used to estimate the model using Generalized Method of Moments (GMM).

6. Limited Information Criterion

• The paper proposes a limited information likelihood based on the moment conditions.
• This likelihood is derived using the Central Limit Theorem and the linear structure of the model.
• The limited information criterion is used to perform Bayesian model selection and averaging.
• Unlike other approaches that use quasi-likelihoods, this method uses a simpler Bayesian procedure to construct the likelihood.

Simulation Results

• The methodology is evaluated using Monte Carlo simulations.
• Results show that LIBMA performs well asymptotically in both model selection and model averaging.
• It accurately recovers the data generating process with high posterior inclusion probabilities for relevant regressors and parameter estimates close to true values.
• The method is robust to non-Gaussian errors.

Application to Trade Gravity Model

• The paper applies LIBMA to a dynamic gravity model for bilateral trade.
• This application demonstrates the practical usefulness of the LIBMA approach in real-world econometric modeling.
• The model includes trade flows, distance, GDP, population, and common language as variables.
• The results suggest that LIBMA is well-suited for inference in dynamic panels with endogenous regressors under model uncertainty.

Conclusion

• The paper concludes that LIBMA is a robust and effective method for handling model uncertainty in dynamic panel data models.
• It is particularly useful when endogenous regressors are present and when the number of time periods is small.
• The method avoids the computational burden of full Bayesian inference while still incorporating model uncertainty through posterior model probabilities.

Key Information

• Methodology: Limited Information Bayesian Model Averaging (LIBMA)
• Application: Dynamic trade gravity model
• Key Features:
  • Uses moment conditions for likelihood construction
  • Incorporates Bayesian model selection and averaging
  • Handles endogenous regressors
  • Reduces computational burden
• Performance:
  • Performs well asymptotically
  • Provides accurate parameter estimates
  • Has good robustness to non-Gaussian errors
• Priors:
  • Uses unit information prior
  • Avoids sensitivity to prior parameters
• Model Structure:
  • Includes lagged dependent variables
  • Includes exogenous and endogenous variables
  • Uses GMM estimation framework
• Simulation:
  • Used to evaluate LIBMA performance
  • Showed superior predictive performance compared to model selection alone

References and Appendices

• The paper includes references and appendices with detailed moment equations and likelihood derivations.
• Appendix I provides the matrix representation of the moment conditions for the dynamic panel model.