MODEL UNCERTAINTY IN CROSS-COUNTRY GROWTH REGRESSIONS

MODEL UNCERTAINTY IN CROSS-COUNTRY GROWTH REGRESSIONS

January 20, 2001 | CARMEN FERNÁNDEZ, EDUARDO LEY, MARK F. J. STEEL
This paper investigates model uncertainty in cross-country growth regressions using Bayesian Model Averaging (BMA). The authors find that the posterior probabilities are widely distributed among many models, suggesting the superiority of BMA over selecting a single model. Out-of-sample predictive results support this claim. Unlike Levine and Renelt (1992), the authors' results broadly support Sala-i-Martin (1997b)'s more optimistic conclusion that some variables are important for explaining cross-country growth patterns. However, the methodology used should be carefully considered. The proposed approach is grounded in statistical theory and provides clear posterior and predictive inference. The paper focuses on a large set of possible models, allowing for any subset of 41 regressors, leading to over two trillion different models. Novel Markov chain Monte Carlo (MCMC) techniques, specifically the Markov chain Monte Carlo Model Composition (MC\textsuperscript{3}) sampler, are used to handle the computational challenge. The findings are based on the same data as Sala-i-Martin, covering 140 countries over the period 1960-1992. The authors identify a relatively large number of variables as important for growth regression, differing slightly from Sala-i-Martin's results. They advocate using BMA rather than selecting a subset of regressors, providing a clear interpretation of the results and a formal statistical basis for inference. The paper also discusses the predictive performance of different regression strategies, showing that BMA outperforms other methods in terms of predictive accuracy.This paper investigates model uncertainty in cross-country growth regressions using Bayesian Model Averaging (BMA). The authors find that the posterior probabilities are widely distributed among many models, suggesting the superiority of BMA over selecting a single model. Out-of-sample predictive results support this claim. Unlike Levine and Renelt (1992), the authors' results broadly support Sala-i-Martin (1997b)'s more optimistic conclusion that some variables are important for explaining cross-country growth patterns. However, the methodology used should be carefully considered. The proposed approach is grounded in statistical theory and provides clear posterior and predictive inference. The paper focuses on a large set of possible models, allowing for any subset of 41 regressors, leading to over two trillion different models. Novel Markov chain Monte Carlo (MCMC) techniques, specifically the Markov chain Monte Carlo Model Composition (MC\textsuperscript{3}) sampler, are used to handle the computational challenge. The findings are based on the same data as Sala-i-Martin, covering 140 countries over the period 1960-1992. The authors identify a relatively large number of variables as important for growth regression, differing slightly from Sala-i-Martin's results. They advocate using BMA rather than selecting a subset of regressors, providing a clear interpretation of the results and a formal statistical basis for inference. The paper also discusses the predictive performance of different regression strategies, showing that BMA outperforms other methods in terms of predictive accuracy.
Reach us at info@study.space
[slides and audio] Model uncertainty in cross-country growth regressions