Bayesian Model Averaging and Endogeneity Under Model Uncertainty : An Application to Development Determinants

Recent approaches to development accounting reflect substantial model uncertainty at both the instrument and the development determinant level. Bayesian Model Averaging (BMA) has been proven useful in resolving model uncertainty in economics, and we extend BMA to formally account for model uncertainty in the presence of endogeneity. The new methodology is shown to be highly efficient and to reduce many instrument bias; in a simulation study we found that IVBMA estimates reduced mean squared error by 60% over standard IV estimates. We also introduce Bayesian over and under-identification tests that are based on model averaged predictive p-values. This approach is shown to mitigate the reduction in power these tests experience as dimension increases. In a simulation study where the exogeneity of the instrument is compromised we show that the classical Sargan test has a power of 0.2% while our Bayesian over-identification test has a power of 98% at detecting the violation of the exogeneity assumption. An application of our method to a prominent development accounting approach leads to new insights regarding the primacy of institutions. Using identical data and robustness specifications we find support not only for institutions, but also for geography and integration, once both model uncertainty and endogeneity have been jointly addressed