estimation and inference with weak instruments and near exogeneity

Empirical economic studies are often confronted by the joint problem of weak instruments and near exogeneity, such as labor economics and empirical economic growth theory. This dissertation presents new evidence and solutions on estimation and inference with weak instruments and near exogeneity. Chapter 1 reexamines the effect of institutions on economic performance in Acemoglu, Johnson and Robinson (2001) where the measurement of current institutions is instrumented by European settler mortality rates. Since many economists argue that the settler mortality rates can possibly affect economic performance through other channels, I reexamine the effect of institutions by considering near exogeneity. I provide some evidence to show that the effect of institutions is not significant in many regression specifications when the settler mortality rates are used as the main instrument.Chapter 2 studies estimation and inference with weak instruments and near exogeneity in a linear simultaneous equations model. I show that near exogeneity can exaggerate asymptotic bias of the TSLS and the LIML estimators. When using critical values from chi-square distributions, Anderson-Rubin and Kleibergen tests under exogeneity have a large size distortion. I propose the delete-d jackknife based Anderson-Rubin and Kleibergen tests to automatically reduce the size distortion in finite samples without a need for any pretest of exogeneity.Chapter 3 extends estimation and inference with weak identification and near exogeneity into a GMM framework with instrumental variables. A GMM framework allows nonlinear and nondifferentiable moment conditions. I examine asymptotic results of one-step GMM estimator, two-step efficient GMM estimator and continuously updating estimator with weak identification and near exogeneity. Near exogeneity can produce relatively large bias for all these estimators. The Anderson-Rubin type and the Kleibergen type tests under near exogeneity converge in distribution to nonstandard distributions, which creates large size distortion when using critical values from chi-square distributions. The delete-d jackknife based approach can reduce the size distortio