Instrumental variables (IV) regression is widely used in economics for drawing causal inferences from observational data. The instrumental variables that are available are often weak, meaning that they provide little information, relative to the sample size, for estimating the parameters of interest. Standard first order asymptotic theory has been shown to provide a poor approximation to the finite sample behavior of IV estimators based on weak instruments. This dissertation develops improved inferential procedures for IV regression using weak instruments, concentrating specifically on the linear structural equations model and an autoregressive model for panel studies.; We consider two main problems. The first problem is construction of confidence intervals using weak instruments. The test statistics that are commonly used to form confidence intervals are highly non-pivotal in the presence of weak instruments. This causes difficulties for both standard asymptotic and bootstrap methods. We develop a resampling method to address the problem of non-pivotality that is based on the hybrid resampling approach of Chuang and Lai (1998, 2000). Simulation studies demonstrate that our approach significantly improves over standard asymptotic and bootstrap methods in the setting of weak instruments and performs comparably when the instruments are strong.; The second problem we consider is how to select among candidate variables and instruments when using IV regression. We develop a criterion-based model selection procedure that addresses shortcomings of several approaches proposed in the econometrics literature. Our procedure captures the tradeoff between potential bias and variance with respect to a sensible loss function. Theoretical analysis shows that our method has better large sample properties under strong instruments than a previously proposed approach in the econometrics literature. Simulation studies show that our method performs reasonably well in the presence of weak instruments and yields significant improvement over previous approaches.; We apply our methodology to an IV regression arising in an empirical study of the impact of agricultural commercialization on nutrition. We also develop an approach to analyzing the sensitivity of inferences to violations of the assumptions behind IV regression and use the empirical study to illustrate our approach.
展开▼
