On post dimension reduction statistical inference (Q2196236)

This paper develops a general approach to assess how the error induced by dimension reduction propagates into the subsequent model estimation. To do that, the matrix-valued statistical functional of the sufficient dimension reduction method and the vector-valued statistical functional of the estimation method are employed. The influence functions of the statistical functionals are combined to determine the post dimension reduction asymptotic distribution. Within this general framework, the influence functions for particular sufficient dimension reduction methods and particular model estimation methods are explicitly derived. These particular sufficient dimension reduction methods include sliced inverse regression, sliced average variance estimation, two forms of principal Hessian directions and directional regression. The model estimation methods include differentiable estimating equations, nondifferentiable estimating equations and the generalized method of moments. Based on the derived post dimension reduction asymptotic distribution, specific methods for conducting statistical inference are derived. Confidence intervals and test statistics are built and the asymptotic null and local alternative distributions of the test statistics are determined. Then, the asymptotic covariances of the estimated parameter with and without taking into account the error induced by dimension reduction are compared. As for the finite-sample performance of the inference method, simulations are employed. A real data analysis of the BigMac dataset is also presented.