ESTIMATION OF SEEMINGLY UNRELATED REGRESSIONS WITH AN INCOMPLETE SET OF DATA
Abstract
Several authors have considered the problem of estimating the linear regression equation when the data set is characterized by a number of missing observations. The techniques developed so far have been confined to single equation methods. The purpose of this study is to extend the literature on missing data estimation to a system of grouped equations. An existing procedure for estimating the coefficients and missing observations on the dependent variable in a single equation will be combined with a separate regression having a complete set of data into one model. Under the assumption that the disturbance terms in the two equations are contemporaneously correlated, the best linear unbiased estimators for both the structural parameters and the missing observations in the resulting Seemingly Unrelated Regression model are given. A statistic designed to test the significance of the correlation coefficient is developed according to the Likelihood Ratio principle. Two empirical studies illustrate the prediction method. The first focuses on the debate in the field of finance concerning asset pricing theory. It is claimed that the statistical model presented here is an improvement over a popular methodology used in the analysis of market efficiency. Further, a test is suggested for examining the residuals of the single market factor equilibrium model. The second investigation examines the construction of economic index numbers when there is an insufficient supply of data available for the direct computation of one of the standard formulas. The estimation is based on a regression equation formed from a stochastic Cobb-Douglas production function. In both empirical studies evidence is provided to substantiate the gains in efficiency accruing to the grouped equation system.
Degree
Ph.D.
Subject Area
Finance
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.