Linear Regression with Random Coefficients: The Finite Sample and Convergence Properties

Abstract

This article studies sampling properties of several unconstrained estimators in the random coefficients model. It is found that the ranking of alternative estimators is sometimes sensitive to selection of true values of the model's parameters. Further, the frequency of negative estimates increases with the size of true variance. Though all estimators converge in 80 to 95 percent of cases, there is generally no efficiency gain at any stage of iterations, possibly because large numbers of negative estimates of variances are obtained. Sampling properties of alternative estimators for variance of the mean response estimates are also examined.