On the computational competitiveness of full-information maximum-likelihood and three-stage least-squares in the estimation of nonlinear, simultaneous-equations models

Abstract

While full-information maximum-likehood (FIML) estimation has long been considered an important theoretical econometric estimation technique, computational considerations have greatly restricted its use in practice. Recent advances in numerical analysis and in computational software, however, have combined to provide algorithms capable of carrying out the FIML calculations quite efficiently relative to past standards. This paper compares the computational competitiveness of FIML with its most popular competitor, 3SLS, in the estimation of a variety of linear and non-linear (in parameters and variables) models. The nonlinear full-information maximum-likelihood (NLFIML) estimator is described and a computatíonally efficient approximation, TRUNFIML, is defined. Nonlinear three-stage least-squares (NL3SLS) is accomplished by the method of Jorgenson-Laffont. Comparisons are made on the basis of numbers of iterations to convergence, number of function evaluations, and total computer CPU time required, this latter figure being most relevant to a comparison of computational effort and cost.