Consistent maximum-likelihood estimation with dependent observations. The general (nonnormal) case and the normal case (Q1112529)

Conditions for existence and weak consistency of the maximum likelihood estimators (mle's) are established in a multiparameter framework assuming neither the independence nor the identical distribution of the observations. The results justify the universal consensus among econometricians that the ml method yields consistent estimators under mild but not trivial conditions. Virtually all attempts at establishing conditions under which the mle is consistent are based on \textit{H. Cramér}'s approach [Mathematical methods of statistics. Princeton (1946)] or \textit{A. Wald}'s approach [Ann. Math. Statist., Baltimore Md. 20, 595-601 (1949; Zbl 0034.229)]. Cramér assumed differentiability but Wald did not assume differentiability of the log-likelihood, and Wald's approach is adopted in the present paper. There are two main parts to the paper. In the first part, two theorems are presented on the weak consistency of the mle obtained from observations that are dependent in general: Theorem 1 contains necessary and sufficient conditions while Theorem 2 contains stronger but more readily applicable conditions. In the second part, a set of observations \(y=(y_ 1,y_ 2,...,y_ n)\) is assumed to be normally distributed: \(y=N(\mu (\gamma_ 0),\Omega (\gamma_ 0))\). The weak consistency of the mle obtained from normal (dependent) observations is proved in Theorem 3, and it is proved under stronger conditions in Theorem 4. Two examples - the linear regression model with a general covariance structure and the nonlinear regression model with first-order autocorrelation - are presented.