Approximate maximum likelihood estimation in linear regression (Q1260703)

Consider the linear regression model \(y_ j=x_ j'\beta+e_ j\), \(1\leq j\leq n\), where \(\beta\in R^ m\) and \(e_ j\) are unobserved errors. The application of the ML method in the linear regression model requires a parametric form for the error density. When this is not available, the density may be parametrized by its cumulants \((\kappa_ 1)\) and then the ML applied. In this article, results are derived under the assumption (assumption 3) that the standardized cumulants \(\gamma_ i\) satisfy \(\gamma_ i=\kappa_{i+2}/\kappa_ 2^{(i+2)/2}=O(\nu)\) as \(\nu\to 0\) for \(i>0\). From this assumption, the distribution associated with the log likelihood of the disturbances \(e_ i\) can be obtained. In addition, when other assumptions (assumption 1 and 2) are satisfied, a simple form of the inverse AIM (approximate information matrix) can be determined and the AML (approximate maximum likelihood) estimator (denoted by \(\hat\theta)\) of \(\theta\) can be obtained by using iteration methods. Using assumptions 1, 2 and 4, then as \(n\to\infty,\) \[ n^{1/2}(\hat\theta-\theta)@>d>>N[0, \lim V(\theta')]. \] The author gives the asymptotic relative efficiency \(e\) of the AML estimator with respect to the LS estimator. When \(e>1\), the length of the AML confidence interval for the slope parameters is shorter than the length of the LS confidence interval and using the AML estimators to test the hypotheses on the slope parameters is more powerful than using the LS estimators to test the hypotheses. The author also derives the bias of the AML estimators. The spirit of the AML method is to use a simple approximation to reproduce the distribution of the disturbance and then to estimate the parameters.