Empirical standard errors for longitudinal data mixed linear models (Q1775976)

The model under consideration is a special case of linear regression with independence among subjects. Such estimators as empirical covariation estimators and empirical standard errors (ESE) are widely used in statistical inferences connected with the mentioned regression models. The authors extend work of \textit{J. G. MacKinnon} and \textit{H. White} [J. Econom. 29, 53--57 (1985)] to the longitudinal data mixed linear models by considering alternative approaches to estimation of regression coefficient standard errors. The first type of the proposed estimators is jackknife, the second one is based on standardized residuals. The residual based estimators occur to be unbiased in many cases. Monte Carlo simulation experiments demonstrated that both alternatives have better small sample properties in hypothesis testing than the usual ESE, regardless whether heteroscedasticity is present. Tests based on both alternatives behave well in terms of size and power even when the covariance matrix is completely misspecified.