We consider M-estimators for a class of semiparametric mixed-effect models without time-dependent covariates and show that the simple marginal estimation method is generally better than the same M-estimator applied to the de-correlated response based on a known or estimated covariance matrix for each subject.