A partial profile empirical likelihood for a semiparametric mixture model (Zou et al., 2002) is shown to originate in a conditional likelihood involving additional nuisance parameters. The partial likelihood is the conditional likelihood with the nuisance parameters replaced by their estimators from the full likelihood. The conditional likelihood suggests alternative estimators. We demonstrate that the partial likelihood estimator is more efficient than an estimator for which the nuisance parameters are known. The practical implications of this counter-intuitive result are discussed. Copyright Biometrika Trust 2002, Oxford University Press.
