Efficient Semiparametric Marginal Estimation for Longitudinal/Clustered Data

Abstract

There has been substantial recent interest in investigating the performance of non- and semi-parametric marginal estimation using kernel methods. Most approaches adopt the strategy of pretending observations within the same subject are independent. A result supporting the use of this "working independence" strategy indicates that under the traditional estimation structure, correct specification of the correlation actually diminishes the asymptotic efficiency. In this presentation, an alternative kernel weighted estimating equation that accounts for the within subject correlation is proposed. Asymptotically, the proposed method uniformly outperforms the most efficient working independence approach in nonparametric estimation. For the estimator of the parametric component, this new estimator is semiparametric efficient.