Abstract: In a study with longitudinal outcomes, the outcome nonresponse mechanism often depends on the observed or unobserved value of the outcome. When nonresponse is monotone in the sense that a subject having a missing outcome at time is not observed after time , Tang, Little, and Raghunathan (2003) developed a semiparametric pseudo-likelihood method for the estimation of parameters of interest. In practice, however, nonresponse is often not monotone and a direct application of their method discards observed data from subjects having nonmonotone nonresponse, which may result in inefficient estimators. We extend the idea in Tang et al. (2003) to nonmonotone nonresponse and construct a semiparametric pseudo-likelihood that utilizes all observed data. Asymptotic normality of the maximum pseudo-likelihood estimators is established. An application is made to the household income data from the Health and Retirement Study. Simulation results are also presented to examine finite sample properties of the proposed estimators.
Key words and phrases: Efficiency, nonignorable missing, semiparametric likelihood, sequential estimation.