Strategies for bias reduction in estimation of marginal means with data missing at random

Incomplete data are common in many fields of research, and interest often lies in estimating a marginal mean based on available information. This paper is concerned with the comparison of different strategies for estimating the marginal mean of a response when data are missing at random. We evaluate these methods based on the asymptotic bias, empirical bias and efficiency.We show that complete case analysis gives biased results when data are missing at random, but inverse probability weighted estimating equations (IPWEE) and a method based on the expected conditionalmean (ECM) yield consistent estimators. While these methods give estimators which behave similarly in the contexts studied they are based on quite different assumptions. The IPWEE approach requires analysts to specify a model for the missing data mechanism whereas the ECMapproach requires a model for the distribution of auxiliary variables driving the missing data mechanism. The latter can be a challenge in practice, particularly when the covariates are of high dimension or are a mixture of continuous and categorical variables. The IPWEE approach therefore has considerable appeal in many practical settings.