Incorporating a mixture approach to the available case missing variable assumption in pattern mixture models
Abby Flynt (Bucknell University)
Friday 22nd May, 2015 15:00-16:00 Maths 204
Finite mixture models are a class of probabilistic models that assume that a heterogeneous population is actually a weighted combination of homogeneous subpopulations. One type of finite mixture model, known as Pattern Mixture Models (PMMs) was developed for modeling potentially informative missingness in longitudinal data. PMMs assume a different response model for each pattern of missing data and treat the full data as a mixture of these models. Further, PMMs require an assumption for missing data estimation. One such assumption commonly used for its simplicity, is known as the available case missing variable (ACMV) assumption. ACMV uses all available data aggregated over missingness pattern for estimation, thus implying that the underlying models for said missingness patterns are similar. In this talk I will propose an alternative to the ACMV assumption which recognizes the underlying differences in the data based on missingness pattern by using a mixture of these patterns for missing data estimation.