Gong, Jing. Modeling longitudinal data with mixed dropout mechanisms using extended pattern mixture model. Retrieved from https://doi.org/doi:10.7282/T37H1HDD
DescriptionIn physical and mental health research areas, longitudinal studies are popular tool for addressing outcome changes over time within and between individuals. However, monotone-type missing data caused by dropout is unavoidable in many longitudinal studies and may lead to biased inference and incorrect conclusions if the nature of dropout is ignored. The dropout process may cause three types missingness: Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR). Most of existed statistical methods have treated the entire dropout the same, although this assumption may not be true in practice. For a real longitudinal study, based on observed dropout reasons, it may be more realistic to assume the nature of missingness to be a mixture of MCAR, MAR and MNAR. In this thesis, two approaches are proposed to deal the mixed nature of dropout due to different longitudinal dataset’s pattern and assumptions, and both of them are based on Pattern Mixture Model. If the outcome of interest can be assumed to follow a linear trend over time and the detailed dropout reasons for each subject are unknown, EM algorithm method will be added to Pattern Mixture Model to reflect a mixture of missing natures. If the outcome of interest is not linear with respect time and the detailed reasons for each dropout are known, available-case missing value restriction and non-future-dependent missing value restriction will be used within Pattern Mixture Model to identify the distribution of unknown measurements caused by different dropout reasons. Multiple imputation method will be combined to impute multiple complete datasets to reflect the uncertainty caused by missing values.