Abstract:
|
Multiple imputation is often used to handle missing data problems. When producing multiple imputations for the missing values, it is recommended that as many variables as possible be included. However, when the sample size is not large, a model with a large number of variables may easily be over-parameterized. Song and Belin (1999) introduced a method to overcome this difficulty by using a factor analysis model, which can reduce the number of parameters substantially. For longitudinal data, the factor model has the limitation of not reflecting the longitudinal structure. To overcome this deficiency, we develop a longitudinal factor analysis model, combining a factor structure to reflect cross-sectional correlations with a multivariate linear-mixed-model structure to reflect longitudinal correlations. The method is illustrated using data from a study comparing two oral surgery treatments, where several clinical and psychological outcomes were measured longitudinally.
|