1996 Remington Lecture: Modeling multivariate longitudinal data that are incomplete

PURPOSE: We describe the impact that missing data may have on model selecti on for longitudinal multivariate data. METHODS: Maximum likelihood was used to fit several models to ultrasonograp hic measurements from the Asymptomatic Carotid Artery Progression Study (AC APS). Graphical techniques were used to examine evidence concerning the und erlying missing data mechanisms associated with each model. RESULTS: Using statistical methodology that addressed missing data substant ially increased the statisti cal efficiency of our analysis of ultrasonogra phic data. Only complex models that included segment-specific parameterizat ions for longitudinal correlations appeared to allow missing data to be ass umed to occur at random. CONCLUSION: Ignoring the nature of missing data in conducting statistical a nalyses can have serious consequences when missingness is not rare. It may be necessary to fit models of high dimension with maximum likelihood techni ques to address missing data appropriately, however these approaches may im prove statistical efficiency. Ann Epidemiol 1999;9:196-205. (C) 1999 Elsevi er Science Inc. All rights reserved.