B0478
Title: Practical approach for missing data sensitivity analyses in joint modelling of cognition and dementia risk
Authors: Tetiana Gorbach - Umea University (Sweden) [presenting]
James Carpenter - London School of Hygiene and Tropical Medicine (United Kingdom)
Chris Frost - London School of Hygiene and Tropical Medicine (United Kingdom)
Maria Josefsson - Umea School of Business, Economics and Statistics (Sweden)
Amy MacDougall - London School of Hygiene and Tropical Medicine (United Kingdom)
Jennifer Nicholas - London School of Hygiene and Tropical Medicine (United Kingdom)
Lars Nyberg - Umea University (Sweden)
Abstract: Joint modelling of longitudinal cognitive measures and time-to-dementia onset is a natural tool for understanding the relationship between the trajectory of cognitive decline and dementia. In the joint model, the longitudinal data is typically represented through a linear mixed effect submodel, while the time-to-dementia data is represented via the Cox proportional hazards submodel. Both the longitudinal submodel and the survival submodel yield valid inferences when data are missing (censored) at random. Unfortunately, the dropout from the longitudinal studies of ageing might be non-ignorable. A practical imputation-based approach is proposed for exploring the sensitivity of inferences to such non-ignorable dropouts. For the sensitivity analysis: (a) missing longitudinal cognitive measurements are imputed using a pattern-mixture approach applied to the linear mixed effect submodel and while accounting for a possible accelerated rate of cognitive decline after dropout, represented by a sensitivity parameter; contextual knowledge is used to inform the choice of the sensitivity parameter values; (b) the joint model is fitted to each imputed data set and (c) the results using Rubins rules are combined. The approach is used to infer the relationship between memory and the risk of dementia in the Betula longitudinal study. It is shown that the inferences in the Betula study are robust to contextually plausible non-ignorable missing in longitudinal cognitive measures.
