B2314 - Handling multivariate missingness in health and behaviour data via multiple imputation under MNAR Issues in Practice - 31/10/2014
BACKGROUND:
Data collected by ALSPAC to explore potential predictors of adolscent alcohol, tobacco and cannabis consumption exhibit substantial levels of missingness in both outcomes and covariates and are plausibly missing not at random (MNAR). This is also the case for data collected by ALSPAC to explore potential correlates of adolescent mental health. Standard maximum likelihood and Bayesian full probability modelling procedures can quickly become computationally infeasible in the presence of multivariate MNAR missingness. We propose to tackle this difficulty by extending a popular procedure for performing multiple imputation under MAR, Fully Conditional Specification, to handle MNAR missingness mechanisms. Our method will be broadly applicable to future analyses of the ALSPAC data and to both cross-sectional and longitudinal data.
AIMS:
1. To develop a new statistical method for handling datasets with multivariate MNAR missingness.
2. To use this new method to further explore the missing data mechanism for
a. key variables on alcohol, tobacco and cannabis use in teenagers.
b. key variables on mental health outcomes in teenagers.