Allowing for uncertainty due to missing and LOCF imputed outcome data in meta-analysis
Dimitris Mavridis (University of Ioannina)
Friday 19th May, 2017 15:00-16:00 Room 311B, 3rd floor of maths and stats building
Missing outcome data are commonly encountered in clinical trials and may compromise the validity of the estimated treatment effects. We focus on mental health trials where outcome data are rarely missing at random (MAR) and the last observation carried forward (LOCF) is typically used for imputation. Meta-analysts typically assume that the missing data problem has been solved at the meta-analysis level and conduct an available case analysis. However, bias both from missing outcome data and flawed imputation methods in clinical trials is accumulated in a synthesis of those trials using meta-analysis. Our aim is to develop methods for estimating meta-analytic summary treatment effects when some outcome data have been imputed and others are missing. We used a pattern-mixture model to allow for uncertainty in the missing and imputed information and we quantified the degree of departure from the MAR assumption by adding two informative missingness parameters that relate the outcome in the missing data and in the LOCF imputed data to the outcome in the observed data. These parameters are not informed by the data and we can use expert opinion or conduct a sensitivity analysis. We illustrate the method using examples from psychosis and depression trials.