Flavia Remo, Title: Non-ignorable missing data under heterogeneity in a meta-analysis with binary outcomes

Abstract: We estimate the pooled treatment effect size in form of the log odds Ratio and heterogeneity variance in a meta-analysis. With missingness and heterogeneity being the most common challenges that hamper data in a meta-analysis, we employ the method of finite mixture modelling combined with multiple imputation to estimate the model parameters using the EM algorithm. At the same time we impute the data that is missing not at random (MNAR) using the augmentation method simultaneously in line with the response mechanism after the work of Lehmann and Schlattmanm (2017). As we are dealing with binary outcomes, we use the standard method of considering a finite mixture of logistic regression models to estimate the regression parameters which translate into log odds as our effect size estimate. We will illustrate our results with a meta-analysis of RCTs comparing haloperidol with placebo in treatment of schizophrenia (see Higgins, White and Wood, 2008).