Mediation Analysis in Categorical Variables under Non-Ignorable Missing Data Mechanisms

A mediating variable is a variable that is intermediate in the causal path relating an independent variable to a dependent variable in statistical analysis. The mediation analysis of using a categorical predictor, mediator, and outcome variables has been investigated in the literature. It is extremely common to have missing data even after having a well-controlled study. It is also well known that missingness, especially the non-ignorable missing, in a dataset has often been proven to produce biased results. This paper uses the extended Baker, Rosenberger, and Dersimonian (BRD) model to estimate the mediation effect under non-ignorable missing mechanisms. This paper also proposes four identifiable models to estimate the mediation effect for missingness in one categorical variable with two fully observed categorical variables. We reported the relative bias and Mean Square Error to compare the performance of the proposed BRD models against the Complete Case and Multiple Imputation methods in estimating the mediated effect (a^ b^) under the non-ignorable missing mechanism. The application of these models in estimating the mediated effect was demonstrated using the Multiple Risk Factor Intervention Trial datasets.