How and why to follow best practices for testing mediation models with missing data

Mediation models are often conducted in psychology to understand mechanisms and processes of change. However, current best practices for handling missing data in mediation models are not always used by researchers. Missing data methods, such as full information maximum likelihood (FIML) and multiple imputation (MI), are best practice methods of handling missing data. However, FIML or MI are rarely used to handle missing data when testing mediation models, instead analyses used listwise deletion methods, the default in popular software. Compared to listwise deletion, the implementation of FIML or MI to handle missing data reduces parameter estimate bias, while maintaining the sample collected to maximise power and generalizability of results. In this tutorial, we review how to implement full-information maximum likelihood and MI using best practice methods of testing the indirect effect. We demonstrate how to implement these methods using both R and JASP, which are both free, open-source software programmes and provide online supplemental materials for these demonstrations. These methods are demonstrated using two example analyses, one using a cross-sectional mediation model and one using a longitudinal mediation model examining how student-athletes reported worry about COVID predicts their perceived stress, which in turn predicts satisfaction with life.