Tutorial: Using Random Forest Analysis to Identify Auxiliary Variables of Missing Data

Missing data is a pervasive problem in research; prevention science is particularly vulnerable due to the study designs used and types of data collected. Recommended approaches to address missing data include full information maximum likelihood estimation and multiple imputation, both of which rely on identification of auxiliary variables related to missingness. The methodological literature recommends including as many potential auxiliary variables as possible, but in practice, that is often infeasible and a researcher must select a more limited number. Prior studies have shown that traditional methods for identifying auxiliary variables do not perform well when missingness follows a nonlinear functional form, but machine learning methods such as random forest analysis (RFA) perform well at successfully identifying correlates of missingness across a variety of missing at random patterns. RFA models can also provide measures of variable importance, allowing researchers to prioritize inclusion of the most relevant variables. Methods like RFA are less familiar to prevention researchers, which serves as an impediment for researchers to use RFA to identify auxiliary variables. This paper provides a tutorial to use RFA to identify correlates of missingness using a real data sample of $n$ = 215 participants with more than 100 variables measuring demographics, personality, and psychopathology. The tutorial demonstrates how to run an RFA (including using measures of variable importance to select the variables most related to missingness) and how to incorporate the selected auxiliary variables into an analysis.