The Impact of Missing Data Imputation on Model Performance and Explainability

Missing data is a common problem in data analysis processes. Data may be missing due to data collection or recording, which
can occur in any field, such as healthcare, transportation, telecommunications, biology, and more. Several machine learning models require
complete data. Therefore, imputation — the process of filling in missing data — is typically the preferred method. Missing data imputation methods are useful for predictive model analysis by filling in the missing values in the data. However, the impact of imputation techniques on model explainability is not yet known. In this study, we investigate the impact of white-box and black-box imputation methods on the explainability of the black-box Random Forest classification algorithm. Moreover, we applied the glass box classification model that allows us to see the model’s decisions in a transparent and understandable way, such as the Explainable Boosting Machine, to compare performance differences. We conducted an experiment with seven imputation methods, using three missing levels (10%, 50%, and 70%) and three datasets. We investigated the effect of the imputers on the feature importance of using SHAP values and applying surrogate models per imputed dataset. Summarize findings show that striving for the best predictive model performance, best explainability, or imputation performance leads to a different best choice for the imputer. It is, therefore, important, contrary to common practice, to regard the selection of imputer and classifier as a combined selection problem and to explicitly choose what to optimize for.