Evaluation of imputation strategies for multi-centre studies: Application to a large clinical pathology dataset.

Abstract

As part of a strategy for accommodating missing data in large heterogeneous datasets, two Random Forest-based (RF) imputation methods, missForest and MICE were evaluated along with several strategies to help navigate the inherently incomplete structure of the dataset. A total of 3817 complete cases of clinical chemistry variables from a large-scale, multi-site preclinical longitudinal pathology study were used as an evaluation dataset. Three types of 'missingness' in various proportions were artificially introduced to compare imputation performance for different strategies including variable inclusion and stratification. MissForest was found to outperform MICE, being robust and capable of automatic variable selection. Stratification had minimal effect on missForest but severely deteriorated the performance of MICE. In general, storing and sharing datasets prior to any correction is a good practise, so that imputation can be performed on merged data if necessary.