Challenge of missing data in observational studies : investigating cross-sectional imputation methods for assessing disease activity in axial spondyloarthritis

Abstract

Objectives We aimed to compare various methods for imputing disease activity in longitudinally collected observational data of patients with axial spondyloarthritis (axSpA). Methods We conducted a simulation study on data from 8583 axSpA patients from ten European registries. Disease activity was assessed by the Axial Spondyloarthritis Disease Activity Score (ASDAS) and the corresponding low disease activity (LDA; ASDAS<2.1) state at baseline, 6 and 12 months. We focused on cross-sectional methods which impute missing values of an individual at a particular time point based on the available information from other individuals at that time point. We applied nine single and five multiple imputation methods, covering mean, regression and hot deck methods. The performance of each imputation method was evaluated via relative bias and coverage of 95% confidence intervals for the mean ASDAS and the derived proportion of patients in LDA. Results Hot deck imputation methods outperformed mean and regression methods, particularly when assessing LDA. Multiple imputation procedures provided better coverage than the corresponding single imputation ones. However, none of the evaluated methods produced unbiased estimates with adequate coverage across all time points, with performance for missing baseline data being worse than for missing follow-up data. Predictive mean and weighted predictive mean hot deck imputation procedures consistently provided results with low bias Conclusions This study contributes to the available methods for imputing disease activity in observational research. Hot deck imputation using predictive mean matching exhibited the highest robustness and is thus our suggested approach.