Handling of missing component information for common composite score outcomes used in axial spondyloarthritis research when complete-case analysis is unbiased.

Abstract

Background: Observational data on composite scores often comes with missing component information. When a complete-case (CC) analysis of composite scores is unbiased, preferable approaches of dealing with missing component information should also be unbiased and provide a more precise estimate. We assessed the performance of several methods compared to CC analysis in estimating the means of common composite scores used in axial spondyloarthritis research.

Methods: Individual mean imputation (IMI), the modified formula method (MF), overall mean imputation (OMI), and multiple imputation of missing component values (MI) were assessed either analytically or by means of simulations from available data collected across Europe. Their performance in estimating the means of the Bath Ankylosing Spondylitis Disease Activity Index (BASDAI), the Bath Ankylosing Spondylitis Functional Index (BASFI), and the Ankylosing Spondylitis Disease Activity Score based on C-reactive protein (ASDAS-CRP) in cases where component information was set missing completely at random was compared to the CC approach based on bias, variance, and coverage.

Results: Like the MF method, IMI uses a modified formula for observations with missing components resulting in modified composite scores. In the case of an unbiased CC approach, these two methods yielded representative samples of the distribution arising from a mixture of the original and modified composite scores, which, however, could not be considered the same as the distribution of the original score. The IMI and MF method are, thus, intrinsically biased. OMI provided an unbiased mean but displayed a complex dependence structure among observations that, if not accounted for, resulted in severe coverage issues. MI improved precision compared to CC and gave unbiased means and proper coverage as long as the extent of missingness was not too large.

Conclusions: MI of missing component values was the only method found successful in retaining CC's unbiasedness and in providing increased precision for estimating the means of BASDAI, BASFI, and ASDAS-CRP. However, since MI is susceptible to incorrect implementation and its performance may become questionable with increasing missingness, we consider the implementation of an error-free CC approach a valid and valuable option.

Trial registration: Not applicable as study uses data from patient registries.