Handling missing data in a composite outcome with partially observed components: Application in clustered paediatric routine data.

Abstract

Background: In health care settings, composite measures are used to combine information from multiple quality of care measures into a single summary score. Composite scores provide global insights and trends about complex and multidimensional quality of care processes. However, missing data in subcomponents may hinder the overall reliability of the composite measures in subsequent analysis and inferences. In this study we demonstrate strategies for handling missing data in Paediatric Admission Quality of Care (PAQC) score, a composite outcome which summarizes quality of inpatient paediatric care in low income settings.Methods: We analysed routine data collected in a cluster randomized trial in 12 Kenyan hospitals. Multilevel multiple imputation (MI) within joint model framework was used to fill-in missing values in selected PAQC score subcomponents and partially observed covariates across two levels of hierarchy. We used proportional odds random intercepts and generalized estimating equations (GEE) models to analyse PAQC score before and after multiple imputation. Using a set of simulations scenario, that is, varied proportions of missingness in PAQC score subcomponents of interest under missing at random and missing completely at random mechanisms respectively, we compared the magnitude of bias in parameter estimates obtained under MI and the conventional method in addressing missing data in PAQC score components. Under the conventional method we scored all missing PAQC score components with value 0.Results: Results from observed data showed that multiple imputation of both PAQC score components and covariates yielded more accurate and precise estimates compared to complete case analysis. From the simulation study, the conventional missing data method led to significantly larger biases in estimated proportional log odds of the outcome compared to MI methods. The amount of bias increased with increase in rate of missingness with substantial variation between the missing data mechanisms under the conventional method. Conclusion: In comparison with conventional method, MI produce minimally biased estimates regardless of amount of missing data rate and underlying mechanism. We therefore recommend avoiding the conventional method in favour of multiple imputation; more research is needed to compare different ways of performing multiple imputation at the component and composite outcome level.TRIAL REGISTRATION: US National Institutes of Health-ClinicalTrials.gov identifier (NCT number) NCT02817971 . Registered September 28, 2016-retrospectively registered.