Handling missing data in a composite outcome with partially observed components: simulation study based on clustered...

Area of Science:

• Health Services Research
• Biostatistics
• Pediatric Healthcare Quality

Background:

• Composite scores are valuable for assessing complex quality of care processes.
• Missing data in subcomponents can compromise the reliability of composite measures.
• The Paediatric Admission Quality of Care (PAQC) score is an important ordinal composite outcome measure.

Purpose of the Study:

• To evaluate strategies for handling missing data in the PAQC score.
• To compare a conventional zero-imputation method with a multiple imputation (MI) approach.
• To assess the impact of missing data on the reliability of composite quality of care scores.

Main Methods:

• A simulation study was conducted to assess different missing data handling strategies.
• The conventional method of scoring missing subcomponents as zero was compared to a latent normal joint modeling MI approach.
• The study analyzed bias in parameter estimates and standard errors under various missingness scenarios.

Main Results:

• Multiple imputation (MI) of missing PAQC score elements at the item level resulted in minimally biased estimates compared to the conventional zero-scoring method.
• Regression coefficients were found to be more susceptible to bias than standard errors.
• The extent of bias was influenced by the proportion of missing data and the underlying data generating mechanism.

Conclusions:

• Careful handling of incomplete composite outcome subcomponents is essential to prevent biased estimates and misleading inferences in quality of care assessments.
• The findings highlight the superiority of MI over conventional methods for missing PAQC score data.
• Further research is recommended on alternative imputation strategies at component and composite outcome levels, compatible with the substantive model.