Online Program

Return to main conference page
Wednesday, September 25
Wed, Sep 25, 9:45 AM - 10:30 AM
Marriott Foyer
Poster Session

A Comparison of Multiple Imputations at Item-Level and Score-Level in Missing Patient-Reported Outcomes: A Case Study on EQ-5D-5L Data Collected in a Randomized Clinical Trial (300904)

*Leiya Han, PPD, Inc. 

Keywords: missing data, multiple imputation, patient-reported outcomes, EQ-5D-5L, randomization clinical trial

Background: Missing data is a major challenge to the success and interpretation of any clinical trial.Previous studies showed that missing data in multi-item questionnaires should be handled by multiple imputation. A simulation study suggested that performing item-level imputation provides meaningful power advantages over scale-level imputation. This study aims to compare utility index of EQ-5D-5L scored with imputed item-level responses for missingness and directly imputed score-level utility index against the utility index scored with completed non-missing data collected in a randomization control trial (RCT). Methods: A subset of 79 subjects who had complete responses of EQ-5D-5L questionnaire at all three protocol scheduled study visits (1, 2, and 3) was selected for this study. Item-level missing responses with a monotone pattern at Visits 2 and 3, or 3 were simulated based on missing at random (MAR) assumption. Monotone logistic regression and predictive mean matching (PMM) regression were used for item-level imputation and score-level utility index imputation. The England value set of EQ-5D-5L was used for scoring the utility index. Performance of utility index scored from imputation with that of completed data was assessed using 1) utility index range 2) root mean square error and 3) pooling difference in least squares means with standard error from analysis of covariance model. Results and Conclusion: All item-level imputations are within original response categorical scale and imputed utility index scores are all having the same range of completed data. Overall, the score-level imputations have relative smaller RMSEs. In term of pooling difference in LS means, performance of score-level imputation shows better in data with 5%, 10%, and 40% missing proportions, whilst item-level imputation performs better in data with 20% and 30% missing. Between two multiple imputation strategies, the study shows that none has significant advantage over the other.