Conference Program

Missing Domain Scores of SF-36 Data in Clinical Trial: Half-Scale Imputation Versus Multiple Imputation (303639)

*Leiya Han, PPD, part of Thermo Fisher Scientific 

Keywords: missing data analysis, half-scale rule, multiple imputation, SF-36v2

Background: Half-scale rule is commonly used to score a domain while missing items exist in the domain. Multiple imputation (MI) accounts for uncertainty in missing data and hence enhances robustness in analyses in the presence of missing data. Research behind this poster evaluates how listwise deletion, half-scale and multiple imputations perform on handling missing scores compared to complete data on physical function (PF) domain of SF-36v2 collected in a clinical trial. Data and methods: A complete SF-36v2 dataset from a phase 3 trial including 282 and 92 subjects in Active or placebo group with available data from Baseline, Weeks 16 and 24. To create monotone missing data as missing at random, 5 or less out of 10 items within PF domain were removed in 10% and 30% of randomly selected subjects based on propensity score of missingness in 500 simulations. Half-scale and multiple imputations were implemented on missing data in each simulation. Each simulated dataset was imputed 100 times using a monotone regression for MI. The estimates from ANCOVA with listwise deletion, half-scale, or multiple imputation data against those with complete date at Week 24 were evaluated using RMSE, raw bias, percent bias, rate of percent bias =5%, and coverage rate. Results and conclusion: The estimates from ANCOVA, compared to those from complete data, half-scale imputation has the smallest RMSE, raw bias, and percent bias, as well as the narrowest 95% CI and the highest rate of percent bias =5% for all levels of missingness. Coverage rates for estimates are at 100% in all 3 sets of data when 10% of subjects have missing data and it drops down to 99.8% when using the listwise deletion at 30% missingness level. The research suggests that the imputation using “half-scale rule” should be applied first to score a missing domain and multiple imputation may be used for where a “half-scale rule” is not applicable with higher levels of missing item data.