Imputing physical health status scores missing owing to mortality - Results of a simulation comparing multiple techniques

BACKGROUND. Having missing data complicates the statistical analysis of hea lth-related quality-of-life (HRQOL) data and, depending on the extent and n ature of missing data, can introduce significant bias in treatment comparis ons. OBJECTIVE. We evaluated the bias associated with 4 different imputation met hods for estimating physical health status (PHS) scores missing as a result of mortality. METHODS. A simulation study was conducted in which we systematically varied mortality rates from 0% to 30% and change in PHS scores from -20 to 20 on a 100-point scale for a 2-group clinical trial with follow-up over 18 month s. The 4 imputation methods were last value carried forward (LVCF), arbitra ry substitution (ARBSUB), empirical Bayes (BAYES), and within-subject model ing (WSMOD). Pseudo-root mean square residuals (RMSRs) and differences betw een true and estimated slopes were used to evaluate how well the imputation methods reproduced the true characteristics of the simulated population da ta. RESULTS. ARBSUB and BAYES methods have the smallest RMSRs compared with LVC F and WSMOD across all mortality rates. As the rate of :missing data result ing from mortality increased, all imputation techniques deviated more from population data. The BAYES technique was best at reproducing group slopes i n cases with differential mortality rates or when mortality rates exceeded 15%. WSMOD and LVCF significantly underestimated changes hi PHS. CONCLUSIONS. The different imputation methods produced comparable results w hen there were few missing data. The BAYES approach most closely estimated true population differences and change in PHS regardless of missing data ra tes:. These findings are limited to physical health and functioning measure s.