Keywords: missing data, imputation, AUC1
For dry powder inhalers, clinical endpoint studies are designed to evaluate the bioequivalence between a generic product and its reference listed product, and to show the superiority of active treatment groups over placebo. One of the primary endpoints, recommended by FDA’s product specific guidance, is the area under the serial FEV1 – time curve calculated from time 0 to 12 hours (AUC0-12) on the first day of the treatment. The FEV1 on Day 1 are measured at 10 time points, i.e., 0, 0.5, 1, 2, 3, 4, 6, 8, 10 and 12-hour post-dose. The applicants proposed different methods to impute the missing data. However, it is not clear how missing data affects the validity of AUC analysis and the performance of different imputation methods has not been investigated yet.
In this project, we evaluate the impact of different imputation methods on equivalence test using simulation. Based on the clinical data submitted in ANDAs, we generate the missing data sets based on MCAR, MAR, MNAR and missing at the end. Different imputation methods are applied to the missing datasets to generate the imputed datasets. We generate 10,000 simulated samples of n=1000 subjects (500 subjects in the test group and 500 subjects in the reference group). The performance of different imputation methods is evaluated using four measures: 1) absolute bias 2) empirical standard error 3) percentage of 90% confidence interval within [0.80, 1.25], and 4) the coverage probability of 90% confidence interval containing the true parameter value. The simulation results show that the LOCF method performs well in all four measures compared to other methods Line, Regression and multiple imputation methods are comparable with LOCF. All four methods perform better than the methods applicants proposed.