Compairng Different Methods In Non-inferiority Trials With Missing Data

Objective Three issues will be discussed in this study:1) to explore proper statisticalmethods for non-missing repeated measurement data in non-inferiority clinical trials.2)to explore the most efficient method for data with different types of missingmechanisms.3) to test the difference among these statistical methods that based onLOCF imputation in both superiority and non-inferiority clinical trials.Methods Complete datasets, datasets with different missing mechanisms and missingproportions were simulated using Monte Carlo simulation methods according to thespecific indexes in the clinical trial. Different statistical analysis methods (t Test, tTest-OC, t Test-LOCF, ANCOVA, ANCOVA-OC, ANCOVA–LOCF, MMRM and GEE)were used to the simulated data. The efficacies of different methods were evaluated withthe evaluation index of Power, Type1Error and so on.Results All the methods besides MMRM-AR(1) got around85%for test power andaround2.5%for Type1Error in complete repeated measures data. Similar results wereobtained in simulation with different missing mechanisms and missing proportions thatthe Type1error obtained by MMRM-AR1reached more than8.9%. Besides, the LOCFalso got a higher Type1error and underestimate the within group difference. The extentof underestimation becomes more severe as with the increase of the missing proportion.The test powers of all the methods decrease more or less with the increase of themissing proportion. However, the MMRM and GEE method with unstructured covariance matrix tend to turn down more slowly than t Test-OC and ANCOVA-OC.The MMRM and GEE method with unstructured covariance matrix and GEE withfirst-order autoregressive covariance matrix have better performance than methods withLOCF.Conclusion The MMRM model with specific covariance are not recommended, whenthe covariance structure is not known in non-missing non-inferiority clinical trial data.All the other methods besides MMRM-AR(1) model applied in the study have excellentproperties in this situation. In non-inferiority clinical trials and datasets with differentmissing mechanisms (MCAR or MAR), LOCF become not conservative, as itunder-estimates the within group difference and result into the inflation of Type1Error.LOCF is not recommended in non-inferiority clinical trials with missing data. TheMMRM and GEE methods with unstructured covariance matrix and GEE withfirst-order autoregressive covariance matrix have better performance than others nomatter what the missing proportion is. Therefore, these two methods are recommendedwhen the missing proportion is not very large.