Comparative Study Of Sensitivity Analysis Methods Under Non-random Censoring

Censoring is a unique definition in survival analysis,its essence is missing data.The occurrence of missing data will lead to a decrease in the advantage of randomized trials,a decrease in test efficiency,and bias in trial results,so it’s of great significance to handle missing data reasonably for clinical trials.In clinical trials with survival data as the endpoint,censoring at random(CAR)is usually used as the hypothesis of the primary analysis method,and when the possibility of nonrandom censoring(CNAR)cannot be ruled out,sensitivity analysis based on the CNAR hypothesis is required to evaluate the robustness of the primary analysis conclusions.Pattern mixture model(PMM)is a commonly used sensitivity analysis method,by setting the relationship between censored data and observed data for analysis of CNAR data,and has been widely studied and applied because it’s of easy interpretation theoretically and clinical significance.The PMM framework combined with multiple imputation(MI)has developed a variety of censored data processing methods,Kaplan-Meier multiple imputation(KMMI),Cox model multiple imputation(COXMI),Piecewise exponential model multiple imputation(PCEMI)based on δ-adjusted method will be applied to clinical trials with survival data as the endpoint index in this paper,and the application of tipping point analysis method in survival data is realized by adjusting the value of delta.This paper will compare the statistical performance of δ-adjusted method,including KMMI,COXMI,PCEMI,and reference-based method,including Jump to reference(J2R),Last hazard carried forward(LHCF),Copy increments in reference hazard(CIR)under different sample sizes and different censoring rates,so as to provide a basis for the selection of sensitivity analysis methods under CNAR in clinical trials.The simulation method proposed by Bender was used to generate data sets obeying the Weibull distribution under different sample sizes and different censoring ratios,the primary analysis method is Cox regression based on observation data under CAR hypothesis,the sensitivity analysis method was based on the CNAR hypothesis and handle the censored data with δ-adjusted methods,including KMMI,COXMI and PCEMI,and reference-based methods,including J2R,LHCF and CIR.The statistical performance of the sensitivity analysis methods were evaluated with the hazard ratio(HR),its 95%confidence interval width and estimated standard error,and the data set from the German breast cancer research group was used to conduct a case study to verify the reliability of the simulation conclusions.The reference-based J2R,LHCF and CIR methods have good statistical performance under different datasets with different sample sizes and different censoring ratios,and the obtained HR and 95%confidence interval widths are relatively stable,and the estimation standard error can be well controlled,but it should be noted that the reference-based J2R method will have too conservative results when the censoring ratio is large.δ-adjusted KMMI has better statistical performance when the censoring ratio is small,but poor statistical performance when the censoring proportion is large.δ-adjusted COXMI has good statistical performance when the sample size is large and the censoring proportion is small,and the statistical performance is poor under other feature datasets.δ-adjusted PCEMI cannot control the estimation standard error well when the sample size is small and the censoring ratio is large,and has good statistical performance under other feature datasets.Case studies based on the Breast Cancer Research Group’s dataset have drawn conclusions consistent with simulation studies.It is of great significance to select appropriate sensitivity analysis methods according to different feature datasets to evaluate the robustness of the primary analysis conclusions,among which KMMI,COXMI and PCEMI based on δ-adjusted realize the application of tipping point analysis methods in survival data,which is more flexible than the reference-based method,and the deviation from CAR hypothesis to CNAR hypothesis can be realized by adjusting the δ value.