Research On The Application Of Generalized Linear Models In Clinical Trials Of Medical Devices

Background:The longitudinal data of quantitative indicators were widely used in clinical trials of medical devices.Correlation existed because the data within clusters or subjects had the same attributes,did not meet the independent hypothesis,and it was relevant data.Traditional analytical methods such as group t-test,ANOVA analysis and general linear model,which required data to be independent,ignored the correlation of data within clusters or subjects,could underestimate the standard error of parameters,and were prone to false positive errors,and might result in approving invalid devices for market.Generalized estimating equations and mixed effects model were two different analysis methods for the analysis of the correlated data,which were developed on the basis of generalized linear model.The former was the overall average model,the latter was a specific individual model.The different methods of parameter estimation also could lead to different conclusions,and existing study had reported that the standard error of parameter estimation was smaller for generalized estimating equations.In addition,traditional statistical analysis methods,generalized estimation equation and mixed effect model had been applied to the longitudinal data analysis of quantitative indicators in clinical trials of medical devices now.The application of statistical analysis methods were still different.How to select appropriate analytical methods,how to effectively and fully analyze the longitudinal data and correctly evaluate the efficacy of medical devices were urgent questions to be solved and was also a concern of the national regulatory authority.Objective:By random simulation method,under different sample size schemes,the combination of different data correlation degree and different proportion was considered to simulate the correlation data,we comprehensively compared the estimation effects of generalized linear model,generalized estimation equation and mixed effect model in longitudinal data of quantitative indicators in trial of medical devices,and combining with case analysis to explore appropriate statistical analysis methods and provide reference.Methods:Taking non-inferiority stent trial design as an example and combining with the application of Monte Carlo random simulation method,four different sample sizes corresponding to the design of non-inferiority trial were constructed by setting different non-inferiority margin values,different power levels and different standard deviations.The sample size of experimental group and control group were 100,120,140 and 200,respectively.Under each sample size scheme,the correlation degree of data were set as uncorrelated,low correlated,moderate correlated and high correlated,respectively.The proportion of correlated data was set as 0.05,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9 and 1,respectively,a total of 11 cases.Relevant data of simulation study was generated,and each case was simulated 1000 times.The actual powers of the general linear model,the generalized estimation equation and the mixed effect model when the positive results were obtained were calculated respectively,and the accuracy of the application of the model was evaluated by examining whether it was consistent with the pre-specified power degree set in the study design and comparing the false positive error rate between different methods.Among them,the generalized estimation equation adopted independent and exchangeable intra-group correlation structure for analysis,which were expressed by generalized estimation equation(independent)and generalized estimation equation(exchangeable),respectively.Results:Simulation results showed that when data was uncorrelated,low,moderate and high uncorrelated,the actual power of the general linear model,generalized estimation equation(exchangeable)and mixed effect model showed a trend of increasing gradually as the proportion of correlated data increased,but the latter two methods were both smaller than the general linear model.When the proportion of relevant data was less than or equal to 5%,the power obtained by the generalized estimation equation was closer to the pre-specified power level.When the proportion of relevant data was more than 30%,the power obtained by mixed effect model was closer.When the proportion of relevant data was 5%～30%,which method was closer to the pre-specified degree of power,the conclusion would be different depending on the degree of data correlation.When data was uncorrelated,low and moderate uncorrelated,the actual power of the generalized estimation equation(independent)was not much different from that of the generalized estimation equation(exchangeable),but the power obtained by the generalized estimation equation(independent)tended to decrease first and then increase,and was lower than pre-specified power level when the data was highly correlated.The false positive error rate obtained by general linear model showed a trend of increasing gradually with the proportion of relevant data increased and was higher than the significance level of 0.05.Compared with the general linear model,the false positive error rates obtained by generalized estimation equation and mixed effect model both were relatively smaller and stable around the significance level of 0.05.Among them,the false positive error rates obtained by generalized estimation equation(independent)and generalized estimation equation(exchangeable)were not much different,and obtained by mixed effect model was the lowest.Conclusion:In the analysis of longitudinal data of quantitative indicators in clinical trials of medical devices,traditional statistical methods should not be used.The correlation of data within cluseters and subjects in the longitudinal data should be considered,and the generalized estimation equation and mixed effect model should be used for analysis.When sample size was designed based on two independent samples of quantitative indicator,whereas actual data were longitudinal correlation data,suggest to adpot generalized estimation equation(exchageable)when the proportion of relevant data was less than or equal to 5%,and adpot mixed effect model when the proportion of relevant data was more than 30%.When the proportion of relevant data was 5%～30%,the appropriate analysis method should be selected to analyze the longitudinal data according to the characteristics and correlation degree of the actual data.Ensure the accurate estimation of the true efficacy between groups,objectively evaluate the effect of products,avoid approving invalid devices for market,and provide reference for the ational regulatory authority.