The Evaluation Of Statistical Methods On Adaptive Seamless Phase â…¡/â…¢ Clinical Trial Using Short-term Outcome For Treatment Selection

The seamless phase â…¡/â…¢ design is one of the most attractive adaptive designs,which is a combination of traditional phase â…¡ and phase â…¢ trials, aiming at achievingthe primary objectives normally achieved through the conduct of separate phase â…¡ andphases â…¢ trials, and it would use data from patients enrolled before and after theadaptation in the final analysis. Seamless phase â…¡/â…¢ designs allow strong control ofthe family-wise type I error rate when the most promising one of a number ofexperimental treatments is selected at an interim analysis to continue along with thecontrol treatment. If the primary endpoint is observed only after long-term follow-upit may be desirable to use correlated short-term outcome data at the interim analysis toconduct the treatment selection. Therefore in an adaptive seamless phase â…¡/â…¢ design,the treatments are compared at the end of stage1（exploratory stage） based on anshort-term outcome measure, and then one or more of the experimental treatments areselected to continue into stage2（confirmatory stage） of the trial for further testingagainst the control treatment. Finally the confirmatory testing is exclusively based onthe long-term outcome, which is the primary endpoint of the trial.In this research, we focused on evaluation of three statistical methods onseamless phase â…¡/â…¢ clinical trial using short-term outcome in early stage fortreatment selection, and long-term outcome in later stage for confirming the effect ofthe most effectiveness treatment. The basic principle of seamless design and threestatistical methods for integrating the results of2stages were briefly introduced.Simulation studies were then used to understand the power and the type I error ofthese methods in different scenarios. The main research contents are as follows: 1. To explore the influence of the correlation of short-term and long-termoutcomes on the family-wise type I error rate:We simulated a study starting with3treatment groups and1parallel controlgroup with equal sample size in each group.The effective rate was set as the primaryendpoint. According to a one-sided significant level of0.025and a power of80%, Nindividuals were randomly allocated to each group. In simulation studies, we denotedθ_i as the effect of the long-term outcome,θ_i^*as the effect of the short-term outcome,and as the correlation of the long-term and short term outcomes. An interim analysiswas planned at1/2information time. First of all, we calculated the family-wise type Ierror rate of each group when assuming the global null hypothesisθ_i=0andθ_i^*=0.Secondly, the simulations were conducted under the global null hypothesisθ_i=0anddifferent values ofθ₁^*, we investigated the influence of different scenarios of andθ₁^*on the family-wise type I error rate for each method.2. To examine the testing power of3methods:Firstly, we estimated the sample size according to a one-sided alpha0.025leveland a power of80%and other different parameters. We performed an interim analysisat1/2information time, and this was repeated10,000times. The power was estimatedunder different values of the correlation coefficient of short-term and long-termoutcomes; secondly, by re-allocating the sample size, we compared these threemethods with each other according to the alteration of power. Thirdly, we kept thecorrelation of short-term and long-term outcomes constantly, and set up differenteffects of the treatmentsθ_i^*and θ_i to find out the differences of the3statisticalmethods.3. Compare three methods using quantitative data:The procedure was similar to the qualitative data. We supposed that data wasnormal distributed with variance known, and evaluated the change of the type I errorand the power under different correlation.The main conclusions of this research are as follows: When choosing some short-term outcomes as the surrogates of long-termoutcomes in an adaptive seamless phase â…¡/â…¢ design, the type I error of each methodwould be well controlled at the nominal level and as decreases, the method becomesmore and more conservative. The power is optimal when the short-term and long-termoutcomes are highly correlated and have consistent effects; otherwise it will be low.