Comparative Study On Different α Spending Strategies In Adaptive Design

With the increasing risk and cost of new drug research and development,the focus of clinical research all over the world is to seek innovative trial design and statistical analysis methods.As a new design,the idea of adaptive design further broadens the space for the development of drug clinical trials.At the same time,due to its high requirements in various aspects,it also brings greater statistical challenges to researchers.Therefore,it is often necessary to carefully plan the trial design and statistical analysis methods in advance in practical applications.The main research contents of this paper are as follows:Firstly,based on the O’Brien-Fleming spending function method and the inverse normal combined p-value method,three alpha spending strategies were proposed to control the familywise type Ⅰ error rate(FWER)caused by the change of sample size.Taking adaptive two-stage double-arm clinical trial as an example,the nominal significance levels for the first interim analysis of three alpha spending strategies are all determined by the O’BrienFleming spending function method.But the nominal significance levels for the final analysis are different.Among them,strategy one and strategy three continue to use the nominal significance levels of the final analysis initially determined by the O’Brien-Fleming spending function method.While strategy two uses the nominal significance levels adjusted according to the changing information time.In addition,there are differences in p-value calculations among alpha spending strategies.Among them,strategy one uses the inverse normal combined p-value method to determine the test p-value of each stage.While strategy two and strategy three combine the data of each stage to calculate the test p-value of each stage.Secondly,three alpha spending strategies were applied to the sample size re-estimation design based on nuisance parameter/treatment effect re-estimation.And three alpha spending strategies were compared by using Monte Carlo simulation to evaluate their statistical characteristics.Finally,three alpha spending strategies were applied to adaptive enrichment design.Under the condition that the sample size was redistributed and re-estimated in the second stage,three alpha spending strategies were compared by using Monte Carlo simulation to evaluate their statistical characteristics.The main research conclusions of this paper are as follows:When the standard deviation under normal data was re-estimated based on sample size re-estimation design,it can be found that the type Ⅰ error rate control ability of each alpha consumption strategy was basically the same.In terms of the power,it can be found that the power loss of the first strategy was the largest,the second strategy was the second,and the third strategy achieved the highest power.When treatment effect under normal data and binary data was re-estimated,the results showed that the type Ⅰ error rate control ability of three alpha spending strategies was decreased in turn,and the power of achieved was increased in turn.When analyzing the enrichment subgroups of the second stage under normal data based on adaptive enrichment design.In terms of the type Ⅰ error rate control,it was found that the type Ⅰ errors rate of strategy two and strategy three had no obvious relationship with the positive rate of the subgroups.The type Ⅰ errors rate of strategy one had obvious relationship with the positive rate of the subgroups.The type Ⅰ errors rate of strategy one decreased with the increase of the subgroup positive rate.But the type Ⅰ error rate of strategy one can swell to more than 10%,so it can’t effectively control the type Ⅰ error rate.The type Ⅰ error rate control ability of strategy two was better than strategy three,but the type Ⅰ error rate of strategy two and strategy three generally fluctuated around 0.025(one-sided test),which was within the acceptable range.In terms of the power,it can be found that the power of each alpha spending strategy increased with the increase of the subgroup positive rate.Except for strategy one,when the subgroup positive rate was low,the power can’t reach the expected value,and the others could reach 80%.Finally,through the analysis of all the simulation trial results in this paper,strategy two is a better choice.