Optimizing Trial Design Sequential, Adaptive, and Enrichment Strategies
Optimizing Trial Design Sequential, Adaptive, and Enrichment Strategies
复制标题
DOI：
10.1161/circulationaha.108.809707
复制
发表时间：
2009-02-03
期刊：
影响因子：
37.8
通讯作者：
Ware, James H.
中科院分区：
文献类型：
--
作者：
Mehta, Cyrus;Gao, Ping;Ware, James H.
检测20%风险降低所需的合并样本量（即，相对风险ρ0. 8)单侧水平-0.025检验的80%把握度，根据公式1计算为N8236。这种方法有2个重要的局限性：把握度随安慰剂组事件发生率和治疗效果的大小而变化。例如，如果πc 7%而不是8%，则在80%把握度下检测20%风险降低所需的样本量从N8236增加到N9503。此外，如果实际风险降低仅为18%，则所需样本量进一步增加至N11 841。图1显示了一项8000例患者研究的把握度，3种不同安慰剂组事件发生率和风险降低15%-22%，其中风险降低100（1ρ）%。该研究的功效从53%到93%不等。在随后的章节中，我们将重点关注中间曲线，对应于8000例患者的固定样本试验，安慰剂组事件发生率为8.7%。对于该事件发生率和相对风险ρ0。8，8000例患者产生的功效为82%。当我们根据假设的安慰剂组事件发生率和治疗效果设计一个固定样本量的试验时，我们面临的风险是，实际情况下，试验的效力不足。当终点测量为至事件发生的时间而不是事件发生率时，可以通过继续试验至预定的事件数D来消除把握度对安慰剂组事件发生率的先验知识的依赖性。单侧水平-α检验检测具有1 × 10把握度的风险比ρ所需的事件数由下式给出：
bined sample size needed to detect a 20% risk reduction (ie, a relative risk of ρ0. 8) with 80% power with a 1-sided level-0.025 test is calculated from Equation 1 as N8236. This approach has 2 important limitations: The power varies as a function of both the placebo group event rate and the magnitude of the treatment effect. For instance, if πc 7% instead of 8%, the sample size needed to detect a 20% risk reduction with 80% power increases from N8236 to N9503. If, in addition, the actual risk reduction is only 18%, the required sample size increases even further to N11 841. Figure 1 displays the power of an 8000-patient study for 3 different placebo-group event rates and risk reductions between 15% and 22%, in which the risk reduction is 100 (1ρ)%. The power of the study varies from 53% to 93%. In subsequent sections, we focus on the middle curve, corresponding to an 8000-patient fixed sample trial with an 8.7% placebo-group event rate. For this event rate and a relative risk of ρ0. 8, 8000 patients yield power of 82%. When we design a trial with a fixed sample size based on an assumed placebo-group event rate and treatment effect, we are at risk of mounting a trial that is underpowered for the actual situation. When the end point is measured as the time to the event rather than the event rate, the dependence of power on prior knowledge of the placebo-group event rate can be eliminated by continuing the trial to a prespecified number of events, D. The number of events needed for a 1-side level-α test to detect a hazard ratio ρ with 1ß power is then given by