在藥物開發中，在沒有證據證明其治療效果的情況下，將患者納入研究藥物的大規模試驗既不可取也不合乎道德。因此，通常我們會進行設計過的二期試驗，為更大規模的三期試驗選擇一個或幾個有希望的劑量組。臨床二期試驗的研究往往會面臨到隨訪時間長、樣本量大但缺乏患者資源或研究用品昂貴等困難。與傳統的藥物開發計劃相比，無縫二/三期試驗具有節省開發成本和縮短新化合物上市時間的潛力，因此受到了廣泛關注。在本文中，我們介紹了Shun (2007)與Friede (2011)兩種適應性無縫二/三期試驗的方法，都是在二期階段基於早期終點數據進行試驗組，特別考慮了在期中分析時未觀察到最終結果的情況。在這些情況下，我們考慮適當的“替代指標”做為早期終點進行期中分析。最後依據主要終點進行最終統計推論。因為期中分析會造成型一誤差的膨脹，Shun (2007)提出兩階段勝者設計，根據最終檢定統計量的分配，反求檢定在指定顯著水下的臨界值；Friede (2011)利用封閉檢定法來控制因劑量選擇所造成的型一誤差膨脹，再使用逆常態加權組合函數來結合根據兩階段的主要終點所得的p值進行統計推論。本文中透過模擬確認兩種方法是否能控制型一誤差與比較檢定力上的表現。;It is a common practice to perform well-designed phase II trials to select one or a few promising doses for larger scale phase III confirmatory trials. Certain phase II studies that aim at clinical events often face difficulties, such as long-term follow-up, large sample size but lack of patient resources, and expensive study supplies. Seamless phase II/III designs have gained much attention because of their potential to save development costs and to shorten time-to-market of a new compound compared to conventional drug development programmes with separate trials for individual phases. Two methods of adaptive seamless phase II/III trials, Shun (2007) and Friede (2011), the phase II trials are based on early endpoint data in, with special consideration given to the fact that no primary endpoint were observed in the interim analysis. In these cases, we considered appropriate "surrogate measures" as early endpoints for interim analysis. Finally, final statistical inferences were made based on the primary endpoint at two stages. Because the interim analysis will inflate the type-one error, Shun (2007) proposed a two-stage winner design, according to the distribution of the final test statistics, then find the critical value of the test under the specified significant level. In Friede (2011), closed testing procedure and weighted inverse normal combination function are used to control the inflation of overall type I error rate due to dose selection and p-values from two stage subjects at the final analysis. In this paper, the simulation is used to confirm whether the two methods can control the type-one error and compare the performance on the power.
