In practice, patient response is often measured some time after treatment commences. If data are analysed group sequentially, there will be subjects in the pipeline at each interim analysis who have started treatment but have yet to respond. If the stopping rule is satisfied, data will continue to accrue as these pipeline subjects respond. Standard designs stipulate that the overrun data be excluded from any analysis. However, their inclusion may be mandatory if trial results are to be included in a filing for regulatory approval. Methods have been proposed to provide a more complete treatment of the pipeline data, for example Whitehead (1992) and Faldum & Hommel(2007), although several issues remain unresolved. The work presented in this thesis provides a complete framework for dealing systematically with delayed responses in a group sequential setting. We formulate designs providing a proper treatment of the pipeline data which can be planned ahead of time. Optimal versions are used to assess the benefits for early stopping of group sequential analysis when there is a delay in response. Our new tests still deliver substantial benefits when the delay in response is small. While these fall as the delay increases, incorporating data on a highly correlated short-term endpoint is found to be effective at recouping many of these losses. P-values and confidence intervals for on termination of our delayed response tests are derived. We also extend our methodology to formulate user-friendly error spending tests for delayed responses which can deal with unpredictable sequences of information. Survival data are a special type of delayed response, where the length of delay is random and of primary interest. Deriving optimal survival tests, we conclude that tests minimising expected sample size for “standard” data are also highly efficient survival trials, achieving a rapid expected time to a conclusion.
|Date of Award||1 Nov 2008|
|Supervisor||Christopher Jennison (Supervisor)|