Biostatistics: Design and Analysis of Experimental Studies; and Evaluating Classifier Tests.

点击翻译按钮获取中文摘要

This project will address estimation problems in two related areas of biostatistics: Design and Analysis of Experimental Studies, and Evaluating Classifier Tests. Both topics require the development of innovative statistical models.****Topic #1: In this initiative, the applicant is considering the impact of preference effects in randomised experiments. Study outcomes may be affected by 1) preferences held by participants for the alternative interventions (inducing a selection effect); and 2) whether participants actually receive their preferred intervention (a preference effect). These effects cannot be estimated in a conventional experiment (where only the intervention effect is identifiable), but one can consider an alternative the two-stage randomised design. The applicant has developed an analytic model that uses two-stage data from participants who are either randomised or who choose their intervention, in order to estimate all these effects, and has shown how design efficiency can be maximised. Current objectives are to enhance the two-stage design, to allow for participants who cannot decide between interventions. Longer term objectives include optimising design efficiency, and quantifying the required sample size and study power for the enhanced design.***Topic #2: While experimental designs are sometimes used, evaluation of classifier tests often requires other approaches. Classifiers are used to distinguish items with or without some attribute. There is often no error-free test available, so an imperfect reference standard must be used instead, subject to error. The applicant has developed latent class models to assess classifier accuracy in this situation, including models to evaluate tests for genetic polymorphisms, and to compare international biochemistry labs.****More recently the applicant has considered the impact of screening on prostate cancer mortality. A European study has shown a 21% mortality reduction, but a similar study in the USA showed no benefit. This has led to inconsistencies in recommendations about screening, which is now of major scientific interest. Adjudication of causes of death is subjective, and there is no gold standard. The European investigators therefore challenged the applicant to formulate an innovative model to assess adjudicator accuracy, and subsequently to correct the study results for bias due to measurement error. A longer term objective is to develop methods to estimate the rate of screening “over-detection”, this being the identification of non-lethal cancers. ***Both of these areas have been investigated by the applicant for some time, and form part of his long term research agenda. They are also very suitable areas for graduate students in statistics or biostatistics, disciplines for which there is currently a shortage of HQP in Canada. As such, training of HQP in these areas will be addressing a clear need in the Canadian context. ******