Biostatistical issues in the design and analysis of experimental and observational studies, and in evaluating test accuracy

本研究涉及观察性研究和随机实验中的原理和应用。它是由重要的实际问题推动的，其结果应广泛应用于卫生、自然科学和社会科学的研究。主要主题包括： 1.1识别参与者偏好的实验设计：在常规实验中，所有参与者都是随机的。已经提出了替代设计，允许一些人选择他们的干预。然后，适当的数据分析可以估计1)治疗的效果，2)参与者的理想治疗，3)首选的治疗是否是接受的治疗。后两种效应可能很重要，但在传统实验中是不可估量的。我们将研究替代设计的优化；效率；样本大小；功率；以及实际实施。 1.2测量误差：我们早些时候证明，在平衡随机实验中，对基线变量中的测量误差进行校正会导致治疗效果中的偏差。相比之下，在观察性研究中，直觉表明，未能纠正测量误差可能会产生偏差。比较这些情况，我们推断，基线差异必须有一个阈值，定义何时值得进行修正。我们建议比较治疗效果的校正和未校正估计值，定义决策阈值，并提供何时校正测量误差的实用建议。 2.测试准确度的评估通常没有金标准可用于评估新的分类器(如诊断或筛查测试)，因此必须替换不完美的参考标准。目前还不清楚这可能会对评估新测试产生什么影响。我们将调查当参考标准不完美时，表观测试精度可能会如何产生偏差。

This research concerns principles and applications in observational studies and randomised experiments. It is motivated by important practical questions, and the results should be widely applicable in research in health, and natural and social sciences. The major themes will be: 1.1 Experimental designs recognising participant preferences: In conventional experiments, all participants are randomised. Alternative designs have been proposed to allow some people to choose their intervention. Suitable data analysis can then estimate the effects of 1) treatment, 2) the participant's ideal treatment, and 3) whether the preferred treatment is the one received. The two latter effects are potentially important, but they are not estimable in conventional experiments. We will investigate optimization of the alternative designs; efficiency; sample size; power; and practical implementation. 1.2 Measurement error: We demonstrated earlier that in balanced randomised experiments, correction for measurement error in a baseline variable induces bias in the treatment effect. In contrast, in observational studies intuition suggests that failing to correct for measurement error may yield bias. Comparing these scenarios, we infer that there must be a threshold on the baseline difference, defining when correction is worthwhile. We propose to compare the corrected and uncorrected estimators of treatment effect, define the decision threshold, and provide practical advice about when to correct for measurement error. 2. Evaluation of test accuracy There is often no gold standard available for evaluating a new classifier (such as a diagnostic or screening test), so an imperfect reference standard must be substituted. It is not clear what impact this may have on assessing the new test. We will investigate how the apparent test accuracy may be biased when the reference standard is imperfect.