Statistical Inference in Diagnostic Studies: Tackling Boundaries and Imperfect Measures

这个项目是生物计量学和临床流行病学研究所和柏林慈善医院Universitätsmedizin心身医学系的合作成果。我们的共同目标是开发新的统计方法，以在诊断研究中实现精确和正确的统计推断，特别是(1)当模型参数接近其边界或(2)当诊断工具没有完美的准确性时。虽然这两个问题很常见，但经常被忽视；因此，来自诊断和流行病学研究的估计和统计推断可能存在很大的偏差，对后续试验和医疗常规产生负面影响，并有很大的风险造成研究人员和卫生政策制定者的错误信息。我们采取一般的生物计量学和流行病学疾病特定的角度来解决这两个问题。首先，我们将开发健壮的统计程序来控制名义误差水平，即使真实参数接近其边界（例如，AUC >.9），包括AUC方差的无偏估计和广泛设计的置信区间。我们将通过开发无偏方差估计器（使用u统计量）以及凸组合方法来实现这些目标，该方法将标准（近似）与新的倒分数检验相结合，用于计算置信区间。结果的质量将在广泛的模拟研究中从理论上和经验上进行调查。其次，我们将适当地解释在抑郁症的诊断和流行病学研究中使用贝叶斯潜类模型的不完善的诊断准确性。在这里，关键的挑战是开发适当的信息先验分布，适当地反映参考标准和常用诊断测试的诊断特性。我们将通过综合有关参考标准诊断准确性的现有证据，并将其纳入诊断荟萃分析，开发新的方法来优化基于不完善诊断测试的患病率估计，例如，通过避免筛查测试的二分类，以及通过调查这些方法在荟萃分析和基于人群的抑郁症患病率研究中的影响，来实现这一目标。我们将共同建立一个基本的统计推断框架，以及免费提供的实现，明确考虑到不完善的参考标准。因此，我们将使研究人员能够在广泛的应用中得出准确的统计推断，例如，小样本的诊断研究以及基于人群的研究，这些研究通常使用不完善的诊断工具，因此极大地有助于在许多不同领域的诊断和流行病学研究中准确分析数据。

This project is a collaborative effort of the Institute for Biometry and Clinical Epidemiology and the Department of Psychosomatic Medicine at Charité Universitätsmedizin Berlin. Our common objective is to develop novel statistical methods to achieve precise and correct statistical inference in diagnostic studies, in particular (1) when model parameters are either close to their boundary or (2) when diagnostic tools do not have perfect accuracy. Although common, both issues are frequently ignored; estimates and statistical inference from diagnostic and epidemiological studies can hence be substantially biased, negatively impacting subsequent trials as well as medical routine, and bearing a significant risk of misinformation of researchers and health policy makers. We take a general biometric and an epidemiological disease-specific perspective to tackle both issues. First, we will develop robust statistical procedures that control the nominal error levels even when the true parameters are close to their boundary (e.g., AUC > .9), including unbiased estimates of variance of the AUC and confidence intervals for a wide range of designs. We will achieve these goals by developing unbiased variance estimators (using U-statistics) as well as a convex combination approach, which combines standard (approximate) with newly inverted score tests, for the computation of the confidence intervals. The quality of the results will be investigated both theoretically as well as empirically in extensive simulation studies. Second, we will properly account for imperfect diagnostic accuracy using Bayesian Latent Class Models in diagnostic and epidemiological studies of depression. Here, the key challenge is to develop appropriate informative prior distributions appropriately reflecting diagnostic properties of both reference standards and commonly used diagnostic tests. We will achieve this by synthesizing the available evidence on diagnostic accuracy of reference standards and incorporating those into diagnostic meta-analysis, developing novel methods to optimize prevalence estimation based on imperfect diagnostic tests, e.g., by avoiding dichotomization of screening tests and by investigating the impact of these methods in meta-analysis and population-based studies of depression prevalence. Together, we will establish a fundamental statistical inference framework along with freely available implementations that explicitly takes imperfect reference standards into account. We will therefore enable researchers to draw accurate statistical inference in a wide range of applications, e.g., diagnostic studies with small samples as well as population-based studies which commonly use imperfect diagnostic tools, and therefore greatly contribute to the accurate analysis of data in diagnostic and epidemiological studies across many different fields.