Meta-analysis of the validity of binary diagnoses based on dichotomized cirteria

在项目的第一阶段，开发了两个新的诊断研究元分析模型，明确地纳入了临界值问题。第一个模型是比例风险模型（PHM），基于SROC曲线的lehman族，而第二个模型使用对数转换的敏感性和特异性的二元正态分布。在该项目的第二阶段，这些模型被扩展到允许协变量信息，以解释随机效应模型中的异质性。在模型中引入了参数化的连杆函数。因此，这种方法被嵌入到更大的一类具有相当灵活的SROC曲线的模型中。为了使这些方法易于为更大的用户社区所用，开发了r包mada。迄今为止开发的模型允许分析观察到的异质性（通过协变量）以及未观察到的异质性（通过随机效应模型）。在第三阶段，有一年的时间，未观察到的异质性将进一步量化。正如所收集的荟萃分析的多样性所显示的那样，几乎所有可用的诊断研究的荟萃分析都显示出某种形式的未观察到的异质性。然而，在不同的荟萃分析中，异质性的强度差异很大。传统上，异质性的大小是通过异质性测量I2来测量的，这是由Higgins和Thompson（2002）提出的。然而，这个数量并不是一个经典的效应测量，因为它不是独立于单个研究的样本量。在这里，一个改进的异质性测量建议用于诊断荟萃分析，它不受上述限制。这种新的异质性度量的发展是通过项目第二阶段的发现而出现的。此外，还将通过分析和模拟研究来研究这种新的异质性测量的性质和质量。

In the first phase of the project two new meta-analytic models for diagnostic studies were developed that explicitly incorporated the cut-off value problem. The first model, a proportional hazards model (PHM), is based upon the Lehmann-family for SROC curves, whereas the second model uses a bivariate normal distribution of logarithmically transformed sensitivities and specificities. In the second phase of this project, these models were extended to allow for covariate information, to explain heterogeneity in the random effects model. Furthermore, a parametric link function was included in the PHM. Hence this approach is embedded in a larger class of models with rather flexible SROC curves. To make these methodologies readily available for a larger user community the R-package mada has been developed.The models developed so far allow the analysis of observed heterogeneity (by means of covariates) as well as unobserved heterogeneity (by means of random effect models). In this third phase, for which one year is available, unobserved heterogeneity will be further quantified. As the diversity of collected meta-analyses shows, almost every available meta-analysis of diagnostic studies shows some form of unobserved heterogeneity. However, the strength of heterogeneity varies considerably across the various meta-analyses. Conventionally, the magnitude of heterogeneity is measured by means of the heterogeneity measure I2, suggested by Higgins und Thompson (2002). This quantity, however, is not a classical effect measure as it is not independent of the sample sizes of the individual studies. Here, a modified heterogeneity measure is suggested for diagnostic meta-analyses, which does not suffer under the above limitation. The development of this new heterogeneity measure has emerged through the findings of the second phase of the project. Furthermore, the properties and qualities of this new heterogeneity measure will be investigated analytically and by means of simulation studies.