Statistical Methods in Diagnostic Medicine and Measures of Agreement

在流行病学、统计学和预防研究部的许多研究中，诊断预测指标的发展以及评估评分者或设备之间的一致性是重要的。这些数据的特点很复杂，经常需要新的统计方法来充分处理这些分析问题。 自然驾驶研究的一个目标是预测在事故发生时或之前不良驾驶行为的青少年发生撞车或险些撞车的情况。开发和评估预测器的性能是复杂的，因为测量是在很长一段时间内实时进行的，而且一些人比其他人有更多的崩溃。此外，预测器的质量本身可能与在给定青少年身上观察到的险些相撞的次数有关(对于那些撞车次数更多的人来说，预测的质量可能会更差)。我们计划开发一种簇内重采样的方法来开发这个预测值。虽然制定预报器的方法已经确立(尽管这将是这一方法的一种新应用)，但在制定验证预报器的方法学方面还需要进一步研究。我们将研究这个问题的不同交叉验证程序，并总结我们的工作，以便制定指导方针，当事件数量具有信息性时，根据纵向数据开发和验证预测因子。 预测排卵的发生和时间是生殖流行病学的一个重要目标。BioCycle研究收集了大约250名女性两个月经周期的排卵发生和时间的信息。我们将使用多变量纵向生物标记物数据来开发这一事件的预测因子。将需要新的统计方法来估计这些预测模型以及验证它们的预测性能。 评分员之间的一致性衡量标准是针对一次评分员和一段时间内的评分员制定的。最近，Endo的研究遇到了一个尚未解决的分析问题，其中的一致性正在单调增加。具体地说，在一系列时间点中的每一个时间点，一组专家评分者都会看到一个案例，信息量不断增加。人们的兴趣是测试协议是否会随着时间的推移而增加，因为有了这些额外的信息。在该项目下，将为这一问题制定新的统计方法。

The development of diagnostic predictors as well as assessing agreement between raters or devices is important for many studies in the Division of Epidemiology, Statistics, and Prevention Research. Features of these data are complex and new statistical methodology is often required to adequately address these analytic issues. A goal in the natural driving study is to predict the occurrence of crashes or near crashes among teenagers from poor driving behavior at or before the time of the crash. Developing and assessing the performance of a predictor is complex since measurements are taken in real time over a long period, and some individuals have many more crashes than others. Also, the quality of the predictor may itself be related to the number of near crashes observed on a given teenager (the quality of the prediction may be worse for those who have more crashes). We plan on developing a within-cluster resampling approach in developing this predictor. Although the methodology for developing the predictor is established (although this would be a novel application of this methodology), further research will be needed in developing methodology for validating the predictor. We will examine different cross-validation procedures for this problem and generalize our work in order to develop guidelines for developing and validating predictors from longitudinal data when the number of events is informative. Predicting the occurrence and time of ovulation is an important goal in reproductive epidemiology. The biocycle study collects information on the occurrence and time of ovulation for two menstrual cycles on approximately 250 women. We will use multivariate longitudinally biomarker data in developing a predictor of this event. New statistical methodology will be required for estimating these predictive models as well as to validate their predictive performance. Measures of agreement between raters have been developed for raters at a single time as well as raters over time. Recently, the ENDO study encountered an unsolved analytic problem in which the agreement is monotonically increasing. Specifically, a group of expert raters are presented with a case with an increasing amount of information at each of a series of time points. Interest is on testing whether the agreement is increasing over time with this additional information. New statistical methodology will be developed for this problem under this project.