Department of Public Health and Surveillance, Sciensano, Brussels, Belgium.
Interuniversity Institute for Biostatistics and Statistical Bioinformatics, Data Science Institute, Hasselt University, Hasselt, Belgium.
Epidemiol Infect. 2020 May 4;148:e151. doi: 10.1017/S0950268820000916.
With a case-crossover design, a case's exposure during a risk period is compared to the case's exposures at referent periods. The selection of referents for this self-controlled design is determined by the referent selection strategy (RSS). Previous research mainly focused on systematic bias associated with the RSS. We additionally focused on how RSS determines the number of referents per risk, sensitivity to overdispersion and time-varying confounding.We illustrated the consequences of different RSS using a simulation study informed by data on meteorological variables and Legionnaires' disease. By randomising the events and exposure time series, we explored statistical power associated with time-stratified and fixed bidirectional RSS and their susceptibility to systematic bias and confounding bias. In addition, we investigated how a high number of events on the same date (e.g. outbreaks) affected coefficient estimation. As illustrated by our work, referent selection alone can be insufficient to control for a time-varying confounding bias. In contrast to systematic bias, confounding bias can be hard to detect. We studied potential solutions: varying the model parameters and link-function, outlier-removal and aggregating the input-data over smaller areas. Our simulation study offers a framework for researchers looking to detect and to avoid bias in case-crossover studies.
采用病例交叉设计，将病例在风险期内的暴露情况与对照期内的暴露情况进行比较。这种自我对照设计的对照选择是由对照选择策略（RSS）决定的。先前的研究主要集中在与 RSS 相关的系统偏差上。我们还关注了 RSS 如何确定每个风险的对照数量、对过离散的敏感性以及时变混杂。我们使用气象变量和军团病数据的模拟研究说明了不同 RSS 的后果。通过随机化事件和暴露时间序列，我们探讨了与时间分层和固定双向 RSS 相关的统计功效，以及它们对系统偏差和混杂偏差的敏感性。此外，我们还研究了同一天内大量事件（例如暴发）如何影响系数估计。正如我们的工作所表明的，仅参考选择本身可能不足以控制时变混杂偏差。与系统偏差不同，混杂偏差很难被发现。我们研究了潜在的解决方案：改变模型参数和链接函数、剔除异常值以及在较小的区域上汇总输入数据。我们的模拟研究为研究人员提供了一个框架，用于检测和避免病例交叉研究中的偏差。
