United States Renal Data System, Minneapolis Medical Research Foundation, Minneapolis, Minnesota 55404, USA.
Kidney Int. 2012 Feb;81(4):341-50. doi: 10.1038/ki.2011.388. Epub 2011 Nov 16.
In observational studies, treatment is often time dependent. Mishandling the time from the beginning of follow-up to treatment initiation can result in bias known as immortal time bias. Nephrology researchers who conduct observational research must be aware of how immortal time bias can be introduced into analyses. We review immortal time bias issues in time-to-event analyses in the biomedical literature and give examples from the nephrology literature. We also use simulations to quantify the bias in different methods of mishandling immortal time; intuitively explain how bias is introduced when immortal time is mishandled; raise issues regarding unadjusted treatment comparison, patient characteristics comparison, and confounder adjustment; and, using data from DaVita Inc., linked with the Centers for Medicare & Medicaid Services end-stage renal disease database, show that the severity of bias and the issues described can occur in actual data analyses of patients with end-stage renal disease. In the simulation examples, mishandling immortal time led to an underestimated hazard ratio (treatment vs. control), thus an overestimated treatment effect, by as much as 96%, and an overestimated hazard ratio by as much as 138%, depending on the distribution of 'survival' time and the method used. Results from the DaVita data were consistent with the simulation. Careful consideration of methodology is needed in observational analyses with time-dependent treatment.
在观察性研究中，治疗通常是时间依赖性的。如果在开始随访到开始治疗的时间处理不当，可能会导致一种称为不朽时间偏差的偏差。进行观察性研究的肾脏病学研究人员必须意识到如何在分析中引入不朽时间偏差。我们回顾了生物医学文献中事件时间分析中的不朽时间偏差问题，并从肾脏病学文献中举例说明。我们还使用模拟来量化不同处理不朽时间不当方法的偏差；直观地解释了当不朽时间处理不当时如何引入偏差；提出关于未调整治疗比较、患者特征比较和混杂因素调整的问题；并使用来自 DaVita Inc. 的数据，与医疗保险和医疗补助服务中心的终末期肾病数据库相关联，表明严重程度偏差和描述的问题可能会出现在终末期肾病患者的实际数据分析中。在模拟示例中，处理不朽时间不当会导致风险比（治疗与对照）低估，从而治疗效果高估，最大可达 96%，风险比高估最大可达 138%，具体取决于“生存”时间的分布和使用的方法。DaVita 数据的结果与模拟结果一致。在时间依赖性治疗的观察性分析中需要仔细考虑方法学。
