Segmented Regression and Difference-in-Difference Methods: Assessing the Impact of Systemic Changes in Health Care
Segmented Regression and Difference-in-Difference Methods: Assessing the Impact of Systemic Changes in Health Care
复制标题
DOI：
10.1213/ane.0000000000004153
复制
发表时间：
2019-08-01
影响因子：
5.7
通讯作者：
Sessler, Daniel I.
中科院分区：
文献类型：
--
作者：
Mascha, Edward J.;Sessler, Daniel I.
围手术期的研究者和专业人士越来越多地寻求评估是否实施系统的做法改变改善结果相比，以前的例行。整群随机试验是评估系统性实践变化的最佳设计，但往往不切实际;因此，研究人员经常选择前后设计。在这轮统计大循环中，我们首先讨论了前后设计中固有的偏差，包括由于时间段完全分开而导致的混杂、回归均值、霍桑效应等。许多这些偏见可以至少部分地解决通过使用适当的设计和分析，我们讨论。我们的重点是分段回归中断的时间序列，这不需要一个并发的控制组，我们还提出了替代设计，包括差异中的差异，阶梯楔形，和集群随机化。要很好地进行分段回归，需要在每个周期内有足够数量的时间点，沿着一组强大的潜在混淆变量。该方法比较干预前后随时间的变化、干预开始时结果的差异以及有干预措施时观察到的趋势与无干预措施时预测的趋势。如果做得好，所讨论的方法允许对干预的效果进行强有力的推断，尽管仍然需要假设并具有局限性。方法证明使用中断的时间序列研究中，麻醉医生负责从内科医生的成人医疗急救队，试图改善结果。
Perioperative investigators and professionals increasingly seek to evaluate whether implementing systematic practice changes improves outcomes compared to a previous routine. Cluster randomized trials are the optimal design to assess a systematic practice change but are often impractical; investigators, therefore, often select a before-after design. In this Statistical Grand Rounds, we first discuss biases inherent in a before-after design, including confounding due to periods being completely separated by time, regression to the mean, the Hawthorne effect, and others. Many of these biases can be at least partially addressed by using appropriate designs and analyses, which we discuss. Our focus is on segmented regression of an interrupted time series, which does not require a concurrent control group; we also present alternative designs including difference-in-difference, stepped wedge, and cluster randomization. Conducting segmented regression well requires a sufficient number of time points within each period, along with a robust set of potentially confounding variables. This method compares preintervention and postintervention changes over time, divergences in the outcome when an intervention begins, and trends observed with the intervention compared to trends projected without it. Difference-in-difference methods add a concurrent control, enabling yet stronger inference. When done well, the discussed methods permit robust inference on the effect of an intervention, albeit still requiring assumptions and having limitations. Methods are demonstrated using an interrupted time series study in which anesthesiologists took responsibility for an adult medical emergency team from internal medicine physicians in an attempt to improve outcomes.