Learning From an Association Analysis Using Propensity Scores.
Learning From an Association Analysis Using Propensity Scores.
复制标题
使用倾向得分从关联分析中学习。
DOI：
10.1097/pcc.0000000000002842
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Kreif N
中科院分区：
文献类型：
--
作者：
Kreif N
在本期《儿科重症监护医学》中，Geneslaw等人(1)旨在评估有创机械通气(IMV)对因呼吸道疾病入院的儿童神经发育结局和精神健康障碍的影响。作者发现，与一组住进综合医院而没有接受IMV的儿童相比，对被诊断为患有严重呼吸系统疾病的儿童进行IMV支持与随后的不良神经发育和心理健康结果有关。随之而来的自然问题是：这种关联是否可以通过IMV或其他ICU治疗的因果有害影响来解释，以及临床医生是否可以减轻IMV对患有严重呼吸道疾病的儿童的这种后果？重要的是，Geneslaw等人(1)没有将他们的分析定为因果关系。作者报告了与因果影响相反的关联，并得出结论，所观察到的关联可能是由基础诊断的特征驱动的，这些特征需要使用IMV，例如缺氧和炎症。尽管如此，读者和媒体可能会忍不住将这些联系解读为部分受到IMV干预的因果影响的推动。尽管措辞谨慎，但作者确实给出了这样一种解释的一些理由。例如，IMV的一个潜在的有害连锁反应是由于使用镇静/止痛药而导致的精神错乱，这是这项工作的驱动假设之一。其次，他们使用倾向得分匹配(PSM)这样的技术作为主要的分析方法，经常用于估计观察性研究中可能的因果效应[2]，在这方面，作者确实使用了因果分析中熟悉的术语，如“混杂因素”。对于纯粹的关联分析，PSM不是必要的，而是可以使用机器学习方法来预测不利的心理健康结果。读者还将在随附的社论中看到，ShPitser等人(3)概述了根据有向非循环图的框架，给予回溯性研究因果解释所需的假设。这两位编辑强调了两个阻碍IMV研究进行因果解释的主要陷阱：1)未观察到的混杂因素的存在；2)出院时的审查问题。
In this issue of Pediatric Critical Care Medicine, Geneslaw et al (1) aimed to evaluate the consequences of invasive mechanical ventilation (IMV) on neurodevelopmental outcomes and mental health disorders in children admitted to hospital with a diagnosis of respiratory illness. The authors found that IMV support for children with a diagnosis of severe respiratory illness was associated with subsequent adverse neurodevelopmental and mental health outcomes, compared with a cohort of children who were admitted to general hospital and did not receive IMV. Natural questions that follow are: whether any of this association may be explained by a causal—harmful—effect of IMV or other ICU therapies, and whether clinicians can mitigate such consequences of IMV in children with severe respiratory illness?Importantly, Geneslaw et al (1) do not frame their analyses as causal. The authors report associations as opposed to causal impacts and conclude that it is likely that the observed association may be driven by features of the underlying diagnosis that necessitates use of IMV, for example, hypoxia and inflammation. Still, it may be tempting for readers and the media to interpret these associations as being partly driven by the causal impact of the IMV intervention. In spite of careful language, the authors do give some reasons for such an interpretation. For example, one potential harmful “knock-on” effect of IMV is delirium resulting from the use of sedative/analgesic medications, which is one of the driving hypotheses of the work. Second, they use a technique such as propensity score matching (PSM) as the main analytical method, frequently used to estimate possible causal effects in observational studies (2), and, in this regard, the authors do use familiar terms from causal analysis such as “confounders.” For a purely associational analysis, PSM would not be necessary, but instead machine learning approaches could be used to predict adverse mental health outcomes. Readers will also see in an accompanying editorial that Shpitser et al (3) outline the assumptions required to give a retrospective study a causal interpretation, relying on the framework of directed acyclic graphs. The editorialists highlight two major pitfalls that prevent the IMV study from having a causal interpretation: 1) the presence of unobserved confounders and 2) the issue of censoring upon hospital discharge.