In a previous commentary, we considered potential positive contributions of a modern causal inference framework for studying causes of developmental disabilities. The study by Schwartz and Ries demonstrates the power of these methods in the real world. The authors estimate the effects of rectus femoris transfer surgery in children with cerebral palsy based on observational data, and compare the results to previously reported effect estimates based on a randomized control trial (RCT). Previously, we touched on three critical points to consider in the design and execution of a study if the results are to be later interpreted causally. Assuming the study is designed and executed appropriately, there are also assumptions – in some cases quite strong assumptions – that need to be made about the data and the data-generating mechanisms in order to interpret the results as evidence of a causal link between treatment and outcome. These are: (1) Positivity: all individuals have a non-zero chance of receiving the treatment of interest. (2) Consistency: if individuals receive the treatment, then we will observe their outcome under treatment. (3) Exchangeability: treatment assignment is independent of the potential outcome (thus, if all patients receiving and not receiving the treatment were exchanged, the effect would be the same). Positivity is straightforward, and even in observational studies can often be justified. Consistency is essentially the philosophical axiom of identity: people who get treated will respond how treated people would, and those who do not get treated will respond how untreated people would. So long as the treatment is uniform over time and place, this is usually a safe assumption. The final assumption, exchangeability, is the most problematic. It is a strong assumption, and often philosophically difficult to justify in observational studies. Exchangeability is the property that all RCTs possess that makes them the criterion standard in causal inference in medicine: the treatment assignment in an RCT is assigned completely independently of any patient characteristics, measured or unmeasured. Thus, exchangeability in an RCT is not an assumption, it is a feature of the study design. So long as treatment assignment is random, any meaningful difference in outcome between the treated and untreated can be reasonably considered to have been caused by (not merely associated with) the treatment. In observational studies it is almost never safe to assume exchangeability. There are many reasons for this, including that physicians generally strive to prescribe therapies they think are most likely to benefit their patients, a problem known as confounding by indication (from ‘indication for treatment’). Should this occur, treatment assignment is not random, and is instead conditional on patient characteristics, confounding estimates of treatment effect. The purpose of propensity score matching by Schwartz and Ries is to break the dependence of treatment assignment on patient characteristics, i.e. to remove confounding by indication. This is achieved by making the treatment assignment conditionally exchangeable, given the propensity score. Even with suitable propensity score matching, conditional exchangeability remains a strong (and often unverifiable) assumption when making causal inferences based on observational data. Schwartz and Ries have made encouraging progress in overcoming this hurdle, and in answering in the affirmative the question posed in our title, by demonstrating good agreement with a previously-published RCT. The use of propensity score methods with observational data will never be a substitute for a well-designed and well-executed RCT. However, RCTs alone are often insufficient to provide definitive answers about the efficacy of new therapies, and they can be time-consuming, expensive, and sometimes impossible to conduct. We see great potential for the modern framework for causal inference to increase the value of observational clinical experience in medical sciences.

中文翻译：

---

脑瘫股直肌转移手术：可以从观察数据中做出因果推断吗？

在之前的评论中，我们考虑了现代因果推理框架对研究发育障碍原因的潜在积极贡献。Schwartz 和 Ries 的研究证明了这些方法在现实世界中的威力。作者根据观察数据估计了股直肌转移手术对脑瘫儿童的影响，并将结果与​​之前基于随机对照试验 (RCT) 报告的效果估计值进行了比较。之前，我们谈到了在研究的设计和执行中要考虑的三个关键点，如果以后要对结果进行因果解释。假设研究设计和执行得当，还需要对数据和数据生成机制做出一些假设——在某些情况下是相当强的假设，以便将结果解释为治疗和结果之间因果关系的证据。它们是： (1) 积极性：所有个体都有非零机会接受感兴趣的治疗。(2)一致性：如果个体接受治疗，那么我们将观察他们在治疗下的结果。(3) 可互换性：治疗分配与潜在结果无关（因此，如果所有接受和未接受治疗的患者都进行了交换，效果将相同）。积极性是直截了当的，即使在观察性研究中也常常是合理的。一致性本质上是身份的哲学公理：被对待的人会回应被对待的人，那些没有得到治疗的人会像未经治疗的人一样做出反应。只要治疗在时间和地点上是统一的，这通常是一个安全的假设。最后一个假设，可交换性，是最成问题的。这是一个强有力的假设，并且在观察性研究中通常在哲学上难以证明其合理性。可交换性是所有 RCT 都拥有的特性，使它们成为医学因果推断的标准：RCT 中的治疗分配完全独立于任何患者特征，测量的或未测量的。因此，RCT 中的可交换性不是假设，而是研究设计的一个特征。只要治疗分配是随机的，接受治疗和未接受治疗的结果之间的任何有意义的差异都可以合理地认为是由治疗引起的（不仅仅是与治疗有关）。在观察性研究中，假设可交换性几乎是不安全的。造成这种情况的原因有很多，包括医生通常会努力开出他们认为最有可能使患者受益的疗法，这个问题被称为“适应症混淆”（来自“治疗适应症”）。如果发生这种情况，治疗分配不是随机的，而是取决于患者特征，混淆治疗效果的估计。Schwartz 和 Ries 进行倾向评分匹配的目的是打破治疗分配对患者特征的依赖性，即消除适应症的混淆。这是通过在给定倾向分数的情况下使治疗分配有条件地可交换来实现的。即使有合适的倾向得分匹配，在基于观察数据进行因果推断时，条件可交换性仍然是一个强有力的（通常是无法验证的）假设。Schwartz 和 Ries 在克服这一障碍方面取得了令人鼓舞的进展，并通过证明与先前发表的 RCT 的良好一致性，肯定地回答了我们标题中提出的问题。使用具有观察数据的倾向评分方法永远不会替代设计良好且执行良好的 RCT。然而，单独的随机对照试验通常不足以提供关于新疗法疗效的明确答案，而且它们可能耗时、昂贵，有时甚至无法进行。