Centre for Market & Public Organisation, University of Bristol, 1TX, UK.
Biostatistics. 2010 Oct;11(4):756-70. doi: 10.1093/biostatistics/kxq024. Epub 2010 Jun 3.
Structural mean models (SMMs) were originally formulated to estimate causal effects among those selecting treatment in randomized controlled trials affected by nonignorable noncompliance. It has already been established that SMMs can identify these causal effects in randomized placebo-controlled trials under fairly weak assumptions. SMMs are now being used to analyze other types of study where identification depends on a no effect modification assumption. We highlight how this assumption depends crucially on the unknown causal model that generated the data, and so is difficult to justify. However, we also highlight that, if treatment selection is monotonic, additive and multiplicative SMMs do identify local (or complier) causal effects, but that the double-logistic SMM estimator does not without further assumptions. We clarify the proper interpretation of inferences from SMMs by means of an application and a simulation study.
结构均值模型（SMM）最初是为了估计在受不可忽视的不依从性影响的随机对照试验中选择治疗的人群中的因果效应而制定的。已经证明，在相当弱的假设下，SMM 可以在随机安慰剂对照试验中识别这些因果效应。现在，SMM 正被用于分析其他类型的研究，这些研究的识别取决于无效应修正假设。我们强调了这个假设如何取决于产生数据的未知因果模型，因此很难证明其合理性。然而，我们也强调，如果治疗选择是单调的，加性和乘性 SMM 确实可以识别局部（或依从者）因果效应，但双对数 SMM 估计器在没有进一步假设的情况下无法识别。我们通过应用和模拟研究来澄清 SMM 推断的正确解释。
