[论文解读] Mimicking counterfactual outcomes for the estimation of causal effects

Judith J. Lok|arXiv (Cornell University)|Sep 3, 2004

Advanced Causal Inference Techniques参考文献 15被引用 3

一句话总结

本文通过证明结构嵌套模型可在不强制要求所有反事实结果与观测数据之间存在确定性关联的情况下估计治疗效应，放松了时间依赖性治疗因果推断中对确定性反事实假设的依赖。其核心贡献在于，在治疗与协变量之间存在反馈回路的情境下，使更符合实际的异质性治疗效应建模成为可能。

ABSTRACT

Large observational studies have become commonplace in medical research. Treatment may be adapted to covariates at several instances without a fixed protocol. Estimation or even definition of treatment effect is difficult in that case. Treatment influences covariates, which influence treatment, which influences covariates, etcetera. To distinguish between these options, even the famous time-dependent Cox-model cannot be used. Robins (1992, 1998), Keiding (1999) and Lok (2001, 2004) study Structural Nested Models to estimate treatment effects even in this difficult setting. Their methods are based on so-called counterfactuals: the outcome a patient would have had if treatment was withheld after a certain time. It is clearly impossible for these outcomes to be observed in all patients. Yet we will show how counterfactual thinking is a very helpful tool to study estimation of treatment effect in the presence of time-dependent covariates. Previous work on these models was usually based on the assumption that the correct model combined with observations made it possible to calculate all counterfactuals for each patient. This assumption was considered not plausible, since it assumes the exact same treatment effect for each patient. This paper provides the cornerstone for the relaxation of treatment effects in e.g. Robins (1992, 1998) or Keiding (1999) in that, at least if there is no censoring, the assumption that counterfactuals are connected with the observed data in a deterministic way is not necessary. We hope that this will contribute to the discussion about causal reasoning.

研究动机与目标

• 解决在治疗与协变量相互影响的观察性研究中估计因果效应的挑战。
• 放松先前在结构嵌套模型中要求的强假设，即反事实结果必须与观测数据确定性关联。
• 通过允许反事实与观测数据之间存在随机关联，支持对异质性治疗效应的更真实建模。
• 为在标准模型（如时间依赖性Cox模型）失效的复杂纵向情境中因果推理的理论基础做出贡献。

提出的方法

• 提出放松结构嵌套模型中对确定性反事实假设的限制，允许反事实以非确定性方式与观测数据关联。
• 证明在无删失条件下，Robins（1992, 1998）和Keiding（1999）的核心识别框架依然有效，无需精确的、确定性的反事实预测。
• 使用反事实思维作为概念工具，即使反事实结果未被观测且无法确定性预测，也能定义治疗效应。
• 依赖结构嵌套模型框架，通过建模不同治疗策略下的潜在结果来估计因果效应。
• 确立治疗效应的识别不依赖于反事实完全由观测数据和模型结构决定的假设。
• 为在治疗与时间依赖性协变量之间存在反馈的情境中使用反事实推理提供了理论依据。

实验结果

研究问题

• RQ1 结构嵌套模型是否可以在不假设反事实结果由观测数据和模型参数确定性决定的情况下，估计因果效应？
• RQ2 在治疗与协变量相互影响的纵向研究中，如何应用反事实推理？
• RQ3 放松确定性反事实假设对复杂动态治疗情境下因果推断有何影响？
• RQ4 当反事实未与观测数据确定性关联时，在何种条件下仍可识别治疗效应？

主要发现

• 本文证明，先前在结构嵌套模型中被认为识别所必需的确定性反事实假设，并非有效因果效应估计所必需。
• 即使反事实未被观测，只要不存在删失，治疗效应仍可在结构嵌套模型框架下一致估计。
• 放松确定性反事实假设，使得在个体间对异质性治疗效应的建模更加符合实际。
• 即使无法从观测数据中确定性重构反事实结果，反事实思维依然是定义和估计因果效应的有效且有用的工具。
• 通过证明识别不依赖于确定性反事实映射，本研究加强了结构嵌套模型的理论基础。
• 本研究支持在治疗与协变量动态演变且相互作用的复杂医疗场景中更广泛地应用因果推断方法。