In a recent International Journal of Epidemiology article, Tennant et al. (2021) asserted that regressions with change score estimators (i.e. ‘gain scores’, ‘difference scores’) do not estimate causal treatment effects in observational data.¹ Their argument—aided by directed acyclic graphs (DAGs)—relies on within-person models with a single treatment measure and causally related outcomes. Whereas this framing is conventional for change score analyses,² it may not apply to other settings that are commonly represented in panel data: that is, change score analyses with dyadic (two-person) data can yield valid treatment estimates and may be preferable to more general methods. I motivate this claim with a common study design: sibling comparison analysis.

中文翻译：

---

变化分数分析的机会

在最近的国际流行病学杂志文章中，Tennant等人。(2021) 断言，使用变化分数估计量（即“获得分数”、“差异分数”）的回归不能估计观察数据中的因果治疗效果。¹他们的论点——在有向无环图 (DAG) 的帮助下——依赖于具有单一治疗措施和因果相关结果的人体内模型。虽然这种框架对于变化分数分析来说是传统的，²它可能不适用于面板数据中通常表示的其他设置：也就是说，使用二元（两人）数据进行的变化评分分析可以产生有效的治疗估计，并且可能比更一般的方法更可取。我通过一个共同的研究设计来激发这一主张：兄弟姐妹比较分析。