Comments
Comments
复制标题
DOI：
10.1080/19345747.2012.688436
复制
发表时间：
2012-07
影响因子：
1.8
通讯作者：
Michael H. Seltzer
中科院分区：
文献类型：
--
作者：
Michael H. Seltzer
我要感谢本期特刊的编辑和《教育有效性研究杂志》的编辑们给我写这篇评论的机会。此外，我要祝贺Steve Raudenbush、Sean Reardon和Takako Nomi（以下简称RRN）的杰出文章。在教育研究中越来越流行的多站点试验提供了一种极有价值的设计选择。它们使我们能够调查项目在不同地点的影响程度，当有明显的异质性的证据时，它们鼓励我们调查为什么一个项目在某些地点特别有效，而在其他地方却不是。此外，我们看到教育研究者对研究治疗影响结果的机制越来越感兴趣（即，努力估计因果中介效应）。作者提供了一种统计方法，将这些方法论的重点结合在一起，做出了极有价值的贡献。此外，他们还将他们的方法应用于几乎所有现场实验中出现的一个重要问题，即对处理分配的依从性通常不太完美。估计中介(M)因果效应的主要挑战是个体不是随机分配到不同水平的M，因此观察到和未观察到的混杂因素成为主要关注的问题（即与M和结果Y相关的预处理协变量）。工具变量（IV）方法，当其假设成立时，提供了克服这一挑战的手段；在一组假设（见下文）下，它们使我们能够获得M对Y的影响的估计，该影响依赖于M和Y中由随机分配治疗引起的外生变化的部分。RRN提出了一个基于iv的建模框架，在该框架中，他们在站点内建立中介模型，并将其应用程序中关键中介的因果效应的大小（编译器的因果效应）视为不同站点的不同。他们的框架的另一个重要特征是假设学生层面的异质性，从而允许一些学生可能比其他学生更有动力参与，一些学生可能比其他学生从治疗中受益更多。他们仔细地详细说明了他们的方法的假设，并提出了三种策略来估计中介的平均因果效应，以及中介效应在不同地点的异质性。他们的策略之一——选项c——特别有趣。它包括回归特定地点对分配到治疗(T)对结果(Y)的影响的估计（即β s），对特定地点对T对中介(M)的影响的估计（即γ s），产生M的平均因果效应的估计（注意，对于每个地点，随机分配的T被视为导致Y和M的一些变化的工具变量）。如果RRN框架的假设成立，选项C，以及选项A和选项B，提供了对M的因果效应的估计，在某种意义上屏蔽了M的影响
I wish to thank the editor of this special issue and the editors of Journal of Research on Educational Effectiveness for the opportunity to write this commentary. Furthermore, I wish to congratulate Steve Raudenbush, Sean Reardon, and Takako Nomi (henceforth RRN) on their outstanding article. Multisite trials, which have become increasingly popular in educational research, provide an extremely valuable design option. They enable us to investigate the extent to which the effects of programs of interest vary across sites, and when there is evidence of appreciable heterogeneity, they encourage us to investigate why a program may be particularly effective in some sites but not others. In addition, we have seen increasing interest among educational researchers in investigating the mechanisms by which treatments impact outcomes of interest (i.e., efforts to estimate causal mediation effects). The authors have made an extremely valuable contribution by providing a statistical approach that brings these methodological thrusts together. Moreover, they have applied their approach to an important problem that arises in virtually all field experiments, that is, compliance with treatment assignment is generally less than perfect. The major challenge in estimating the causal effect of a mediator (M) is that individuals are not randomly assigned to different levels of M, and thus observed and unobserved confounders become a major concern (i.e., pretreatment covariates related to M and the outcome Y). Instrumental Variable (IV) methods, when its assumptions hold, provide a means of overcoming this challenge; under a set of assumptions (see below), they enable us to obtain an estimate of the effect of M on Y that relies on the portion of the variation in M and Y that is exogenously induced by random assignment to treatment. RRN present an IV-based modeling framework in which they pose mediation models within sites, and treat the magnitude of the causal effect of the key mediator in their application—the causal effect for compliers—as varying across sites. Another important feature of their framework is the assumption of heterogeneity at the student level, thus allowing for the likely possibility that some students may be more motivated to participate than others, and some students may benefit more from the treatment than others. They carefully detail the assumptions of their approach and present three strategies for estimating the average causal effect of the mediator, and the amount of heterogeneity across sites in the effects of the mediator. One of their strategies—Option C—is of particular interest. It involves regressing sitespecific estimates of the effect of assignment to treatment (T) on the outcome (Y) (i.e., β̂s), on site-specific estimates of the effect of T on the mediator (M) (i.e., γ̂s ), yielding an estimate of the average causal effect of M. (Note for each site, T, which is randomly assigned, is viewed as an instrumental variable that causes some of the variation in Y and in M.) If the assumptions of the RRN framework hold, Option C, as well as A and B, provide estimates of the causal effect of M that are in a sense shielded from the impact of