Abstract
We regret an error in the article. In correcting this error, we no longer recommend the methods in Section 4 of the article. The rest of the article and, most importantly, the sensitivity analyses proposed in Section 3 for moderators observed in the RCT but not in the target population, are unaffected.
Introduction
Generalizing treatment effects from a Randomized Controlled Trial (RCT) to a broader target population is a central challenge in evidence-based medicine and social science. This process often involves adjusting for "moderators"—variables that influence the magnitude of the treatment effect. In their 2017 work, the authors proposed sensitivity analyses for two scenarios: the $V$ case (moderators measured in the RCT but missing in the target population) and the $U$ case (moderators completely unobserved in both). This paper serves as a critical correction, retracting the methods proposed for the $U$ case due to a fundamental statistical oversight involving collider bias.
The Collider Bias Trap
The core error identified in the original proposal for the $U$ case stems from the assumption that an unobserved moderator ($U$) could remain independent of observed variables ($Z$) in both the general population and the RCT sample. However, if both $Z$ and $U$ influence the probability of an individual participating in the trial ($S$), then trial membership acts as a collider.
As shown in the figure above, conditioning on trial participation ($S=1$) creates a mathematical association between $Z$ and $U$, even if they are independent in the general population ($p_{z,u}$). This phenomenon occurs because knowing that someone is in the trial ($S=1$) provides information about $U$ if we already know $Z$. For instance, if $Z$ and $U$ both increase the likelihood of participation, and we observe an individual in the trial with a low value of $Z$, we can infer they likely have a high value of $U$ to compensate. This induced correlation means the joint distribution $p_{z,u|S=1}$ will differ from $p_{z,u}$, invalidating the independence assumptions required for the previously proposed bias-formula and weighting-plus-bias methods.
Retraction of Methods
The authors formally retract the techniques described in Section 4 of the original 2017 paper. Specifically:
- Bias-formula-based sensitivity analysis for unobserved moderators is no longer recommended.
- Weighting-plus-bias-formula-based sensitivity analysis is similarly retracted.
The authors also clarify that "regressing out" observed variables does not ensure independence from unobserved moderators; it only ensures they are uncorrelated. This distinction is vital for the mathematical validity of weighting methods in causal inference. Because weighting relies on the correct specification of the selection mechanism, the induced dependence between $Z$ and $U$ within the trial sample ($S=1$) breaks the logic of the proposed hybrid approach.
Conclusion and Scope of Validity
Importantly, this correction does not affect the methods proposed for the $V$ case. When moderators are measured within the RCT, the sensitivity analyses described in Section 3 remain sound and are still recommended for researchers. The empirical examples provided in the original paper, which focused on the $V$ case, also remain valid. While the problem of completely unobserved moderators remains a significant challenge, this correction ensures that future research avoids the "collider bias trap" in generalization efforts. Researchers are encouraged to consult Nguyen et al. (2018) for further details on the application of $V$ case methods.
Summary updated