https://orcid.org/0000-0002-7534-0945
-
Views
-
Cite
Cite
Ludovic Trinquart, Sandro Galea, TWO AUTHORS REPLY, American Journal of Epidemiology, Volume 188, Issue 8, August 2019, Pages 1–2, https://doi.org/10.1093/aje/kwz129
Close - Share Icon Share
Extract
We thank Dr. Hamra (1) for his interest in our work. Fundamentally, Dr. Hamra expresses skepticism that the E value can be informative for epidemiologic work. While we point out several limitations of the E value in our paper (2), we disagree with this broader assertion.
Dr. Hamra’s first point is that there are multiple potential biases that can influence interpretation of findings from an epidemiologic study (1). We agree. In the Discussion section of our paper (2), we addressed this point and noted that biases other than confounding, including reporting bias and measurement error, could distort the effect size estimate in selected individual studies. We concluded that unmeasured confounding could explain away most of the observed associations, in the absence of other forms of bias (2).
Dr. Hamra further notes that it is essential to consider which adjustments have been performed in order to interpret the E value within the context of a specific study (1). We are not aware of a universally recognized method for assessing whether confounding control was sufficient in a given study. One approach would be to define a minimum set of essential confounders. This task would require a comprehensive overview of confounders operating for each specific exposure and health outcome. It would also require a consensus process to avoid the inherent subjectivity of the task. For example, a survey of epidemiologists showed little consistency in choosing confounders to control for when designing an epidemiologic study (3). Dr. Hamra considers it likely that most studies within a field have not accounted for a subset of essential confounders (1). In Dr. Hamra’s meta-analysis on the association between outdoor particulate matter exposure and lung cancer (4), we found that only 2 of the 18 selected studies had adjusted for smoking, socioeconomic status/income, education, and occupation. We agree with the essential observation that confounders are typically left unmeasured and argue that if unmeasured confounding is likely, E values can help assess the robustness of studies across the field by aiding researchers in considering whether the magnitude of unmeasured confounding is plausible.
