Data, design and background knowledge in etiologic inference. Epidemiology

• Main Author: James M. Robins
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Published: 2001
• Subjects: Greenland
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.329.8983; http://www.biostat.jhsph.edu/~dscharf/Causal/00001648-200105000-00011.pdf

I use two examples to demonstrate that an appropriate etiologic analysis of an epidemiologic study depends as much on study design and background subject-matter knowledge as on the data. The demonstration is facilitated by the use of causal graphs. (Epidemiology 2001;11:313–320) Key Words: inference, etiology, study design, data collection, data analysis, epidemiologic methods Greenland et al 1 discussed the use of causal graphs in epidemiologic research. A limitation of that paper was that it was lacking concrete examples designed to help the reader see how to take one’s knowledge of study design, temporal ordering, basic biology, and epidemiologic principles to construct an appropriate causal graph. Here I present two epidemiologic thought experiments that make the point that the choice of an appropriate etiologic analysis depends as much on the design of the study and background subject-matter knowledge as on the data. Specifically, in the first, I provide a single hypothetical dataset and three differing study designs, each of which plausibly could have given rise to the data. I show that the appropriate etiologic analysis differs with the design. In the second, I revisit a well-known epidemiologic controversy from the late 1970s. Horowitz and Feinstein 2 proposed that the strong association between postmenopausal estrogens and endometrial cancer seen in many epidemiologic studies might be wholly attributable to diagnostic bias. Others disagreed. 3–5 Part of the discussion centered on the issue of whether it was appropriate to stratify on vaginal bleeding, the purported cause of the diagnostic bias in the analysis. The goal here is to show, using causal graphs, that the answer depends on underlying assumptions about the relevant biological mechanisms.