Today, my paper "Is caviar a risk factor for being a millionaire?" was published in the Christmas Edition of the BMJ (formerly the British Medical Journal). The paper is available at http://www.bmj.com/content/355/bmj.i6536 but it is unfortunately behind a paywall. I am hoping to upload an open access version to a preprint server but this needs to be confirmed with the journal first.
In this paper, I argue that the term "risk factor" is ambiguous, and that this ambiguity causes pervasive methodological confusion in the epidemiological literature. I argue that many epidemiological papers essentially use an audio recorder to determine whether a tree falling in the forest makes a sound, without being clear about which definition of "sound" they are considering.
Even worse, I argue that epidemiologists often try to avoid claiming that their results say anything about causality, by hiding behind "prediction models". When they do this. they often still control extensively for "confounding", a term which only has a meaning in causal models. I argue that this is analogous to stating that you are interested in whether trees falling in the forest causes any human to perceive the qualia of hearing, and then spending your methods section discussing whether the audio recorder was working properly.
Due to space constraints and other considerations, I am unable to state these analogies explicitly in the paper, but it does include a call for a taboo on the word risk factor, and a reference to Rationality: AI to Zombies. To my knowledge, this is the first reference to the book in the medical literature.
I will give a short talk about this paper at the Less Wrong meetup at the MIRI/CFAR office in Berkeley at 6:30pm tonight.
(I apologize for this short, rushed announcement, I was planning to post a full writeup but I was not expecting this paper to be published for another week)