Hi Elan,
good questions. Please see below in line with your questions and let me know if this helps. best wishes, june
Q4 and Q6 – do you want us to write the study objective considering the KM or Cox model (ie risk or rate)? And, which one is the "main analysis" that you want us to discuss?
JMC: For objective, this is the overall objective laid out by the authors at end of the Intro. The two models you mention are specific analysis approaches they used to achieve the objective. We were going for more of a higher level description of the research question and hypotheses, to set the stage for this paper in general. For 'main analysis' - I realize in hindsight this is less evident in the way the paper is written than I perceived it, as it was happening. When I write "main analysis" - I mean the analyses shown in Figs 1 and 2. As I think I mentioned during lecture, the Cox model and data in Table 4 were done after the paper was already submitted/reviewed, as a sensitivity analysis. I have made a note to clarify this in the Instructions next time; in prior years we asked for both, so this question didn't come up. Thank you for pointing this out.
Q7 – the class example notes that there is selection bias because study inclusion for both users and non-users required living to the point of receiving the first survey. This would imply that study inclusion is influenced by the outcome but if this is independent of the exposure, how does this lead to selection bias? Wouldn't we have to assume that the exposure is also influencing selection in order for bias to result? It makes sense that the study population (people who live 1 year) is different from the target population (all people with prostate cancer) so in this sense, it is understandable that there could be selection bias but I was having difficulty conceptualizing the DAG.
JMC: To answer question 7, please draw out the lines showing when T0, E, and A are assigned and think about which, if any, scenario it most resembles, and then answer from there. We are looking for a time sequence line diagram (like the examples in class) for this question, not a DAG. This question is also focused on the ever/never exposure categorization, which is different than what we went over in class (updating of exposure annually for the Cox model). If you are asking a general question (and not specific to q7) - note that we will be going over the DAGs for Emulation Failures 1 and 2 in class on 2/25. For Emulation Failures 3 and 4 (those with immortal time issues) - there is not consensus from experts on how to draw such DAGs so we are opting not to go over this to avoid confusion. In general, though, what you write is on track for how selection bias could happen when E is before time 0, as in Emulation Failure 1.
