The person who taught your epidemiology course is incorrect: As Ilya correctly points out, differential misclassification can certainly occur even in a prospective cohort study. Unfortunately, this exact confusion is very common in epidemiology.
Some reading on how to reason about mismeasurement bias using causal graph is available in Chapter 9 of the Hernan and Robins textbook, which is freely available at http://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/ .The chapter contains all the relevant principles, but doesn’t explicitly answer your questions. I also have a set of slides that I use for teaching this material, these slides contain some directly relevant examples and graphs. I can send these to you if you contact me at ahuitfeldt@mail.harvard.edu.
The distinction between “cohort” and “case-control” is not relevant here. The professor is using it as shorthand for retrospective/prospective. The most useful definition of “prospective” and “retrospective” is that in a prospective study, the exposure variable is measured before the outcome variable is instantiated. This is a useful definition because under this definition of prospective, there cannot be a directed path from the outcome to the measurement error on the exposure, which reduces the potential for bias. However, there can still be common causes of the outcome and the measurement error on the exposure, which will results in differential misclassification of the exposure.
Thank you, I hope I indeed follow through on it! My interest in epi stems from an interest in stats, which was sparked from reading about Bayesian statistics through LW and being utterly overwhelmed from it!
Thanks—I was worried I was missing something. Incidentally, I wrote something that you might be interested in on missing data under MNAR that is generalizable to some measurement error contexts.
The answer given was that according to the differential measurement lecture, differential measurement of exposure has to
be dependent on the outcome for there to be error, that’s not going to happen for cohort study cause it’s not till years later
that the outcome is known.
How are exposures set in this study? What if the final outcome depends on an unobserved cause (health status maybe?), and that cause also influences an intermediate outcome that does determine the measurement of some exposure along the way (via doctor assigning the exposure based on it, maybe?)
Or am I misunderstanding the question? (This is entirely possible, I don’t fully understand epi lingo, I just construct counterexamples via d-separation/d-connection in graphs directly).
Where are you taking this class, if you don’t mind me asking?
In cohort studies, the experimenter doesn’t set exposures
Yes I understand, but somehow they are set (maybe by Nature?) The real question I was getting at is whether they were randomized at all, or pseudo-randomized somehow. I was guessing not, so you get time-varying confounding issues alluded to in my earlier post.
So by unobserved you’re referring to say, self report of health status?
Well, if it’s self-report you observe a proxy. I meant actually unobserved (e.g. we don’t even ask them, but the variable is still there and relevant).
In epi this is meets the causal pathways definition for a confounder, if I’m not mistaken.
You are right, in this case, but should be careful about the definition of a confounder, see:
Did you mean “confounding” rather than “confounder”? The difference is important (the former is much easier to define, it is just related to what is called conditional ignorability in epi, the latter is quite tricky).
Is there another question you might be getting at that I can answer without identifying myself?
this was a misleading comment, removed and replaced by this placeholder comment
The person who taught your epidemiology course is incorrect: As Ilya correctly points out, differential misclassification can certainly occur even in a prospective cohort study. Unfortunately, this exact confusion is very common in epidemiology.
Some reading on how to reason about mismeasurement bias using causal graph is available in Chapter 9 of the Hernan and Robins textbook, which is freely available at http://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/ .The chapter contains all the relevant principles, but doesn’t explicitly answer your questions. I also have a set of slides that I use for teaching this material, these slides contain some directly relevant examples and graphs. I can send these to you if you contact me at ahuitfeldt@mail.harvard.edu.
The distinction between “cohort” and “case-control” is not relevant here. The professor is using it as shorthand for retrospective/prospective. The most useful definition of “prospective” and “retrospective” is that in a prospective study, the exposure variable is measured before the outcome variable is instantiated. This is a useful definition because under this definition of prospective, there cannot be a directed path from the outcome to the measurement error on the exposure, which reduces the potential for bias. However, there can still be common causes of the outcome and the measurement error on the exposure, which will results in differential misclassification of the exposure.
this was an unhelpful comment, removed and replaced by this comment
I think it would be very valuable for thinking about bias etc. in epi studies to learn about d-separation, good work on being proactive about it!
Thank you, I hope I indeed follow through on it! My interest in epi stems from an interest in stats, which was sparked from reading about Bayesian statistics through LW and being utterly overwhelmed from it!
Thanks—I was worried I was missing something. Incidentally, I wrote something that you might be interested in on missing data under MNAR that is generalizable to some measurement error contexts.
How are exposures set in this study? What if the final outcome depends on an unobserved cause (health status maybe?), and that cause also influences an intermediate outcome that does determine the measurement of some exposure along the way (via doctor assigning the exposure based on it, maybe?)
Or am I misunderstanding the question? (This is entirely possible, I don’t fully understand epi lingo, I just construct counterexamples via d-separation/d-connection in graphs directly).
Where are you taking this class, if you don’t mind me asking?
this was an unhelpful comment, removed and replaced by this comment
Yes I understand, but somehow they are set (maybe by Nature?) The real question I was getting at is whether they were randomized at all, or pseudo-randomized somehow. I was guessing not, so you get time-varying confounding issues alluded to in my earlier post.
Well, if it’s self-report you observe a proxy. I meant actually unobserved (e.g. we don’t even ask them, but the variable is still there and relevant).
You are right, in this case, but should be careful about the definition of a confounder, see:
http://arxiv.org/abs/1304.0564
Did you mean “confounding” rather than “confounder”? The difference is important (the former is much easier to define, it is just related to what is called conditional ignorability in epi, the latter is quite tricky).
No, that was enough information, thank you.