general question about likely pitfalls when studying a single disease

general question about likely pitfalls when studying a single disease

by Laura Koth -

would appreciate clarification regarding a recurring theme in epi 203 and 207 classes which I think most recently came up with the firefighter dag in last hw and the obesity paradox presented ~58-59 min into Dr. Mayeda's QBA talk, which is the following:

based on the dags related to these two examples, is it correct to infer from this that anytime one performs disease focused research, you will most likely have collider bias (due to  restricting your analysis to people with the disease, like heart failure in the obesity example) resulting in spurious associations between your exposure of interest with your outcome of interest through spurious pathways?

This is the message that I am taking away from this content and want to confirm this is accurate reasoning.

thank you



Re: general question about likely pitfalls when studying a single disease

by June Chan -

Hi Laura, thank you for this question. I will think about it some more and try to catch up with the TA's and Dr. Glymour as I wasn't there last week and not sure if this is prompted by some additional examples that may have been discussed. 


In general though, I think case-only studies have different bias vulnerabilities (esp selection bias and immortal PT bias), and this is why we raise this in week 2 (the trial emulation failure examples; and raising the case-only cohort idea). However, each scenario still warrants individual evaluation and I'd like to think that it is not all futile.  If you have not looked at it yet, I encourage you to read the 2016 J Clin Epid paper by hernan and Hernandez-Diaz from Week 2 (optional reading).


Hernán MA et al, Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses J Clin Epid, 2016. https://www.ncbi.nlm.nih.gov/pubmed/27237061


In the second half of this paper, they discuss methodologic approaches to try and mitigate these biases, and I believe these are covered in a Spring biostats course (bio215).


best, 

JMC

Re: general question about likely pitfalls when studying a single disease

by Laura Koth -

thank you for this answer!! very helpful and I will look into all of your suggestions!

Re: general question about likely pitfalls when studying a single disease

by June Chan -

Hi Laura, 

here are further thoughts from Dr. Glymour:


"Anytime you conduct disease-outcome focused research, i.e., you are studying only a group of people already diagnosed with a condition, there is the potential for collider bias. This potential only becomes a a reality if the exposure you are interested in influences the risk of disease incidence and also some other factor that influences disease incidence also influences the outcome you are interested in. Even then, it may not be a big problem (we have lots of biases we live with because they are not big enough to meaningfully change interpretation of the results) unless the exposure and that other factor interact when determining disease incidence.  By "interact" here we mean specifically interact on a  multiplicative scale.  We can walk through the math of that, but it's a bit of a detail.  The big picture is - yes, always be worried about the potential for collider bias in disease prognosis research if your exposure influences disease incidence; - consider whether there are any other major factors with a strong influence on disease incidence that also influence your outcome; - if there are such factors, ask whether the influence of those factors on disease incidence might be really different for people with vs without the exposure.  You usually won't be sure about the last two parts of this, so you may have to use a sensitivity analysis."

cheers & happy friday!


June