Thread by @evpatora

Coincidentally (not?), I was just telling at my intro to bias talk at @DCEAarhus, that this image makes somewhat regular rounds on here

I took this coincidence a bit too seriously and made a 🧵
#selectionbias #epitwitter
1/n

#selectionbias may be one of the more challenging concepts to understand in epidemiology and ever

#selectionbias #epitwitter
2/n

Nothing helps to explain a concept like a good xkcd comic though xkcd.com/2618/

#selectionbias #epitwitter
3/n

More casual (did you read this right?😄) & intuitive examples are a good medium to explain complex concepts. Visualizing distributions always helps too.

#selectionbias #epitwitter
4/n

One of the most intuitive explanations of the selection bias due to collider stratification I have ever encountered was by @MariaGlymour using as an example the association between height and speed when conditioning on being a basketball player. slideshare.net/mglymour/quick…

5/n

The structural definition of selection bias: “conditioning on a common effect of (a cause of) the exposure and (a cause of the) outcome” is a must-read @_MiguelHernan
tinyurl.com/selectionbiass…

#selectionbias #epitwitter
6/n

Examples of paradoxes in epidemiologic studies arising due to conditioning on a collider or its descendant are plenty. Here are several that improved my understanding and some of them are specific to reproductive epidemiology - a field where colliders are certainly tricky.

7/n

To start, the famous birth weight paradox. When investigating the association between smoking and infant mortality, conditioning on birth weight leads to a paradoxical reversed association within the stratum of low birth weight.
tinyurl.com/LBWparadox
#selectionbias
8/n

The biased estimate suggests that smoking is associated with a reduced risk of infant mortality for LBW babies (which is certainly not true).

#selectionbias
9/n

Conditioning on LBW opens a biasing path via other and more severe causes of infant mortality (U) like major congenital malformations (smoking -> LBW <- U -> mortality).

#selectionbias
10/n

Another paper, which lays out collider stratification on multiple nodes neatly is by @liew_zeyan, @beate_ritz and @oacarah et al. What happens when conditioning on live birth in pregnancy cohorts? academic.oup.com/ije/article/44…

#selectionbias
11/n

But selection bias can arise as well in the absence of conditioning on a collider: academic.oup.com/aje/article/18…

#selectionbias
12/n

It can be confusing! This paper by @LuHaidong @EpidByDesign et al. explains the two types of selection bias: type 1 due to conditioning on a collider or a descendant of a collider and type 2 due to restriction to a particular level(s) of an effect measure modifier.

13/n

#selectionbias
14/n

Thinking about selection bias is important in any line of research. Here is the example with covid-19 vaccination and subsequent infection and the severity-related covid-19 outcomes

#selectionbias
15/n

Finally, “Ignoring selection bias can delay advances in medical science.” tinyurl.com/obesityparadox…
Hard to argue with that!


#selectionbias
16/16

Feel free to add your favourite sources on #selectionbias 🤓 #epitwitter