,
21 tweets,
5 min read
Read on Twitter
I found the discussion of Quetelet and populations at the beginning of Chapter 3 of Epidemiology and the People’s Health particularly instructive, especially for genetic epidemiology. #epibookclub #epipeopleshealth #gwas #genepitwitter 1/18 
The question “who—or what—determines populations or groups that merit comparison” is an important but tetchy one. 2/18
The concept of “population stratification bias” in genetic epidemiology is usually introduced using a toy example: say we’re studying two populations, with random mating within but no mating across populations. 3/18
If the avg value for the trait under study differs between these two pop’ns, then any allele that has diff’t frequencies across pop’n will appear assoc’d with the trait—even if the allele is not causally related to the trait (eg no trait-allele assoc’n within either pop’n). 4/18
While this might be a good thought experiment for teaching, it is an extreme oversimplification of real human population demography and genetics. 5/18
I don’t care how you draw imaginary partitions around people in the real world, folks will not be randomly mating within your boundaries and there will be some (history of) mating across groups. 6/18
To make it more complicated, real-world mating patterns will be strongly influenced by things that also affect human health: class, caste, war, imperialism, slavery... 8/18
In practice, from the technical standpoint of gene mapping, we may primarily want an analytic approach that avoids false positive associations due to population stratification. 9/18
In that case, defining groups with broadly similar ancestry and running analyses within these group strata (using principal components to adjust for residual structure) could work just fine. 10/18
(For those who haven’t seen the sausage being made, these groups are often defined by plotting subjects in genetic PC space, then drawing ellipses around different clusters, becuase... well, because, it’s obvious.) 11/18
This procedure arguably worked fine over the first decade of genome-wide association studies, thanks to the very stringent significance thresholds set for locus discovery. Any residual bias would not be strong enough to pass those thresholds. 12/18
Residual population stratification bias may be more of in issue now with GWAS sample sizes pushing 1,000,000+ and with polygenic score analyses relaxing significance thresholds. 13/18
Anyway, back to groups and populations: a challenge with this technically-useful approach to defining strata for analysis is the temptation to reify these arbitrary groupings into something they’re not. 14/18
This is not helped by our use of convenience labels for these strata: e.g. “white” or “European-Ancestry” for clusters where most subjects self-identify as white, “African-American,” “Asian,” etc. 15/18
These convenience labels suggest that individuals are defined by their group membership, rather than groups being defined by who’s been placed into them. This can be taken as support for “race” as a biologically meaningful entity. 16/18
(I may be rehashing ebirney’s argument for using “subjects of European-ancestries” rather than “European-ancestry” ...) 17a/18
(... and some of the arguments from this piece.) 17b/18 buzzfeednews.com/article/bfopin…
So what to do? Suggestion welcome! At a minimum: be mindful about what you are doing, why, and what it means. “Race is a confounder and I want to adjust for it”—are you using “race” as a proxy for racism (individual or systemic)?Can your results be misinterpreted? 18a/18
And while the discussion of Quetelet made me think of issues around the use of race” in genetic epidemilogy, these are not restricted to genetic epi. 18b/18 reallyfin
Whoops! A couple of goofs to fix on that lest thread. “ebirney” is @ewanbirney (if it wasn’t obvious). And I cut off some important text from the paragraphs on Quetelet (now attaches). (I’m letting the typos go).
Ack. “Attached.” It’s like Zeno’s paradox but for tweets...
