Thread by @ewanbirney

Had another moment of "well, yes, but people *are* different" and "you geneticists use continental groups in your analysis" as we skirted around discussions of ethnicity / race in health impacts. TL;DR Partially correct but the underlying mindset that ethnicity=genetics is wrong

Let's deal with the correct things first. Yes, people are different partly (sometimes mainly) due to genetics. Visibly, eg height, weight, hair colour, skin colour, smoking habits + invisibly, eg cholesterol levels, heart trabeculation levels, likelihood of getting breast cancer

Some of these visible differences we integrate into the gestalt assessment of ourselves and others for ethnicity, as represented by self identified ethnicity boxes which people tick, eg "Black British, White English, British Indian, British xxx", gloriously variable by society

Also correct is that geneticists have (relatively) complex procedures in analysis which confusingly often use a short hand of the results with similar sounding language to ethnicity labels, eg, "European American" +non-random correlation to self identified ethnicity

The most "classic" use is the selection of subsets of cohorts of people where geneticists use wide genetic relatedness to select a group of people such that (a) that their environment, most importantly social (b) that genetic properties (eg, LD) are reasonably homogenous

We do this by using a longstanding statistical approach to analyse variation (Principal component analysis) to project the high dimensional genetics into a lower space +often use the first 2 dimensions (easy to visualise) to crudely selected people (draw a circle around the blob)

It is worth stressing there is nothing magical about PCA, and each PCA is unique to the *samples*. The "central blob" being European American in US cohorts is due to the sampling in the US; in Japanese cohorts the central blob is "Japanese" etc.

Although the selection of the biggest blob in the 2-D PCA nearly always associates with a self identified ethnicity (eg. "White" in UK BioBank) *plenty* of people who tick that box don't make it into the blob. This is not because they "got it wrong" or have some family secret ...

This is because we want to be conservative in the sub-selection of people to meet some statistical criteria - we're not trying to describe ancestry in the population genetics use of that term (PCA is not a good representation of this) nor predict ethnicity.

Focusing on the social environment component first, because I think it is most important, this is leveraging the fact that recent ancestry is a good marker for social environment - eg, crudely Black British people experience a different social environment to White British people

(it's weird, but we're using genetics to "predict" social environment features here)

What we're aiming to achieve is twofold - the most important is getting to pseudorandomisation of the environment with respect to genetics - if this holds, our downstream GWAS will hold. The second is limiting variation from environment (easier to spot the genetic variation).

It is also useful we're selecting for a relatively homogenous and random genetics - for example that linkage disequilibrium patterns of SNPs are well modelled as coming from one distribution.

Frustratingly we (as geneticists) don't study much this procedure about *why* it works and what precisely it is modelling (or rather, the properties of the people not selected vs selected) - partly because this walks you into some messy societal things, often not measured, also>>

because by definition the subselection is the majority of the people and the people left out of the blob are hetreogenous probably both in their social environment as well as their genetics - the former being the big headache.

As we (@aylwyn_scally @AdamRutherford @JenniferRaff @minouye271 ) have outlined in this preprint (arxiv.org/abs/2106.10041) we think geneticist's use of language in this setting and other language usage is confusing and can cause harm.

It is clunky, but we suggest using more technical precise terms (eg, first usage, "The European-associated PCA
cluster, which aims to minimise variation in non-genetic factors and genetic factors" ; repeated usage "PCA-selected European subset"

We welcome feedback from genetics and non-genetics colleagues on how best to craft our language to explain what we're doing.

Back to my conversation - given geneticists do these sorts of subselections in their analysis, are we not saying there is *something* in the genetics of continental groupings / continental ancestry?

No, we're not. We're using (recent) continental ancestry to predict the *social environment* that people experience in our societies, and making using that the genetics we do our association studies with are randomised to these large societal differences.

It also gives us smoother LD properties and some other nice features, but for me this is secondary to the "elephant in the room" which is social stratification.

(I can sense some of my genetics colleagues now saying 'well what about genetic background, from "real" epistasis through to background linear polygenic scores? Surely *some* of this PCA selection/PCA covariates are modelling this? >>

This deserves its own thread, but the simple answer; given the elephant in the room of social stratification we have to be conservative, and accept we will "model out" some "real" genetic background.

In fact, I think in human genetics there is far less "genetic background" a la laboratory mouse/maize land races/farm animal genetics but this needs more words than a tweet!)

So.... Ethnicity is associated with a little bit genetics inherently (visible characterisitics) and depressingly alot of recent ancestry *via* the social stratification that associated with ethnicity, but this doesn't make ethnicity a crude proxy for genetics.

In both research (basic+clinical) and clinical practice we can't be "colour blind" - human society nearly everywhere is too stratified impacting all sorts of things via many routes - but *genetics* is always best done by measuring *DNA* not self identified ethnicity box ticking