11 Apr Within- and Between-Family Prediction in TEDS
A new paper by Selzam, Ritchie, Pingault, Reynolds, O’Reilly and Plomin has just been published on bioRxiv.
They show results of an analysis that I have been looking forward to for a long time– looking at the performance of GPS for a variety of physical and psychological phenotypes between and within pairs of DZ twins in TEDS. The headline, it seems to me, is that for the psychological characteristics– I am going to focus on IQ– the performance of the GPS is much smaller within pairs. For the uninitiated, here is what this means. In a typical GPS study there is only one individual per family, so the correlation between the GPS and the phenotype includes the effect of every socioeconomic difference among the families in the presumably genetic effect. But in siblings or DZ twins, one can separate the effect into two parts: a between family part that includes the family differences, and a within family part that examines the relation between the difference in GPS scores and the difference in phenotype. This relationship controls for between family differences, because they are held constant in the twin pairs.
For IQ, the Beta between families is .263, and within families it is .137. The authors describe this as a 50% reduction, but on a more appropriate R2 scale it is much larger than that: 6.9% of the variance between, and 1.9% within, something like a 70% reduction.
Stuart Ritchie thinks I am being Debbie Downer about a study that, yes, has a whole lot of results in it:
But come on, the study has to be seen in the context in which it was conducted. This is Robert Plomin’s Fortune Teller we are talking about here! The GPS and cognitive performance performance people can’t have it both ways, dancing in the end zone about the performance of GPS when they look big enough to be useful, but then discreetly looking the other way when it turns out that the genetic Fortune Teller has been performing some sleight of hand with her cards. Most of the performance of the IQ GPS is not actually genetic, what is left over after you take out the socioeconomic part is not large enough to be practically useful, and furthermore you have to remember that even the within family part of the prediction is not exactly established as some kind of hard-wired gene to neuron to IQ pathway. Red-haired kids, if they are out there (and of course they are, in one form or another), are included in the within-family correlation.
Stuart Ritchie disagrees with me that the within-family GPS for IQ is the appropriate headline here, using a standard modern genomics argument: Yeah the effect may be small now, but just you wait until we have bigger samples!
The GWAS on which the IQ GPS is based is not exactly small. In any case, I promise to write about the new improved within-family GPS if and when it happens. I agree that the between-within comparison is the interesting part, and it seems to me the conclusion is very clear.
So for example, a common conclusion from this lab is the GPS are about to revolutionize education. The conclusion from the between-within comparison is that making educational decisions about children on the basis of their GPS score is mostly a matter of tracking poor kids into special classes. Do we really want to do that? The enthusiasm for (behavioral) GPS needs to be tempered by what we are learning about it.
Tomorrow: enough already with “quantile analysis.”