Biological and statistical interactions \thread part 2 (for part 1 see quoted tweet)
If you know something about biology you can hypothesize about the form of the statistical interaction you expect to see—but (a) you need to know a lot already to do this, and (b) not seeing the expected pattern does not automatically mean your hypothesized mechanism is wrong.
You could also use what you (think you) know about biological mechanism to improve estimation precision when your sample size is small—e.g. if you are interested in genotype-specific treatment effects.
.@TimothyLash gave a nice example of this involving tamoxifen pharmacogenetics in a recent @HarvardEpi seminar. Keep an eye out for that. 
Note that all of the examples I’ve given so far require knowing something about the specific changes specific alleles make to gene products and their likely consequences.
“Maybe this intergenic SNP regulates this nearby gene that might do something related to this thing related to exposure” won’t cut it.
So what about going the other way? Can we make inferences about underlying biology from the shape of the relationship between G and E and outcome Y, l(G,E) Not without additional assumptions. There are too many possible paths from G,E to Y.
This is particularly relevant for genome-wide scams for interactions. Say you run a GWAS and you find a statistical interaction between an exposure and a variant G (function unknown). As much as you might like, you can’t really infer much about the function of the variant.
Even where you know the exposure acts thru process X, you can’t automatically conclude that G affects X. Best case scenario: you come up with falsifiable hypotheses about what G is doing (which you then have to try to falsify).
Maybe testing falsifiable hypotheses coming from the few statistical GxE interactions identified via GWAS to date will yield more biological insights than testing hypotheses from the 1,000s of GWAS-identified marginal genetic associations—or maybe not.
To be clear—there are good reasons to study the joint effects of genes and environment on traits in epidemiologic studies. But inferring underlying biology is the heaviest lift.
In general, “statistical interaction” per se does not necessarily directly address the scientific goals of the study.
The form of the analysis and its interpretation should be driven by the scientific question. Horses for courses. \fin
