Isn't the finding of a gene that's reliably associated with a particular lifestyle practice usually a problem though? Even if you know that one gene increases the risk of something, there are probably other genes that might mitigate that risk and others that also increase it (and not all of these genes may be identified yet).
Yea. Not everyone with the gene will exhibit the feature, because of other genes or environment or both. That's assumed to be the case. Its just a probabilistic association. The idea is that you use the gene to divide the population into two groups, and one of the groups has a reliably higher probability of having your behaviour or marker of interest than the other.
Of course the size of the probability difference matters. If its marked, then you have more of a chance of capturing the thing you're interested in. If its small, then it will capture it poorly. But to be fair, these papers always report the strength of the initial association - in the Abstract itself - so the reader can decide for themselves how much weight to put on the results.
The logic part of all this I sort of follow. The genetics bit, not so much. I'm still not sure whether they isolate a single gene, a set of genes or what. And I have no clue how they actually do that part.
It might end up being a bit like the functional MRI's - you can fish around and prove whatever theory you want?
It seems to be mainly used to address questions that have already been asked. You don't just fish around will nilly. Most of the published stuff reports positive results, so there might be a fair bit of publication bias going on there. But there are the odd negative findings which make it into print. And there's a fair few studies that use the technique to find out the direction of an already known association (e.g. does smoking cause anxiety, or does anxiety make you more likely to be a smoker? The answer seems to be the second one).
think even quite a simple mathematical model would start to show this is hard once you include several non-linear functions in the equations. Perhaps you would still get some correlations but I suspect little certainty.
You may be way ahead of me here. But if you're asking whether they control for covariates statistically, then no, they don't. The stats are extremely simple. The method itself is designed to take care of covariates. That's the whole point.
So for example, if you find a reliable fat gene, you assume this gene will vary independently of all other genes across your population. So your two comparison groups should not differ systematically on any other gene other than your fat gene. They should be equally likely to be rich or poor, smokers or nonsmokers. This won't always be entirely true but at least there won't be the same sort of multiple huge confounds between variables you see in conventional epidemiology studies.
It sounds like I'm really in favour of the approach, but I do have some worries (which I'll talk about some other time).