4 ms·

You misunderstand what mendelian randomization and instrumental variable estimation are. I'm not even sure what the methods you propose are, they're definitely

by et2o 8y ago

You misunderstand what mendelian randomization and instrumental variable estimation are. I'm not even sure what the methods you propose are, they're definitely not popular ways to do causal inference in medicine or epidemiology.

AstralStorm 8y ago
Sorry, but I actually understand precisely what mendelian randomization and instrumental variable estimation are. You apply nonlinear least squares estimation to every selected suspected variable which is a bunch of linearization/polynomial approximating cross correlations with normalized by estimated covariances. That is nonlinear IV method. MR is this based on an assumption that genomes are random fields of genes with equal independent probabilities of occurence (known to be false, partly corrected by population statistics which is not enough) giving direct results on marker conditions/correlates but not on endpoints (known to be false, partial attempts to correct for it are made to check for pleiotropy but those checks are weak and linear) which markers are nonlinearly covariate with endpoints (hope you used the right mapping) and that highest magnitudes of the "correlations" say something about causation related to certain endpoints. (not in general) MR shines as a negative method to rule out genetic causation or susceptibility, not positive method to confirm it. It can be used to confirm internal validity of a marker - invariance of it vs genetics. The choice of variables (dimensions) is an educated guess too. The assumptions that are verified is that genetic sequence is not heterogenous, heteroskedastic, there are no hidden linkages etc. The assumptions are not typically met or verified for complex system such as lipid metabolism alone, much less endpoints such as CVD. Polygenetic phenotypes are common and typical, not to say anything about epigenetics. Nothing is said about the choice of the checked variables which is critical, and trying too many gives the problem of multiple measurement reducing power. The strongest assumption is that the variables are not covariant with estimation error. This is not possible to check with MR statistics alone unless you can somehow check those variants experimentally directly. (e.g. via gene splicing or mutation) Good luck with that.
- et2o 8y ago
  No, I don't think you have a clue. You wrote a completely wrong thing, edited it when I pointed out it was incorrect, and then 4 hours later you read up about MR on wikipedia or something and came back here and posted a partially correct answer with a lot of buzzwords and nice-sounding phrases, but with even more straw-man arguments and outright errors. Peace out, I'm done with you.

by et2o 8y ago

AstralStorm 8y ago

Sorry, but I actually understand precisely what mendelian randomization and instrumental variable estimation are. You apply nonlinear least squares estimation to every selected suspected variable which is a bunch of linearization/polynomial approximating cross correlations with normalized by estimated covariances. That is nonlinear IV method. MR is this based on an assumption that genomes are random fields of genes with equal independent probabilities of occurence (known to be false, partly corrected by population statistics which is not enough) giving direct results on marker conditions/correlates but not on endpoints (known to be false, partial attempts to correct for it are made to check for pleiotropy but those checks are weak and linear) which markers are nonlinearly covariate with endpoints (hope you used the right mapping) and that highest magnitudes of the "correlations" say something about causation related to certain endpoints. (not in general) MR shines as a negative method to rule out genetic causation or susceptibility, not positive method to confirm it. It can be used to confirm internal validity of a marker - invariance of it vs genetics. The choice of variables (dimensions) is an educated guess too. The assumptions that are verified is that genetic sequence is not heterogenous, heteroskedastic, there are no hidden linkages etc. The assumptions are not typically met or verified for complex system such as lipid metabolism alone, much less endpoints such as CVD. Polygenetic phenotypes are common and typical, not to say anything about epigenetics. Nothing is said about the choice of the checked variables which is critical, and trying too many gives the problem of multiple measurement reducing power. The strongest assumption is that the variables are not covariant with estimation error. This is not possible to check with MR statistics alone unless you can somehow check those variants experimentally directly. (e.g. via gene splicing or mutation) Good luck with that.

et2o 8y ago
No, I don't think you have a clue. You wrote a completely wrong thing, edited it when I pointed out it was incorrect, and then 4 hours later you read up about MR on wikipedia or something and came back here and posted a partially correct answer with a lot of buzzwords and nice-sounding phrases, but with even more straw-man arguments and outright errors. Peace out, I'm done with you.