6 ms·
For those who arent aware, this entire field is largely pseudoscience; though this paper goes out of its way to be fairly sincere in its modelling, it's still j
by mjburgess 2y ago
For those who arent aware, this entire field is largely pseudoscience; though this paper goes out of its way to be fairly sincere in its modelling, it's still just nonesense methodologically.
The correlation coefficient is not a measure of correlation, but of co-linearity, and coefficients < 0.5 are typically bunk, and the entire system of computing correlation coefficients is bunk if the data is nonlinear. Almost all interesting phenomena outside are non-linear, esp. biology and higher, where nonlinearities arise from the mutual interaction of parts of the system. Here, the correlation coefficients are in the ~0.2 range. You might as well be staring at monkey entrials.
Methodologically, you cannot construct explanatory models of weak-effect non-linear phenomena from observational data. Basically, you can fit any explanatory model you like, since you can make any parts of the system interact with any strength, and since they are non-linear, this will reproduce any distribution you so wish.
You can entirely reproduce any heritability distribution, genetic covariance, "shared" parent-child, "unshared" child-child, etc. you want by changing this model. Observational data here is basically useless at discriminating. ie., i can make a model where genes are 100% irrelevant, or 100% determinative, entirely consistent with the observational data.
The only method which can distinguish here is interventional, ie., you have to actually control the causes of the system. However, since we cannot breed different groups with different genes; nor take the same person and run their life with differnet parents, friends, etc. you're basically out of luck.
I'd prefer we closed this whole field down, and any person mentioning "heritability" outside of a wheat breeding lab, shuffled off to some discipline less catastrophically detrimental to social policy.
- mucle6 2y agoCan you share some resources or topics to research for someone who doesn't know what a weak effect or co-linearity? If I google those topics, I don't know if I'll find anything relating to a criticism of this field.
- mjburgess 2y agoSpecifically for the area mentioned: https://gwern.net/doc/genetics/heritable/2015-burt.pdf https://gwern.net/doc/genetics/heritable/2015-burt.pdf But if you want a general criticism of these academic linear methods of observational analysis: https://arxiv.org/abs/2001.10488 https://arxiv.org/abs/2001.10488
- mucle6 2y agoDidn't expect to see statistical consequences of fat tails! I bought that book and it was soooooo dense. I had trouble keeping up with it even after rereading it. Its basically a textbook
- mjburgess 2y agoRead the chapter, "a non-technical overview - the darwin college lecture" He's quite a bad communicator; that happened to be written by students attending his lecture.
- whimsicalism 2y agoYou're clearly arguing from the conclusions you want to be true. Nobody who is in the know and actually is skeptical of this branch of work (ie. people like Sasha Gusev) thinks it is as useless as you are making it out to be. Adoption studies produce causality and even the GWAS (genome-wide association studies) stuff, while not necessarily causal, is pretty interesting.
- mjburgess 2y agoI'm presenting using heritability as an initial toy problem area for exploring building PPL/agent-based scenario models. I havent yet got to the point of building the model out, but it's clear very early on that any observational distribution can be reproduced. I will have a more refined view when I've done that. re Sasha Gusev: disease, to some degree, can be studied with these methods if you have interventional evidence of a genetic cause. But since all non-genetic factors are almost obviously mutually interacting, with "compound interest", no explanatory model can be fit. eg., suppose your personality is entirely genetic: then why variation? Well, because of gene-environment interactions. eg., suppose your personality is 0% genetic: when why genetic co-variation? Because people's genes covary geographically, and due to the trait selection of partners... ie., families live in the same geography, and people of similar traits breed together. Why (very weak) correlation with genetic similarity? Because "compound interest" applies to the interacting factors, so even extremely weak geographical selection will compound. It is entirely unclear to me how you're meant to discriminate between teh infinite family of models on either side. Certainly, with correlation coefficients of ~.2, you're talking about pseudoscience either way. The causes would have to be extremely linear for such trivial correlations to mean anything at all, and they certainly arent.
- whimsicalism 2y agoI understand your objection and of course you can't just naively get causality from observational GWAS. That is true. Your rhetoric does not match the strength of your scientific objection. Let's imagine the worst case - these studies are absolutely useless for indicating any causality whatsoever. Even then, they are still extremely useful because they can explain away lots of environmental variation in a behavior of interest and increase the power of studies of human behavior. That's if they are purely non-causal and genetics actually has no impact whatsoever on behavior or the brain (a mechanistic stretch). Sorry, but this simply does not match your rhetoric of 'pseudoscience' and closing entire fields down. > Certainly, with correlation coefficients of ~.2, you're talking about pseudoscience either way. The causes would have to be extremely linear for such trivial correlations to mean anything at all, and they certainly arent. I simply don't agree, but I can tell you won't be convinced on this point.
- em-bee 2y agoas i mentioned in another comment, what bothers me the most is that heritability is used as an excuse for not changing certain traits and and as a reason to demand the right to continue bad behaviors
- mjburgess 2y agoWell, remember: heritability of scottish accents is near 100%, and heritability of two eyes is near 0%. Why? The covariance of genes with accent is near 100%: almost everyone who speaks scottish interbreeds within the same geographical area. As for eyes: there is almost no variation in the species between the number of eyes. Since ~everyone has 2, no genetic variation whatsoever discriminates. The co-variation of a gene with an observable trait is irrelevant when vast amounts of cultural (, geographical, social, etc.) traits cause (irrelevant) genetic variation. So "heritability" is basically nonsense outside of cases when the experimenter is actually causing genes to differer between populations (it was invented for the case of engineering crops).
- elevatedastalt 2y agoWill a kid born to Scottish parents who was separated at birth and raised by Chinese parents in China also speak in a Scottish accent? Or are you just giving insanely strawmannish arguments no one actually presents?
- whimsicalism 2y agothey’re making a (correct) point about causality and geographic gene variance
- elevatedastalt 2y agoNo, they are giving an absurd example in order to create the implication that heritability is a meaningless concept.
- 2y ago