dynomight
[Question] Why has the replication crisis affected RCT-studies but not observational studies?
I loved this book. The most surprising thing to me was the answer that people who were there in the heyday give when asked what made Bell Labs so successful: They always say it was the problem, i.e. having an entire organization oriented towards the goal of “make communication reliable and practical between any two places on earth”. When Shannon left the Labs for MIT, people who were there immediately predicted he wouldn’t do anything of the same significance because he’d lose that “compass”. Shannon was obviously a genius, and he did much more after than most people ever accomplish, but still nothing as significant as what he did when at at the Labs.
I thought this was fantastic, very thought-provoking. One possibly easy thing that I think would be great would be links to a few posts that you think have used this strategy with success.
Thanks, I clarified the noise issue. Regarding factor analysis, could you check if I understand everything correctly? Here’s what I think is the situation:
We can write a factor analysis model (with a single factor) as
where:
is observed data
is a random latent variable
is some vector (a parameter)
is a random noise variable
is the covariance of the noise (a parameter)
It always holds (assuming and are independent) that
In the simplest variant of factor analysis (in the current post) we use in which case you get that
You can check if this model fits by (1) checking that is Normal and (2) checking if the covariance of x can be decomposed as in the above equation. (Which is equivalent to having all singular values the same except one).
The next slightly-less-simple variant of factor analysis (which I think you’re suggesting) would be to use where is a vector, in which case you get that
You can again check if this model fits by (1) checking that is Normal and (2) checking if the covariance of can be decomposed as in the above equation. (The difference is, now this doesn’t reduce to some simple singular value condition.)
Do I have all that right?
Thanks for pointing out those papers, which I agree can get at issues that simple correlations can’t. Still, to avoid scope-creep, I’ve taken the less courageous approach of (1) mentioning that the “breadth” of the effects of genes is an active research topic and (2) editing the original paragraph you linked to to be more modest, talking about “does the above data imply” rather than “is it true that”. (I’d rather avoid directly addressing 3 and 4 since I think that doing those claims justice would require more work than I can put in here.) Anyway, thanks again for your comments, it’s useful for me to think of this spectrum of different “notions of g”.
Thanks, very clear! I guess the position I want to take is just that the data in the post gives reasonable evidence for g being at least the convenient summary statistic in 2 (and doesn’t preclude 3 or 4).
What I was really trying to get at in the original quote is that some people seem to consider this to be the canonical position on g:
Factor analysis provides rigorous statistical proof that there is some single underlying event that produces all the correlations between mental tests.
There are lots of articles that (while not explicitly stating the above position) refute it at length, and get passed around as proof that g is a myth. It’s certainly true that position 5 is false (in multiple ways), but I just wanted to say that this doesn’t mean anything for the evidence we have for 2.
Can I check if I understand your point correctly? I suggested we know that g has many causes since so many genes are relevant and thus f you opened up a brain, you wouldn’t be able to “find” g in any particular place. It’s the product of a whole bunch of different genes, each of which is just coding for some protein, and they all interact in complex ways. If I understand you correctly, you’re pointing out that there could be a sort of “causal bottleneck” of sorts. For example, maybe all the different genes have complex effects, but all that really matters is how they affect neuronal calcium channel efficiency or something. Thus, if you opened up a brain, you could just check how efficient the calcium channels are and you’re done. Is that right?
If this is right, I do agree that I seem to be over-claiming a bit here. There’s nothing that precludes the possibility of a “bottleneck” as far as I know, (though it seems sorta implausible in my not-at-all-informed opinion)
I used python/matplotlib. The basic idea is to create a 3d plot like so:
fig = plt.figure() ax = fig.add_subplot(111, projection='3d')Then you can add dots with something like this:
ax.scatter(X,Y,Z,alpha=.5,s=20,color='navy',marker='o',linewidth=0)Then you save it to a movie with something like this:
def update(i, fig, ax): ax.view_init(elev=20., azim=i) return fig, ax frames = np.arange(0, 360, 1) anim = FuncAnimation(fig, update, frames=frames, repeat=True, fargs=(fig, ax)) writer = 'ffmpeg' anim.save(fname, dpi=80, writer=writer, fps=30)I’m sure this won’t actually run, but it gives you the basic idea. (The full code is a complete nightmare.)
Thanks for the reply. I certainly agree that “factor analysis” often doesn’t make that assumption, though it was my impression that it’s commonly made in this context. I suppose the degree of misleading-ness here depends on how often people assume isotropic noise when looking at this kind of data?
In any case, I’ll try to think about how to clarify this without getting too technical. (I actually had some more details about this at one point but was persuaded to remove them for the sake of being more accessible.)
Factors of mental and physical abilities—a statistical analysis
if a trait is 80% heritable and you want to guess whether or not Bob has that trait then you’ll be 80% more accurate if you know whether or not Bob’s parents have the trait than if you didn’t have that information.
I think this is more or less correct for narrow-sense heritability (most commonly used when breeding animals) but not quite right for broad-sense heritability (most commonly used with humans). If you’re talking about broad-sense heritability, the problem is that you’d need to know not just if the parents have the trait, but also which genes Bob got or not from each parent, as well as the effect of dominant genes, epistatic interactions, etc.
Assuming you’re talking about broad-sense heritability, I think a better way of looking at it would be to say that you’ll be 80% more accurate if Bob has an identical twin raised by a random family and you know if that twin had the trait. This isn’t quite right either, but I think it’s valid if you assume that phenotypic traits are the sum of genetic effects and environmental effects and also that genetic effects are independent of environmental effects.
Of course, few people have identical twins raised by random families, and most phenotypes probably aren’t additive in genetic and environmental effects, and those effects probably aren’t independent! Which… is a lot of caveats if you want to know practical applications of heritability numbers.
On the other hand, there is some non-applied scientific value in heritability. For example, though religiosity is heritable, the specific religion people join appears to be almost totally un-heritable. I think it’s OK to read this in the straightforward way, i.e. as “genes don’t predispose us to be Christian / Muslim / Shinto / whatever”. I don’t have any particular application for that fact, but it’s certainly interesting.
Similarly, schizophrenia has sky-high heritability (like 80%) meaning that current environments don’t have a huge impact on where schizophrenia appears. That’s also interesting even if not immediately useful.
My view is that people should basically talk about heritability less and interventions more. In most practical circumstances, what we’re interested in is how much potential we have to change a trait. For example, you might want to reduce youth obesity. If that’s your goal, I don’t think heritability helps you much. High heritability doesn’t mean that there aren’t any interventions that can change obesity—it just means that the current environments that people are already exposed to don’t create much variance. Similarly, low heritability means the environment produces a lot of variance, but it doesn’t tell you anything specific you can actually do!
If you goal is to find interventions, all heritability gives you is some kind of vague clue as to how promising it might be to look at natural environmental variation to try to find interventions.
In principle, I guess you could also think about low-tech solutions. For example, people who want to opt out of alcohol might have some slowly dissolving tattoo / dye placed somewhere on their hand or something. This would eliminate the need for any extra ID checks, but has the big disadvantage it would be visible most of the time.
Thanks. Are you able to determine what the typical daily dose is for implanted disulfiram in Eastern Europe? People who take oral disulfiram typically need something like 0.25g / day to have a significant physiological effect. However, most of the evidence I’ve been able to find (e.g. this paper) suggest that the total amount of disulfiram in implants is around 1g. If that’s dispensed over a year, you’re getting like 1% of the dosage that’s active orally. On top of that, the evidence seems pretty strong that bioavailability from implants is lower than from oral doses, so it’s effectively even less.
Of course, there’s nothing stopping someone implanting 100x as large a dose, and maybe bioavailability can be improved (or isn’t that big a concern). But if not, my impression was that most implants are effectively pure placebo effect.
Very interesting! Do you know how much disulfiram the implant gives out per day? There’s a bunch of papers on implants, but there’s usually concerns about (a) that the dosage might be much smaller than the typical oral dosage and/or (b) that there’s poor absorption.
I specified (right before the first graph) that I was using the US standard of 14g. (I know the paper uses 10g. There’s no conflict because I use their raw data which is in g, not drinks.)
Ironically, there is no standard for what a “standard drink” is, with different countries defining it to be anything from 8g to 20g of ethanol.
I wasn’t (intentionally?) being ironic. I guess that for underage drinking we have the advantage that you can sort of guess how old someone looks, but still… good point.
Done!