|
Adam Green
@
adamlewisgreen
|
|
Interested in behavioral and medical genetics, statistics, bioethics, jazz, kettlebells, ev bio, etc. Attempting to improve my model of the world.
|
|
|
229
Tweetovi
|
510
Pratim
|
115
Osobe koje vas prate
|
| Tweetovi |
|
Adam Green
@adamlewisgreen
|
2. velj |
|
In all seriousness, I think the extemporaneous nature of jazz is what distinguishes from other musical, and maybe all, art forms.
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
Shakespeare ain’t got nothing on Michael Brecker. youtu.be/tbYl_B3YINQ
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
Cool, I'll check it out. I'm just curious how and why you make the jump from your results to "these two findings really undermine narratives that low-SES people do not attain higher education because they are not capable". Maybe I'm missing something...
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
So, if SES measures + parental education capture most c^2; and if there is a lot of a^2 not captured by the PGS: then conditioning on parental education underestimates difference in mean latent PGS (what the PGS would look like if r^2 = h^2_SNP) between high and low SES groups
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
See the crude, unfinished structural causal model below. The green represents everything you eliminate when conditioning on parental EA (which includes some c^2 not captured by SES measures). Unidirectional arrows represent causal effects; dotted double-arrows correlations (3/n) pic.twitter.com/edshLVvnnz
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
You observe SES-group differences in PGS that go away after conditioning on SES and parental education. This conditioning eliminates some shared environmental effects not captured by SES, which is why you do it. But it also eliminates unmeasured (non-PGS) genetic effects (2/n) pic.twitter.com/RrcA34uDSD
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
2. velj |
|
This inference—that the distribution of "genetic capability" is largely the same across SES groups, and therefore, if we were to equalize SES, between-group differences in educational attainment would decrease substantially—isn't justified based on the data you present (1/n)
|
||
|
|
||
| Adam Green proslijedio/la je tweet | ||
|
Jordan Smoller
@jorsmo
|
30. sij |
|
Integration of polygenic risk scores with modifiable risk factors improves risk prediction: results from a pan-cancer analysis biorxiv.org/content/10.110… pic.twitter.com/WDbxqyFAPp
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
31. sij |
|
Cue the Jackson Browne soundtrack...
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
29. sij |
|
Evidently "neobehaviorist" doesn't mean what I thought it does. Perhaps better language is a "blackbox approach" or "Skinnerian approach"
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
Joint-carving (i.e., explicitly representing our model of the world) is only useful insofar as we need to grok these models. But for many use cases, this isn’t necessary. A black box gets the job done.
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
If social sci is ultimately about prediction and modification of behavior (from level of atoms to groups), not joint-carving (which is for the philosophers), then it seems inevitable that ML will soon supersede top-down psych theorizing in this game.
nature.com/articles/s4159…
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
Spicy take. I think ML won’t result in massive changes to our existing psychological ontology until more types of real-world data are collected (not just movie and shopping preferences). When they are, we’ll see a neobehaviorist revolution in social sci
biorxiv.org/content/10.110… twitter.com/talyarkoni/sta…
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
Funny exercise: compare the SE’s of the balance checks to those of the main effects (this might pan out differently when the dependent var is a proportion, though, as it’s bounded by [0,1]):
mobile.twitter.com/jd_wilko/statu…
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
You suggest some of the EA ~ PGS x SES effect might operate through # of books in the home, preschool, and other early investment. If true, wouldn’t we expect to see these sorts of interventions produce consistent, long-term, large effects in RCTs (which we mostly don’t)? pic.twitter.com/Q90MRtUakD
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
Glad to see this paper finally out!
- You say EA score distribution is largely similar across SES groups. What about in the far right tail? Graphs below
suggest potentially big differences in odds ratio. Not sure how to answer this question in a non-parametric (KDE) context (1/2) pic.twitter.com/Rj9W1lllQO
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
This is how the RDD treatment procedure worked. The paper seems pretty convincing, I must say, especially given all the existing literature on the causal lead-crime connection. pic.twitter.com/7vDUAUapnI
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
28. sij |
|
Interesting. 2 concerns:
- Does RDD treatment "randomization" process + controlling for covariates eliminate all selection bias? (treatment: receive 2 blood tests with BLL > 10 µg; control: only 1 test > 10µg)
- Treatment effect heterogeneity: effect only present in >20µg group? pic.twitter.com/HFk7enekYF
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
27. sij |
|
Combine with career incentives (I've built my career studying topic X) and other psychological incentives (my ego is tied up in my work and I want to believe it is meaningful), and you get sclerotic areas of science that don't track reality.
statmodeling.stat.columbia.edu/2020/01/24/the…
|
||
|
|
||
|
Adam Green
@adamlewisgreen
|
27. sij |
|
I think many incentives are psychological. Consider education research: one has deep ethical commitments (egalitarianism) and develops a hypothesis around them (performance diffs are due to unequal opportunity). If hypothesis is falsified, it challenges your model of the world.
|
||
|
|
||