Profile avatar
vincentab.bsky.social
Prof. Most tweets about R. “Polisci, it’s all about what’s going on.” http://arelbundock.com
602 posts 5,707 followers 1,336 following
Regular Contributor
Active Commenter
comment in response to post
One common exception might be factorial experiments in convenience samples, where you want treatment conditions to have the same weights. Finally, note that average quantities have nice mapping to causal quantities in many "traditions"
comment in response to post
If there's any heterogeneity at all and your sample is somewhat representative of a population of interest, I think you should average. Personally, I never create manual grids unless there's fabulous (non-lazy) reasons to care about specific predictor values.
comment in response to post
Consider this vignette. Often, there is massive variation in predictions (or contrasts, risk ratios, slopes, etc.), depending on where you evaluate the quantity on the predictor space. You can pick an arbitrary point/individual, but it won't be representa of pop. marginaleffects.com/bonus/logit....
comment in response to post
We should have orders of magnitude fewer “theories” in PS than we do papers.
comment in response to post
This norm rewards (sometimes dubious) claims of theoretical novelty, rather than encouraging us to build on each other in a cumulative way. They say: send empirical pieces to “specialty journals.” But top journals like AJPS make or break careers, so the incentives are strong (and bad, I think). 2/2
comment in response to post
lol. My favorite new one is this `estimator` argument that facilitates bootstrapping multi-step estimation procedures: marginaleffects.com/bonus/bootst...
comment in response to post
yeah, that's what I meant. I've been trying out claude code and it's kind of amazing (at some things).
comment in response to post
There's lots of stuff to learn, but here's the gist of it
comment in response to post
It'll even install python for you if you don't have it.
comment in response to post
cosign on polars and uv. Fixed the two major reasons I had given up many years ago.
comment in response to post
Thankfully, boilerplate is getting cheaper and easier by the day.
comment in response to post
Also simulating linear systems for DAGs, of course
comment in response to post
Another fun intro example is to have students do Frisch–Waugh–Lovell as two regressions w/ residuals. I find it helps them understand the logic of "control."
comment in response to post
lol
comment in response to post
I don't have reading to recommend, but here's a blog post where I did simulation-based power calculations for g computation estimates: arelbundock.com/posts/money_...
comment in response to post
Sorry about that! We're headed toward version 1.0.0. I'm not planning any big changes until release, and the interface will become very stable from then on.
comment in response to post
Aah, thanks for asking! Website says November, but I've not received official confirmation from the publisher yet, and pre-orders are not yet possible. I'll post incessantly about it when it's available 😉 www.taylorfrancis.com/books/mono/1...
comment in response to post
This looks fascinating! I can't wait to take a closer look!
comment in response to post
Feels like a plot by JSS to artificially pump up their impact factor.
comment in response to post
lol. 🤞
comment in response to post
That's far too sweeping a critique. There are many good use cases for surveys even in low information environments.
comment in response to post
I like that one. It's one of the clearer treatment of the relationship between collapsibility and generalizability/transportability.
comment in response to post
Oh cool. Still, not sure I should impose on a stranger to explain their work to a neophyte like me. 🤷
comment in response to post
Right. But in this Popperian account, severity is not affected by p-hacking, because the friendly-hostile community of critical scientists can always determine the "sincerity" of a test. If that's the world where pre-reg doesn't matter, I don't feel super comforted.
comment in response to post
Not sure I get it. Is the key argument that pre-reg doesn't matter (a) in the Popperian view if science is self-correcting via a community of friendly-adversarial scientists, and (b) in the Mayo view when people don't follow their pre-reg?
comment in response to post
lol yeah, I had forgotten about that! It was my "learn python" project in grad school. Feels like forever ago
comment in response to post
And a real downer for new (better software). Status quo bias is going to be massive.
comment in response to post
Félicitations!!
comment in response to post
Yeah, this is a very cool graph.
comment in response to post
Yep. I need to release `tinytable` before the next version of `modelsummary`, but you can install both from Github now. Sorry for the issue. Please feel free to open a bug report on Github next time you see something weird like that.
comment in response to post
No no. The nudge was very useful. I completely forgot about this when I updated the site.
comment in response to post
Sorry about that. Link is back up here: arelbundock.com/posts/dt_tb_...
comment in response to post
Wow that Danish looks craaazy. (And I'm happy you got to see that show! Must have been cool.)
comment in response to post
it's a good one
comment in response to post
I like this frame a lot! And I'll add another incentive category (what I personally care most about): lowering the cost of good science.
comment in response to post
Right. I'm sure this happens. In such cases, I feel one benefit of prereg is that it turns a routine behavior into something that feels obviously unethical. And I think many people will feel bad / refrain. But maybe I'm too optimistic. Again, I'm super interested to see more systematic evidence.
comment in response to post
Interesting case, and I appreciate your work in that space! But I wonder if the prior you are talking about is one about the bias of norm entrepreneurs, or about the effect size of a given intervention.
comment in response to post
I'd love to see more systematic evidence! FWIW, I've been in the room many times when audience members recommended some weird fishing/phacking trick, and the author said "I wish I could but we didn't preregister that." That's obv weak, unsystematic evidence, so the prior is still doing work for me