Andrew Mercer
@awmercer.bsky.social
6.9K followers 1.3K following 4.1K posts
Principal methodologist at Pew Research Center. He/him
Posts Media Videos Starter Packs
awmercer.bsky.social
Fantastic show at the 9:30 club tonight! You guys killed it! ✌️
awmercer.bsky.social
See, to me it seems pretty harmless. The use cases that baffle me are the ones that could get you sued for malpractice or disbarred if it screws up.
awmercer.bsky.social
Oh, see this totally makes sense to me. Presumably they read all the guidebooks and travel websites, and the stakes are generally pretty low if it gives bad suggestions.
awmercer.bsky.social
Terrific piece from @gelliottmorris.com this morning.

Whenever you ask if something “might” happen or “may” be necessary, it’s very easy for respondents to think of scenarios where the answer is yes, no matter how unlikely.
Why most polls overstate support for political violence
Misperceptions about the popularity of violence increase public support for it — but you can help change that.
www.gelliottmorris.com
awmercer.bsky.social
Depending on how you plan to distribute it, neither would require you to collect anything identifying (like email addresses), although if you are planning to post a public link, there will probably be a lot of spam/bots no matter what.
awmercer.bsky.social
SurveyMonkey and Alchemer are both decent with the former being easier but less flexible and the latter costing a bit more. Qualtrics is very expensive unless you work at a university and get it for free.
Reposted by Andrew Mercer
johngramlich.bsky.social
The US Supreme Court returns today for its new term, two decades after John Roberts became chief justice. As this chart shows, public views of the court are much more negative – and far more politically divided – than 20 years ago. www.pewresearch.org/short-reads/...
Line chart showing how public opinion of the U.S. Supreme Court has changed in recent decades, both among Americans overall and among Republicans and Democrats. The chart is based on Pew Research Center surveys of U.S. adults.
awmercer.bsky.social
If you must, search for “bigot neuron”, but don’t say you weren’t warned.
Reposted by Andrew Mercer
proptermalone.bsky.social
we should invent a Discourse that isn’t increasingly furious people talking past each other
awmercer.bsky.social
This is 100% correct. At this point I would pay extra to get rid of copilot.
awmercer.bsky.social
If you’re lucky, you have either blocked or been blocked by one or more of the relevant parties.
awmercer.bsky.social
Oh I meant “deep cut” in the sense that he was referencing an obscure bit of internet lore from 2008 but that works too.
awmercer.bsky.social
That’s a pretty deep cut but also yes.
awmercer.bsky.social
So tired of all the helicopters. So many helicopters.
awmercer.bsky.social
Agree about target trials. I guess I see Pearl and Robins as connected they’re coauthors some papers related to DAGS and confounding. But you’re right that Robins uses counterfactuals. It’s messy.
awmercer.bsky.social
Published online. It was originally published in 2000.
awmercer.bsky.social
I guess my point is it’s unphysical concepts all the way down. But for some reason people really don’t like this one in particular.

Sure, it’s not the only way to frame causal inference, but I’m not sure it’s inherently worse. Different frameworks mostly emphasize different sets of priorities.
awmercer.bsky.social
If you replace “counterfactual” with “predicted value conditional on having received one treatment or the other” does that solve the metaphysical problem? Predictions aren’t physical either but people seem more comfortable with them.
awmercer.bsky.social
The estimand is the same (population mean, CATE, ATT, whatever), right? I see it more about how you set up your estimator.
awmercer.bsky.social
In sampling though, if you’re willing to embrace unphysical constructs, you can get much more efficient estimates (assuming your model is well specified). This was actually Rubin’s original motivation for the propensity score: improving the efficiency of estimates in RCTs.
awmercer.bsky.social
Ultimately it’s just a way to frame the problem that may be more useful in some contexts than others.

I’ve found the notions of ignorability, etc… helpful in thinking about the problems I deal with, but I can allow that it may be less helpful in other areas.
awmercer.bsky.social
I mean that’s true, but that’s also true for multiple draws from a superpopulation. I can envision it conceptually but it can’t actually exist. I can also envision an alternate universe where someone got a different treatment, even though it didn’t happen.
awmercer.bsky.social
I mean, regression coefficients aren’t “real” either. We don’t have to take any of this stuff literally.