Dan Zeltzer (@danzeltzer) header
Dan Zeltzer (@danzeltzer) profile picture

Health Economist @UCBerkeley Public Health. Studying digital health - telemedicine, AI. IV hobbyist. Ex-Tel Aviv U Econ. Princeton Econ PhD

Berkeley CA
Joined October 2013
Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)Photo by Dan Zeltzer (@danzeltzer)
๐ŸšจNew in AER๐Ÿšจ: "Negative Control Falsification Tests for IV Designs" (@DanieliOren, @DanielNevo, Itai Walk, Bar Weinstein, @danzeltzer). doi.org/10.1257/aer.20240636 Common placebo test implementations may reject valid IVs. We're coming to save your IVs! ๐Ÿงต
1
50
3
170
16,823
Dan Zeltzer retweeted
Most studies of #bias in healthcare #LLMs find the problem. Very few try to fix it. Our new article in @LancetDigitalH maps 201 studies on bias and equity in LLM applications for health care. There are several key gaps: โœ… Most evaluations rely on exam questions or synthetic cases, not real clinical data โœ… Most focus on proprietary models, limiting how far bias can be traced to its source โœ… Mitigation is rare, and usually limited to prompt changes โœ… Few studies explore how LLMs could actively advance health equity The field has become good at spotting bias in model outputs. The next step shall move upstream and closer to the clinic: evaluation grounded in real workflows, shared standards, bias governance across the model lifecycle, and real investment in low-resource languages and settings. Congratulations to the team and thanks to our collaborators across Duke-NUS, NUS, Duke, Penn, Yale, Stanford and Harvard. ๐Ÿ“œRead paper here: thelancet.com/journals/landiโ€ฆ #HealthEquity #LargeLanguageModels #AI #AIinHealthcare #MedicalAI #DigitalHealth #DukeNUS
7
11
494
Our NEJM AI Perspective on the physical limits to AI substitution drew many reactions (which we welcome!). A few comments:
The authors analyze the degree to which clinical specialties are shielded from, or threatened by, AI depending on how much of their work requires in-person physical interaction. Learn more: nejm.ai/3UDkrIv
9
5
3
26
21,193
In sum, AI's potential to replace physicians' work is limited by the physical nature of many specialties.
2
1
1,193
Substituting some work doesn't mean replacing physician specialists - AI substituting certain non-physical specialist tasks can augment their physical work. How AI impacts medicine ultimately depends on how technology is designed, deployed, and paid for.
1
1
1,182
Dan Zeltzer retweeted
I'm looking for a postdoc to work on AI for science and related topics for the 2027-28 academic year. Details here: aeaweb.org/joe/listing.php?Jโ€ฆ
1
46
2
121
15,417
Dan Zeltzer retweeted
My friend sent me this great problem: A random card is drawn from a deck, which you will try to guess. But first, you can ask 1 yes/no question. Is it better to ask if the card is red or if the card is the ace of spades?
236
77
57
3,939
1,200,931
Dan Zeltzer retweeted
Please enjoy! Writeup: dropbox.com/scl/fi/ago5vsi89โ€ฆ Repo: github.com/peterdhull/twfe-rโ€ฆ Planning to expand to other journals soon
Why has no one done the massive AI-assisted "how much do negative weights really matter for difference-in-difference" review paper yet?
9
120
22
534
96,904
Dan Zeltzer retweeted
Looking for postdocs+PhDs to help lead a large-scale study of the effects of long-term AI use, a collaboration between Berkeley+@TransluceAI+others! Funding+positions available (postdocs/visiting researchers/Transluce affiliations). Please reshare; application in next tweet!
6
56
1
254
23,943
Dan Zeltzer retweeted
๐ŸšจNew paper!๐ŸšจWho orders more testing: male physicians or female physicians? And what are the implications on patient outcomes? We examine these questions in the emergency department (ED) in @JAMAInternalMed.๐Ÿงต1/
2
5
14
5,204
Dan Zeltzer retweeted
ืื ื™ ืžื—ืคืฉ ืขื•ื–ืจ.ืช ืžื—ืงืจ ืœืคืจื•ื™ืงื˜ ืขืœ ื™ืฆื™ืื” ืžืขื•ื ื™. ืื ื™ ืžืคืชื— ืžืชื•ื“ื•ืœื•ื’ื™ื” ืืžืคื™ืจื™ืช ื—ื“ืฉื” ื•ืžื—ืคืฉ ืกื˜ื•ื“ื ื˜.ื™ืช ืžืฆื˜ื™ื™ืŸ.ืช ืขื ื™ื›ื•ืœื•ืช ื˜ื•ื‘ื•ืช ื‘ึพR ื•ื‘ืขื‘ื•ื“ื” ืขื ื›ืœื™ AI, ืฉื™ืขื–ื•ืจ ืœืžืžืฉ ืื•ืชื”. ื”ืขื‘ื•ื“ื” ื‘ืขื™ืงืจ ื‘ื™ื•ืœื™โ€“ืกืคื˜ืžื‘ืจ, ืœืคื—ื•ืช 15 ืฉืขื•ืช ื‘ืฉื‘ื•ืข, ื•ืชื™ืžืฉืš ื’ื ื‘ืกืžืกื˜ืจ ืืณ. ืงื•ืดื— ื•ื’ื™ืœื™ื•ืŸ ืฆื™ื•ื ื™ื: [email protected]
3
19
4
31
11,397
Dan Zeltzer retweeted
When are causal parameters identified in instrumental variables models? This paper shows many such models share a common necessary and sufficient condition for identification that is constructive and suggests a simple sample-analogue estimator. econometricsociety.org/publiโ€ฆ
45
1
204
18,619