Michael Heseltine
michaelheseltine.bsky.social
Michael Heseltine
@michaelheseltine.bsky.social
Post-Doc at University of Oxford. Congress, text-as-data, and political communications.
And here's Facebook. Really does seem to be a Covid term!
August 26, 2025 at 3:33 PM
Size matters: LLama 3.1/3.3 70b perform best, along with Qwen2/2.5 72b. Gemma2 and Phi4 stand out as strong mid-sized models. Many smaller models produce poor results. Model selection can have a huge impact on your results!
August 12, 2025 at 4:28 PM
Just started using atrrr for search and user post collection. Firehose would be a massive plus!
August 3, 2025 at 2:41 PM
113th congress starts in '13 is my go-to shorthand!
July 2, 2025 at 1:18 PM
Fresh email this morning: "an account has been created for you in ScholarOne", nooooo!
June 30, 2025 at 11:58 AM
Bit chicken and egg though, did researchers adjust to stargazer or did stargazer adjust its defaults to perceived research practices
May 1, 2025 at 9:26 AM
My hunch would be at least some creep after the rise of stargazer as a default tabulating method, with its default of * for 0.1.
May 1, 2025 at 9:20 AM
Accept it with arms wide open, Kevin...
March 25, 2025 at 5:28 PM
Worked in three reps offices. Member 1 was given weekly tally of calls, never saw emails. Member 2 given no information on either. Member 3 informal updates on calls, never saw emails.
99% is just fielded by interns. Up to the individual Member whether they care to receive any of it.
March 14, 2025 at 3:54 PM
An extra 128gb ram and any 16gb or 24gb Nvidia gpu
... Alternatively, an absolute tonne of gaudy strobe rgb fans
March 11, 2025 at 8:35 PM
@sysilviakim.bsky.social
Coming to a JOP issue near you!
March 11, 2025 at 5:49 PM
I think @jbgruber.bsky.social is the expert on this?
February 26, 2025 at 12:05 PM
Science!
February 9, 2025 at 4:39 PM
Reposted by Michael Heseltine
R&P editors' collection 2024 (2/4)

@michaelheseltine.bsky.social & B. C. von Hohenberg find that GPT-4 coding of political text is highly accurate for shorter texts, but performance drops for longer ones, suggesting consideration should be given to task complexity when using LLM-assisted coding.
Sage Journals: Discover world-class research
Subscription and open access journals from Sage, the world's leading independent academic publisher.
doi.org
January 11, 2025 at 3:31 PM
Also... the social sciences have absolutely destroyed the respondent pool in convenience sample pools around the world through the over-proliferation of survey research. Not sure how to fix that, but running even more surveys not likely the answer.
December 5, 2024 at 4:26 PM
Very feasible scenario is: 1) grad student begs for some money and runs survey and writes paper. 2) reviewer or journal policy says "that's nice, now run it again so I believe you" 3) grad student goes back to begging, or pulls paper?
Massive chunk of the discipline is priced out of experiments atm.
December 5, 2024 at 4:24 PM
Does every self-citation I've ever done count?
November 27, 2024 at 9:49 PM
Having one of the top-tier generalist journals in the discipline being almost unusable for early career scholars is sub-optimal for the discipline, to put it mildly.
November 27, 2024 at 9:08 PM