Josh Pasek

@joshpasek.com

Prof of Comm and Polisci @UMich studying how people get & use political information and social measurement. Competencies: data sci, DIY solar, election analytics, carpooling, #polcom, #polpsych, survey methods, AI literacy (views=own, overuses speech2text)

The big problem though is that weighted inferential statistics are no longer BLUE - there are reasonable approaches to figuring out how big and consistent the difference is after weighting, but I very rarely see them used. Running the weighted regression, however, is usually not appropriate.

Drew Engelhardt@amengel.bsky.social · 2mo ago

Love puzzles with survey weights. There's a well cited paper in a top 3 political science journal that uses an unweighted YouGov sample. When you actually weight, the takeaway isn't that clear. I also know of a paper that "weighted" by using the weights a predictor variable in its regressions.

The critical question of US politics right now is whether the reticence of normally solid Republican voters to pick up the phone and talk to pollsters signals a general disengagement with politics (and thus low turnout in November), or is limited to polling.

1/ So as is my tradition, I was reading the Declaration of Independence today and I was thinking that instead of asking what it says about King George, it might be more interesting to think about what kind of country the founders were envisioning. A 🧵

"We hold these truths to be self-evident, that all men are created equal, that they are endowed by their Creator with certain unalienable Rights" "to secure these rights, Governments are instituted among Men, deriving their just powers from the consent of the governed"

There’s something quite fitting and ironic about celebrating the nation’s 250th birthday with a two day long power outage. An inspiring anniversary dashed by degraded infrastructure…

Power outage map showing an outage that started at 6 PM on July 3 is expected to be restored at 11:30 PM on July 5.

Just inadvertently discovered that a whole bunch of research on AI responses may be invalid. Apparently, even with memory off, ChatGPT is still giving contextual information, including the name of the user! It used my full name casually in conversation; which means that data is in the model.

I asked: Please write a summary from your perspective as an AI on what happened here and keep it short  - Chat response: From my perspective as the AI:

During this conversation, I referred to the user by name even though long-term memory was disabled. I did not retrieve that information from prior chats, but I did have access to hidden contextual metadata that identified the user, and I used it conversationally when I shouldn’t have. That revealed that my responses were conditioned on information beyond the visible transcript. The user pointed out that, regardless of whether the information came from memory, metadata, or retrieval, hidden conditioning changes the effective inputs to the model. For researchers, that creates a reproducibility problem because the conversation transcript alone is no longer a complete specification of what produced the response.