Merry Christmas! 🎅🎄

I remember that when I would urge media to talk more about winning probabilities derived by forecasters/aggregators in their polling coverage, people would tell me the general public would not understand it. I guess the media decided that this is not a problem anymore? www.cnbc.com/2025/12/04/c...

What would once be the dream of any subjectivity Bayesian -- the popularization of using wagering odds as probabilities, is now becoming a real dystopia... 😫

This is some sort of measurement error, just don't ask me which 😅 www.nbcnews.com/news/amp/rcn...

This is an incredible opportunity if you want to work in polling and surveys with some of the best in the field! 😊

For reference, it took Selzer 12 days after the election last year to produce and release a full post-mortem of her infamous poll. Given their background on methodological transparency, I'd be very surprise if we see anything remotely similar from this so-called "America's most accurate pollster".

Did you say "representative sample"?

A year ago today, AtlasIntel said they would deposit at the Roper Center the raw data from their final polls. It's still not there. Something tells me they will not be depositing that data file, neither the ones from this week's polls...

I was just saying this in the other site: Mamdani's share should probably have a heavier weight in the post-election polling analysis than Cuomo's share, or at least it should be put together with Sliwa's, given the likely late shift. And looks like somebody already did it.

Believe or not, the New Jersey poll wasn't the lowest point in US elections for the (former?) "American most accurate pollster", AtlasIntel Take a look at their final poll in the PNP Primaries in Puerto Rico last year: Poll: Pierluisi +14 Result: González +9 That's a 23-points difference!

l guess NJ is a small demographic subgroup for AtlasIntel standards?

Mood da noite

Also, can we all put to bed now the myth of "the most accurate pollster in the US"? I know I've been saying this since after the 2020 election, but I think we can finally agree on that, right?! RIGHT?!?

Wow, turns out weighting the hell out of your sample recruited through river sampling out of ads in social media doesn't work after all, who knew?! 🤭

I've been calling AtlasIntel's poor (to say the least) methodology out since after the 2020 elections, when it was not a popular thing to say. I'm just glad you all caught up and we are all on the same page now!

ANES Data Release! electionstudies.org/data-center/... The 3-wave ANES panel is now available. It merges data from 3 election studies (2016-2020-2024), the first time the ANES has collected interviews of the same respondents across 3 presidential elections.

They literally just published a NYC poll two days ago, but sure, why not?

LMAO That's not how margins of error work. They reflect random sampling error (variance), not systematic errors, which clearly seems to be the case here, when looking across polls. Also, saying with a straight face that what's shady is if the crosstab looks okay. C'mon, you are not even trying! 😅

🚨It's finally here!🚨 AAPOR's Taskforce on 2024 Pre-Election Polling report is out! Full report: /https://aapor.org/wp-content/uploads/2025/10/AAPOR-Task-Force-on-2024-Pre-Election-Polling_Report.pdf Executive summary: aapor.org/wp-content/u...

I don't know who else still needs to hear this, but raking is not post-stratification. Could you all just stop using "post-stratification" as a synonym for "raking", or more generally, "calibration"?

I'm a sampling statistician and a lot of people ask me about "representative samples". Since there is quite some confusion about it, I'm putting here an entire thread with everything you need to know about "representative samples": 🧵 [1/n]

Happy World Statistics Day! 🎉📊🤓 Statistics plays a fundamental role in helping us describe and better understand our world through high-quality data. #WorldStatisticsDay

new polls website just dropped fiftyplusone.news

Oh, I see some new weighting variables here. Cool! And yes, the methods report is the very first thing I look at a new poll 🤓

Interesting result on the effect of presenting a "No opinion" option on this Washington Post-Ipsos poll www.washingtonpost.com/politics/202...

Dr. Sunghee Lee (@sungheelee.bsky.social‬) featured in this NYT article examining the consequences of deep NIH cuts on health disparities research. Her work on dementia in underrepresented populations was defunded despite its critical value. Read more: myumi.ch/W6dZb #HealthEquity #SocialResearch

I got asked to take uncertainty intervals off a chart because the lower tail showed potentially a 0% lift and like nah dawg I will not be doing that. You can take uncertainty out of the chart but you can't take it out of the system --- if you wanna delude yourself, that's between you and god

🚨 All data nerds 🤓: ANES 2024 full release is now available!

🚨 For all data nerds 🤓 out there: incoming announcement for a new data release!

THREAD: There are a lot of unanswered questions 12 hours after President Trump — who, according to the Constitution, has no final authority over the census — released a social media post calling for unprecedented changes to how the government produces data that forms the foundation of U.S. democracy

Even if you dislike the results, don't kill the polls!

I *really* hope ACS never gets to Trump's attention over the next few years...

AAPOR Statement on the Removal of BLS Commissioner McEntarfer aapor.org/statements/a...

Past vote weight-ers, rejoice!!! 🎉 www.pewresearch.org/decoded/2025...

Who in their right mind would poll Mamdani in a Democratic primary for president?! Oh, right, of course! 🤭

Trump is threatening Brazil with 50% tariffs for no reason whatsoever -- It's Brazil that has a trade deficit with the US...

This has to be a joke, right?

This week's E is for Epi newsletter: what the NYTimes' hit piece on Zohran Mamdani can tell us about the importance of good survey design. Read it here: open.substack.com/pub/epiellie...

Since I reposted this yesterday, I feel obliged to also post this important correction to the charts:

Turns out it's much harder to poll when a large part of the electorate is new and/or young voters

@badsurveyq.bsky.social‬ is in the house! Make sure to tag them for all of your favorite (or least favorite) bad survey questions!

This is brilliant! And for photo/image sharing, bring back the fax machine! Also, the survey research in me thinks: "Yay, make RDD great again!" 🥳

I love how journalists and writers are getting an independent space to freely express themselves, but this Substack model is quickly becoming unsustainable for consumers of this type of media (even more so than the streaming model). I hope we start to see some bundling options at some point...

This interview is a great lesson about liberalism and it brings hope that there is an alternative path in politics through grace and empathy.

Reminder that if these polls are using online opt-in panels, results for young people can be misleading. www.pewresearch.org/short-reads/...

One of my son's toys is the survivorship bias aircraft 😍🤓

Strength In Numbers partnered with Verasight to ask a sample of 1,000 Americans to describe the first few months of the Trump presidency in their own words. Republicans used words like "good" and "great." Democrats wrote "terrible" and "treason." More: www.gelliottmorris.com/p/what-voter...

With that said, here would be my advice (that nobody asked for) to pollsters running pre-election polls in the US: if you are not already, try to invest more in election modelling, especially using the voter file. +

We are excited to team up with @um-psc.bsky.social to create an International Hub for the important survey and popul. research that is done in the social sciences. Bill Axinn, founding Director, brings many years of experience to the position. isr.umich.edu/news-events/... @um-src.bsky.social

The issue of evaluating pollsters based on their, so-called, accuracy in pre-election polls is that this is turning much more in a matter of election modelling than survey methods. If you have very good models, you can get pretty decent estimates even with a crappy methodology. 🧵 +