I was really lookng forward to this, but the "other" category misses a lot. If you look at Arizona, the maps combines Navajo County and Apache County. The northern parts of that county are Navajo lands, the southern rural conservative communities. It would be useful to see a breakdown. Same story in Montana. The grouping of regions misses Native American voters, particularly around the Highline (U.S. 2) but across the state.
Ron Brownstein says it's going to be very hard for Republicans to win, in the areas where Trump's approval is under 50%. If next November works out that way, sure looks like a blue tsunami.
I am sure shifting to a focus on individual states is a good idea, but doesn't that level of detail require more people to be polled to ensure that enough people are polled in each of the state levels? I guess my question is how do you determine your polls have polled enough people to validate breaking that data into 50 state pieces or even 535 congressional districts.
It seems to me that the most relevant speculative factor at this point in time is the expectation of turnout by party both in the reelection of Dems for the Senate and the 5 or 6 key potential upsets in NC, OH, ME, NE, IA, Alaska and TX? Do you agree and what are your numbers for each of these states?
I'm a subscriber and I'm trying to figure out who's your intended audience, and under the suspicion that I'm not in it).
At the top of today's 'stack, you use the term MRP without explanation (it's okay, I looked it up). At the top of the nerd section, you use the term again, and at least spell out the acronym, but you never explain what it means.
On your charts, sometimes the units aren't defined (like the numbers printed on the states on the Mondale outcome map -- took me a minute to surmise, correctly on not, that those respresent electors), and sometimes it would be really helpful to have other clarifications along the lines of, e.g., "The gray line shows x," or point out anomalies (still trying to figure out what that "NM" line is on the MRP estimates one!).
I realize it would be a lot of work to make an English majors version, but if you had a non-statistician friend or family member read your non-nerd posts and point out simple additions/fixes, people like me would get a lot more out of it (and be more likely both to subscribe and re-up -- and there are lots more people like me than there are economists/statisticians!). Krugman does an excellent job with super-simple explanations of terms. Regression means different things to different people!~
Excellent notes -- But watch Indiana... A +20 points state for Trump in 2024, but notoriously ready to turn purple. Trump has screwed the soybean farmers, auto industry, and has really offended plain Hoosier folk with his self-love. In particular -- watch Governor Mike Braun -- a STAUNCH Trump supporter as senator and early-term governor -- but smart enough to know wind direction -- especially after Republicans in the state legislature overwhelming stiff-armed Trump on gerrymandering. Also note former Indiana governor Mitch Daniels, is finally, carefully, going public with is negative views of Trump. (He was privately negative on Trump in 2024.) Meanwhile, other previous GOP governor Mike Pence is actively stirring the negative-Trump pot. This all will make for an interesting stew by summer of 2026.
As an Indiana Democrat, I agree with you that the state is trending toward purple, based on about 40% of those casting ballots in the governor's race in 2024 voted for the Democrat. The Indiana Democratic party is its own worst enemy in that it is slow to identify promising potential candidates and to support them. There is also a failure to increase the voter base by reaching out to those who for whatever reason do not vote.
State by state is super helpful l, and I hope all of the campaign data nerds that need to see this are seeing it! Your work is so important. Thank you!!
The shift to focus on individual states rather than national polling is very helpful for working out election strategy at both the state and national level.
Your excellent explanation at the end of your modeling method is way beyond me, but I have one question: if you weight strata by their voting characteristics in the 2024 Presidential election, aren't you building in a pro-GOP (pro Trump actually) bias that will not be present in 2026 when he is not personally on the ballot (although arguably the mid-terms are heavily influenced by attitudes toward Trump himself)? Stated differently, if many GOP voters voted in 2024 because they wanted to support Trump, and simply clicked on GOP candidates on the ballot, won't there be a drop off in 2026 of such voters because their man is not on the ballot?
This is a great question. The simple answer is no, adjusting by past vote does not bias the model toward Republicans. This is because we are only holding their recollection of their past behavior constant with real-world results, and still allowing change in other variables. So if the people who said they voted for Trump in 2024 show higher disaffection rates in 2026/2028, that gets picked up by the model.
Put another way, imagine we have 4 groups of adults: (1) Adults who voted for Trump, (2) Adults who voted for Harris, (3) Adults who voted for someone else, and (4) Adults who did not vote in 2024, either because they were too young or they chose not to. These groups should be roughly stable over time, since turnout and Trump's vote share in 2024 do not change. But forcing the past vote buckets to match proportions in real results does not bias _current_ estimates, because those are allowed to change relative to past targets.
In other words, we aren't forcing the 2026 vote intention for every respondent to match their 2024 vote intention.
*I say "roughly" because actually the non-voter pool should increase slightly over time as 2024 voters age out of the electorate and younger people age in, increasing the non-voter %, but we can take care of that too.
We look at adjustment for (mismeasured) past vote X and ask your question: when does this adjustment pull the results towards that past election ? It depends on the example, e.g. in example 3 not adjusting is actually closer to the past election.
I was really lookng forward to this, but the "other" category misses a lot. If you look at Arizona, the maps combines Navajo County and Apache County. The northern parts of that county are Navajo lands, the southern rural conservative communities. It would be useful to see a breakdown. Same story in Montana. The grouping of regions misses Native American voters, particularly around the Highline (U.S. 2) but across the state.
Ron Brownstein says it's going to be very hard for Republicans to win, in the areas where Trump's approval is under 50%. If next November works out that way, sure looks like a blue tsunami.
This is really cool work, thanks for sharing
I am sure shifting to a focus on individual states is a good idea, but doesn't that level of detail require more people to be polled to ensure that enough people are polled in each of the state levels? I guess my question is how do you determine your polls have polled enough people to validate breaking that data into 50 state pieces or even 535 congressional districts.
It seems to me that the most relevant speculative factor at this point in time is the expectation of turnout by party both in the reelection of Dems for the Senate and the 5 or 6 key potential upsets in NC, OH, ME, NE, IA, Alaska and TX? Do you agree and what are your numbers for each of these states?
English major/former editor alert:
I'm a subscriber and I'm trying to figure out who's your intended audience, and under the suspicion that I'm not in it).
At the top of today's 'stack, you use the term MRP without explanation (it's okay, I looked it up). At the top of the nerd section, you use the term again, and at least spell out the acronym, but you never explain what it means.
On your charts, sometimes the units aren't defined (like the numbers printed on the states on the Mondale outcome map -- took me a minute to surmise, correctly on not, that those respresent electors), and sometimes it would be really helpful to have other clarifications along the lines of, e.g., "The gray line shows x," or point out anomalies (still trying to figure out what that "NM" line is on the MRP estimates one!).
I realize it would be a lot of work to make an English majors version, but if you had a non-statistician friend or family member read your non-nerd posts and point out simple additions/fixes, people like me would get a lot more out of it (and be more likely both to subscribe and re-up -- and there are lots more people like me than there are economists/statisticians!). Krugman does an excellent job with super-simple explanations of terms. Regression means different things to different people!~
I don't like feeling dumb.
Excellent notes -- But watch Indiana... A +20 points state for Trump in 2024, but notoriously ready to turn purple. Trump has screwed the soybean farmers, auto industry, and has really offended plain Hoosier folk with his self-love. In particular -- watch Governor Mike Braun -- a STAUNCH Trump supporter as senator and early-term governor -- but smart enough to know wind direction -- especially after Republicans in the state legislature overwhelming stiff-armed Trump on gerrymandering. Also note former Indiana governor Mitch Daniels, is finally, carefully, going public with is negative views of Trump. (He was privately negative on Trump in 2024.) Meanwhile, other previous GOP governor Mike Pence is actively stirring the negative-Trump pot. This all will make for an interesting stew by summer of 2026.
As an Indiana Democrat, I agree with you that the state is trending toward purple, based on about 40% of those casting ballots in the governor's race in 2024 voted for the Democrat. The Indiana Democratic party is its own worst enemy in that it is slow to identify promising potential candidates and to support them. There is also a failure to increase the voter base by reaching out to those who for whatever reason do not vote.
It's fascinating to see the difference between western rural North Carolina and eastern rural Tennessee
State by state is super helpful l, and I hope all of the campaign data nerds that need to see this are seeing it! Your work is so important. Thank you!!
The shift to focus on individual states rather than national polling is very helpful for working out election strategy at both the state and national level.
Your excellent explanation at the end of your modeling method is way beyond me, but I have one question: if you weight strata by their voting characteristics in the 2024 Presidential election, aren't you building in a pro-GOP (pro Trump actually) bias that will not be present in 2026 when he is not personally on the ballot (although arguably the mid-terms are heavily influenced by attitudes toward Trump himself)? Stated differently, if many GOP voters voted in 2024 because they wanted to support Trump, and simply clicked on GOP candidates on the ballot, won't there be a drop off in 2026 of such voters because their man is not on the ballot?
Hi there Rees,
This is a great question. The simple answer is no, adjusting by past vote does not bias the model toward Republicans. This is because we are only holding their recollection of their past behavior constant with real-world results, and still allowing change in other variables. So if the people who said they voted for Trump in 2024 show higher disaffection rates in 2026/2028, that gets picked up by the model.
Put another way, imagine we have 4 groups of adults: (1) Adults who voted for Trump, (2) Adults who voted for Harris, (3) Adults who voted for someone else, and (4) Adults who did not vote in 2024, either because they were too young or they chose not to. These groups should be roughly stable over time, since turnout and Trump's vote share in 2024 do not change. But forcing the past vote buckets to match proportions in real results does not bias _current_ estimates, because those are allowed to change relative to past targets.
In other words, we aren't forcing the 2026 vote intention for every respondent to match their 2024 vote intention.
*I say "roughly" because actually the non-voter pool should increase slightly over time as 2024 voters age out of the electorate and younger people age in, increasing the non-voter %, but we can take care of that too.
Great question and thoughtful answer !
We explored this a bit here: https://statmodeling.stat.columbia.edu/2025/12/30/survey-statistics-more-adventures-in-mismeasured-x/
We look at adjustment for (mismeasured) past vote X and ask your question: when does this adjustment pull the results towards that past election ? It depends on the example, e.g. in example 3 not adjusting is actually closer to the past election.
Thanks Shira and cool to see you here. I read everything you write.