This is a strong piece, and the core thesis is clearly true: the Michigan survey has become a partisan thermometer, and the post-2024 readings can't be taken at face value. I argued something similar about the survey's inflation-expectations component in a Bloomberg column, so I'm sympathetic. But I have three objections.
First, the adjustments explain the wrong thing. The mode shift and the partisan reweighting both land after mid-2024. The vibecession is older than that. Kyla Scanlon coined the term in 2022, under Biden, when — by your own partisanship chart — the sample wasn't yet Democrat-skewed and the survey was still on the phone. The puzzle that launched four years of commentary was sentiment running below what 3.5% unemployment "should" produce. The corrections pull the recent readings up toward 2013 levels; they don't touch the 2022 divergence. They explain the depth, not the phenomenon.
Second, subtracting nine points to match the old phone series assumes phone was unbiased and web is contaminated. But self-administered modes are well documented to reduce social-desirability bias — people are franker to a screen than to a live interviewer. If web respondents are simply more honest, the level shift is signal, and it was the phone series that ran high. As the piece argues, a 0.97 correlation between modes can't adjudicate this. The same goes for anchoring to a fixed NPORS partisan split: it partly defines away the partisan-driven sentiment you're trying to measure.
Third the piece addresses Hsu's claim that the national number tracks independents. But her main argument, in the May 22 note, is that partisan gaps are "economically important and meaningful," showing up in actual spending, entrepreneurship, and portfolio choice. If partisanship moves behavior, then partisan-skewed sentiment isn't expressive noise to weight away — it's a forecast of consumption. You can't treat the partisan signal as meaningless for the index and meaningful for the economy at once.
Which gets at what the vibecession actually is. These instruments measure central tendency, not perceived variance. "Continuous good times, or periods of widespread unemployment, over the next five years" asks about the mean, not the felt second moment. Someone can rationally say things are fine now but the ground feels less stable than before 2020 — pandemic, Silicon Valley Bank, Iran War, an oil shock, tariffs, AI displacement, DSA making a bid to take over the Democratic party while MAGA is showing signs of life after Trump — and that depresses sentiment without being partisan or irrational.
The vibecession reads more like a fragility premium than a happiness deficit. The clearest evidence is the present-versus-expectations gap: okay now, worried ahead. The Conference Board's split between its Present Situation and Expectations indexes captures that directly.
None of this rescues the headline. 49.5 is not a true Great Recession reading. But "partly an artifact of bad data" is too clean — part of it is a real signal the instruments aren't built to measure.
1) Scanlon's claim about 2022 was flawed and simply didn't build in enough lag which is what Nate wrote in his Times' piece. The more recent data was harder to explain; 2) the point is that you can't make historical comparisons when you're switching modes, the level shift isn't a signal in that context; 3) nobody is saying partisanship doesn't matter for sentiment and important and meaningful, but that doesn't justify simply overestimating the number of Democrats.
Three good points, but I defend 2/3s my comment anyway.
1. I credited Kyla Scanlon with the term Vibecession, but I didn't mean to limit it to her analysis. Lots of people have been saying similar things about how CoVID and the collapse of trust in experts and the unedifying spectacle in Washington since 2016 have left many ordinary people feeling that things are pretty good at the moment, but fragile.
2. You are agnostic, claiming the level shift isn't a signal of anything. I say it's meaningful but ambiguous--it could mean the phone data was reliable, or the Internet data, or neither. It only rules out the "both" option.
3. You're right here, I conflated two issues. Overweighting Democrats is a straightforward error, and I ignored it. But it's also true that Hsu presented evidence that this is not partisan griping to a survey-taker, it's real behavior; and the original post ignored that important point.
It's so disappointing that an academic survey like this can do such shoddy work. I work in healthcare. Imagine that somebody creates a new cancer test. They test it and find that a) it's 100% correlated with the gold standard test and b) it catches 10% fewer cancers than the gold standard. Nobody would accept that and it's disappointing that U of M didn't hit pause or figure out how to adjust.
People would accept your example if the gold standard test was prohibitively expensive, and the secondary test was cheap enough to be mass administered as a "good enough" first pass.
That's the problem polling has. Accurate polling is expensive, and at some point there is a tradeoff where you run out of money and have to decide on what's good enough. In addition to that, the only people with an incentive to fund good poling are PACs and campaigns who will directly benefit from the data. Those groups aren't going to be transparent about sharing their data. The general public might like to see the data, but they aren't going to want to pay for it.
Even if maintaining telephone polling is no longer feasible, the U of M survey has the obligation to make the change transparent and help decision-makers adapt to the change. It isn’t just some poll, it’s the most influential source of consumer sentiment that has been relied on by multiple professions by decades.
The inconsistency of date ranges on the charts makes the analysis raise my eyebrows a bit. Looking at the charts, they go back to:
* 2006
* 2016
* "G.W. Bush", so 2001-2008?
* N/A - not your chart
* 2022
* N/A - point in time
* N/A - not your chart
* 2017
* 2007 - not your chart, but the data is publicly available to make it shorter/longer
* 2009
Some of these may be benign in intent (e.g. you want to highlight a more recent trend, or not all data series are available for the full time window) but shifting the window of analysis on every single chart smells a lot like cherry picking to make the data fit your conclusion.
Take the chart "Consumer sentiment looks less negative after adjusting for mode and partisanship". It goes back to 2017, but the leading text references earlier dates in Obama's term ("depths-of-the-Great-Recession", "when he was re-elected"), so reader's can't even see that in the data.
This reads like… a Silver Bulletin article! Introducing a new author into a small team is a punt I’m sure. You’ve nailed it. Punchy geeks of the world unite!
This is strong work. I'd like to know if UMich. responds. I think another feature of the disconnect is that the people who write about these subjects are disproportionately on the coasts where life is very expensive and the industries they're employed in are not thriving.
Combine the NPR Nazi 24/7 simp-missives and the Michigan survey and you have a narrative that has consistently kicked Democrats square in the balls.
I was a Democrat for 33 straight years. I voted for Clinton, Gore, Kerry, Obama, Biden and Harris. No more. Now, I'm a proud Independent. Why? A tranny, AJ Hernandez, won the Califoria State HS triple jump/long jump/high jump; Florence Joyner, Marion Jones, Tommie Smith, Lee Evans, Rafer Johnson, Bob Mathias, all CIF champs/Olympic Gold take a back seat to a nutless/dickless dude who now has plastic vertical smile. Pathetic.
One thing I worry about, and maybe Silver Bulletin has already explained this, but how do online polls filter for bots? Like, how do pollsters & researchers know that these are human responses online when agentic AI can now act semi-autonomously? If they can use CCs to make purchases, or work social accounts to shitpost & manipulate, then why couldn't they also be instructed to answer online polls to construct a specific narrative?
That doesn't seem beyond the realm of possibility for a foreign government, lobbying organization, or a megacorp to implement.
Really enjoyed this piece. I spent some time earlier this year writing about the consumer sentiment gap with a particular focus on the UMich survey, and this post adds some much-needed clarity to the data.
One thing that stuck with me is where the corrected series lands. Even after both adjustments, consumer sentiment sits at roughly 2013 levels, which was a time when unemployment was around 7.5% and households were still recovering from the GFC. The adjusted sentiment figure seems much easier to believe than the headline data plunging to 2008 lows, but it leaves an open question since today's mix of higher inflation and lower unemployment doesn't neatly map to 2013. When I modeled sentiment against standard macroeconomic predictors, the post-2020 gap came out roughly five times larger than the historical norm. The mode shift clearly accounts for a meaningful piece of that, but I don't believe it accounts for all of this gap.
While it makes sense that the corrected Michigan series converges with trackers like Gallup and YouGov (which all still reflect subdued consumer sentiment themselves), the main reason I suspect a genuine shift in the public mood comes down to the broader happiness data. Sam Peltzman's analysis of the General Social Survey happiness data found that there was a happiness decline in the aftermath of the pandemic and that this decline has spared no party. In fact, happiness for Republicans (who historically have responded as being happier than Democrats) fell further than for Independents and Democrats, which is the opposite of what a Democrat-heavy sample would produce. And the most puzzling pattern of the broader happiness decline is that the people doing the best economically were hit the hardest.
So --- when the reading doesn't look right, the thing to do is just to "adjust" the data? Yeah --- sounds like most polling that goes on in this country. Oh, the poll says the wrong person is likely to win? We can fix that-----
As for the gross undercounting of Republicans, I've known that for ten years, but it just gets worse and worse, looks like. Hey, maybe if you Democrats undercount us enough, we'll all just disappear! Good luck with that.
If one reads what Nate actually has written rather than engage in weird innumerate partisan knee-jerking, the analysis and adjustments on data (and cross benchmarking against other indicators) are data analytical.
Of course frothing political partisans on both edges of the spectrum will search for conspiratorial just-so stories and JAQing.
99.99% sure the author / Nate was not confident enough to believe in or draft their own title. Of all the places to use gpt or a misplaced m-dash titles are the worst ...
I’m at least 20% sure that the team used a LLM to suggest some titles for this piece that are consistent with the Silver Bulletin style that will maximise attention, and then chose the one they liked.
And I’m 90% sure this is a perfectly fine use of AI. Particularly given how overt Nate is about his use of these tools.
This is a strong piece, and the core thesis is clearly true: the Michigan survey has become a partisan thermometer, and the post-2024 readings can't be taken at face value. I argued something similar about the survey's inflation-expectations component in a Bloomberg column, so I'm sympathetic. But I have three objections.
First, the adjustments explain the wrong thing. The mode shift and the partisan reweighting both land after mid-2024. The vibecession is older than that. Kyla Scanlon coined the term in 2022, under Biden, when — by your own partisanship chart — the sample wasn't yet Democrat-skewed and the survey was still on the phone. The puzzle that launched four years of commentary was sentiment running below what 3.5% unemployment "should" produce. The corrections pull the recent readings up toward 2013 levels; they don't touch the 2022 divergence. They explain the depth, not the phenomenon.
Second, subtracting nine points to match the old phone series assumes phone was unbiased and web is contaminated. But self-administered modes are well documented to reduce social-desirability bias — people are franker to a screen than to a live interviewer. If web respondents are simply more honest, the level shift is signal, and it was the phone series that ran high. As the piece argues, a 0.97 correlation between modes can't adjudicate this. The same goes for anchoring to a fixed NPORS partisan split: it partly defines away the partisan-driven sentiment you're trying to measure.
Third the piece addresses Hsu's claim that the national number tracks independents. But her main argument, in the May 22 note, is that partisan gaps are "economically important and meaningful," showing up in actual spending, entrepreneurship, and portfolio choice. If partisanship moves behavior, then partisan-skewed sentiment isn't expressive noise to weight away — it's a forecast of consumption. You can't treat the partisan signal as meaningless for the index and meaningful for the economy at once.
Which gets at what the vibecession actually is. These instruments measure central tendency, not perceived variance. "Continuous good times, or periods of widespread unemployment, over the next five years" asks about the mean, not the felt second moment. Someone can rationally say things are fine now but the ground feels less stable than before 2020 — pandemic, Silicon Valley Bank, Iran War, an oil shock, tariffs, AI displacement, DSA making a bid to take over the Democratic party while MAGA is showing signs of life after Trump — and that depresses sentiment without being partisan or irrational.
The vibecession reads more like a fragility premium than a happiness deficit. The clearest evidence is the present-versus-expectations gap: okay now, worried ahead. The Conference Board's split between its Present Situation and Expectations indexes captures that directly.
None of this rescues the headline. 49.5 is not a true Great Recession reading. But "partly an artifact of bad data" is too clean — part of it is a real signal the instruments aren't built to measure.
1) Scanlon's claim about 2022 was flawed and simply didn't build in enough lag which is what Nate wrote in his Times' piece. The more recent data was harder to explain; 2) the point is that you can't make historical comparisons when you're switching modes, the level shift isn't a signal in that context; 3) nobody is saying partisanship doesn't matter for sentiment and important and meaningful, but that doesn't justify simply overestimating the number of Democrats.
Three good points, but I defend 2/3s my comment anyway.
1. I credited Kyla Scanlon with the term Vibecession, but I didn't mean to limit it to her analysis. Lots of people have been saying similar things about how CoVID and the collapse of trust in experts and the unedifying spectacle in Washington since 2016 have left many ordinary people feeling that things are pretty good at the moment, but fragile.
2. You are agnostic, claiming the level shift isn't a signal of anything. I say it's meaningful but ambiguous--it could mean the phone data was reliable, or the Internet data, or neither. It only rules out the "both" option.
3. You're right here, I conflated two issues. Overweighting Democrats is a straightforward error, and I ignored it. But it's also true that Hsu presented evidence that this is not partisan griping to a survey-taker, it's real behavior; and the original post ignored that important point.
It's so disappointing that an academic survey like this can do such shoddy work. I work in healthcare. Imagine that somebody creates a new cancer test. They test it and find that a) it's 100% correlated with the gold standard test and b) it catches 10% fewer cancers than the gold standard. Nobody would accept that and it's disappointing that U of M didn't hit pause or figure out how to adjust.
People would accept your example if the gold standard test was prohibitively expensive, and the secondary test was cheap enough to be mass administered as a "good enough" first pass.
That's the problem polling has. Accurate polling is expensive, and at some point there is a tradeoff where you run out of money and have to decide on what's good enough. In addition to that, the only people with an incentive to fund good poling are PACs and campaigns who will directly benefit from the data. Those groups aren't going to be transparent about sharing their data. The general public might like to see the data, but they aren't going to want to pay for it.
Even if maintaining telephone polling is no longer feasible, the U of M survey has the obligation to make the change transparent and help decision-makers adapt to the change. It isn’t just some poll, it’s the most influential source of consumer sentiment that has been relied on by multiple professions by decades.
The inconsistency of date ranges on the charts makes the analysis raise my eyebrows a bit. Looking at the charts, they go back to:
* 2006
* 2016
* "G.W. Bush", so 2001-2008?
* N/A - not your chart
* 2022
* N/A - point in time
* N/A - not your chart
* 2017
* 2007 - not your chart, but the data is publicly available to make it shorter/longer
* 2009
Some of these may be benign in intent (e.g. you want to highlight a more recent trend, or not all data series are available for the full time window) but shifting the window of analysis on every single chart smells a lot like cherry picking to make the data fit your conclusion.
Take the chart "Consumer sentiment looks less negative after adjusting for mode and partisanship". It goes back to 2017, but the leading text references earlier dates in Obama's term ("depths-of-the-Great-Recession", "when he was re-elected"), so reader's can't even see that in the data.
The partisan data from UMich was not gathered consistently until recently https://data.sca.isr.umich.edu/fetchdoc.php?docid=81303
This reads like… a Silver Bulletin article! Introducing a new author into a small team is a punt I’m sure. You’ve nailed it. Punchy geeks of the world unite!
(And still waiting for Claire Malone…)
This is strong work. I'd like to know if UMich. responds. I think another feature of the disconnect is that the people who write about these subjects are disproportionately on the coasts where life is very expensive and the industries they're employed in are not thriving.
Combine the NPR Nazi 24/7 simp-missives and the Michigan survey and you have a narrative that has consistently kicked Democrats square in the balls.
I was a Democrat for 33 straight years. I voted for Clinton, Gore, Kerry, Obama, Biden and Harris. No more. Now, I'm a proud Independent. Why? A tranny, AJ Hernandez, won the Califoria State HS triple jump/long jump/high jump; Florence Joyner, Marion Jones, Tommie Smith, Lee Evans, Rafer Johnson, Bob Mathias, all CIF champs/Olympic Gold take a back seat to a nutless/dickless dude who now has plastic vertical smile. Pathetic.
One thing I worry about, and maybe Silver Bulletin has already explained this, but how do online polls filter for bots? Like, how do pollsters & researchers know that these are human responses online when agentic AI can now act semi-autonomously? If they can use CCs to make purchases, or work social accounts to shitpost & manipulate, then why couldn't they also be instructed to answer online polls to construct a specific narrative?
That doesn't seem beyond the realm of possibility for a foreign government, lobbying organization, or a megacorp to implement.
Really enjoyed this piece. I spent some time earlier this year writing about the consumer sentiment gap with a particular focus on the UMich survey, and this post adds some much-needed clarity to the data.
One thing that stuck with me is where the corrected series lands. Even after both adjustments, consumer sentiment sits at roughly 2013 levels, which was a time when unemployment was around 7.5% and households were still recovering from the GFC. The adjusted sentiment figure seems much easier to believe than the headline data plunging to 2008 lows, but it leaves an open question since today's mix of higher inflation and lower unemployment doesn't neatly map to 2013. When I modeled sentiment against standard macroeconomic predictors, the post-2020 gap came out roughly five times larger than the historical norm. The mode shift clearly accounts for a meaningful piece of that, but I don't believe it accounts for all of this gap.
While it makes sense that the corrected Michigan series converges with trackers like Gallup and YouGov (which all still reflect subdued consumer sentiment themselves), the main reason I suspect a genuine shift in the public mood comes down to the broader happiness data. Sam Peltzman's analysis of the General Social Survey happiness data found that there was a happiness decline in the aftermath of the pandemic and that this decline has spared no party. In fact, happiness for Republicans (who historically have responded as being happier than Democrats) fell further than for Independents and Democrats, which is the opposite of what a Democrat-heavy sample would produce. And the most puzzling pattern of the broader happiness decline is that the people doing the best economically were hit the hardest.
This raises the question of whether a deeper trend might help explain both the remaining sentiment gap and the simultaneous drop in happiness post-COVID. (Also if anyone's curious, I wrote up the residual analysis here: https://godisaswingvoter.substack.com/p/the-atomization-will-continue-until)
So --- when the reading doesn't look right, the thing to do is just to "adjust" the data? Yeah --- sounds like most polling that goes on in this country. Oh, the poll says the wrong person is likely to win? We can fix that-----
As for the gross undercounting of Republicans, I've known that for ten years, but it just gets worse and worse, looks like. Hey, maybe if you Democrats undercount us enough, we'll all just disappear! Good luck with that.
If one reads what Nate actually has written rather than engage in weird innumerate partisan knee-jerking, the analysis and adjustments on data (and cross benchmarking against other indicators) are data analytical.
Of course frothing political partisans on both edges of the spectrum will search for conspiratorial just-so stories and JAQing.
Nate didn't write it, so evidently it's YOU should read the article before your knee-jerk, or maybe just jerk, partisan nastiness.
Yes you are right, I only read the text closely and not the byline. My bad indeed. the rest stands.
99.99% sure the author / Nate was not confident enough to believe in or draft their own title. Of all the places to use gpt or a misplaced m-dash titles are the worst ...
I’m at least 20% sure that the team used a LLM to suggest some titles for this piece that are consistent with the Silver Bulletin style that will maximise attention, and then chose the one they liked.
And I’m 90% sure this is a perfectly fine use of AI. Particularly given how overt Nate is about his use of these tools.