concept

Marvin Barth's term for polling errors that run consistently in one direction, which he says overstated Democrats in 2014, 2016, 2020 and 2024. In episode 01 it is the quantity his forecast tries to estimate. In episode 02 he reports that the 2026 errors were as large as he expected, bigger than in 2016 or 2024, but overstated the Republicans by four to five points. Start with episode 01, at about 12 minutes, then episode 02, at about 6 minutes.

Across the episodes

01

Enters as the quantity the forecast tries to size, a systematic miss Barth traces back before Trump

Barth's distinction: random error averages out, but a miss that always points the same way is systematic.

"But the places where people really started to notice these poll errors, and they're not poll errors, they're poll bias. They were in one direction."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 1 (2 Nov 2026), 12:10

By pushing the pattern back to 2014, Barth detaches it from Trump personally.

"2014 actually had even bigger polling errors, polling bias in favor of Democrats than 2016 did."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 1 (2 Nov 2026), 12:34

Why the bias resists correction: it is made of absent respondents.

"Because what you're effectively looking for is people who don't show up in polls."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 1 (2 Nov 2026), 13:42

The bias can be measured only after the votes are counted, which is why Barth's forecast has to be a judgment.

"And so the only way you ever see this is ex post."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 1 (2 Nov 2026), 28:17

Barth's cross-state picture of the bias, the graph he expects to tilt the other way if the left hides its vote.

"Polling errors go up with the share of the Republican vote. That's that right wing populism."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 1 (2 Nov 2026), 1:07:52
02

Flips sign: the bias Barth set out to size arrives at the size he expected but in the Democrats' favour, and he tests it with the same error-versus-bias distinction he drew before the vote

The size of the miss bears out the episode-01 premise that the polls were badly off, even as the call failed.

"This is, I think, the biggest takeaway here is that the polling errors were bigger this year than they were in 2024 or 2016."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 2 (7 Nov 2026), 5:57

The sign reverses: a bias Barth had measured as running for the Democrats ran against them this time.

"And the difference was, is that rather than erring towards overstating Democrats, it overstated Republicans."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 2 (7 Nov 2026), 6:14

Barth puts his own distinction from episode 01, random error against one-directional bias, to the new miss, and answers it with the demographic data.

"And you say, okay, well, maybe that's just a polling error, not bias in the polls."
Marvinasserted#
Marvin, The Grant Williams Podcast: Forecasts and Fortunes, Episode 2 (7 Nov 2026), 6:40

Connections

Barth's explanation of the bias.

Barth's earliest and largest example.

The 2026 errors overstated the Republicans.