Key Takeaways

  • Nate Silver points out that Donald Trump carried an approval rating near 37 percent, matching Harry Truman in 1946 for the worst presidential midterm mark on record.
  • Polling errors do not happen in isolation. If survey models miss high on a candidate in Iowa, demographic overlap means they will likely miss in Kansas and Nebraska too.
  • The belief that surveys always underestimate Republicans is wrong. In off-year elections like the New Jersey and Virginia governor races where Trump was absent from the ballot, pollsters underestimated Democrats.
  • Undercounts of Republican turnout stem specifically from low-propensity voters who only participate when Trump runs. Down-ballot candidates fail to activate this cohort on their own.
  • Scared of public embarrassment, pollsters adjust likely-voter models to match consensus spreads, turning empirical data into reverse-engineered safety blankets.

The Ballot Missing Link

Political operatives treat polling error like a permanent tilt. After misses in 2016 and 2020, conventional wisdom concluded that polls systematically miss Republican voters. Silver rejects that premise. The miss only appears when Donald Trump is literally on the physical ballot.

Trump attracts irregular, low-propensity voters who do not answer phone calls from survey firms. Those voters trust Trump, but they do not care about the Republican Party apparatus. When Trump sits out, they sit out. As Silver observed, “In every election in the Trump era where Trump is not on the ballot, you have not had that error. And, in fact, in like New Jersey and Virginia last year, you had a massive underestimation of Democrats in governor's races.” (play from 8:40)

If you assume an error pattern is structural when it is actually tied to a single charismatic figure, your predictive models will fail in ordinary cycles.

Errors Travel in Packs

Amateur analysts treat 50 state races as 50 independent coin flips. Silver stresses that demographic and economic ties bind state outcomes together. When an analytical model makes an assumption error about suburban white voters or working-class turnout, that miscalculation cascades across state lines simultaneously.

Silver explained the mechanism directly: “So, what we know number one is that polling errors are correlated, that if a poll misses high on the Democrat in Iowa, probably in a similar state like Kansas or Nebraska, all three states are competitive this year, good chance you'll have a correlated error.” (play from 7:46) If a modeler assumes each state is independent, they will assign microscopic odds to a sweep that is actually quite likely.

Herding Destroys Survey Independence

The dirtiest secret in public polling is that firms look at each other before publishing. Running a polling firm is a commercial business built on client perception. If you publish an outlier result showing a Democrat up by eight points and you are wrong, you face reputational ruin. If you publish a tight two-point race and miss alongside everyone else, nobody fires you.

Because of that dynamic, pollsters manipulate their turnout filters until their numbers blend into the pack. Silver notes that pollsters operate under intense anxiety: “Election polls are made in an environment where pollsters are actively fearful of missing, particularly missing high on Democrats.” (play from 22:04) Instead of letting the sample speak, “They decided what result they felt comfortable with and reverse-engineered.” (play from 23:17) When every researcher herds toward safety, the published aggregate stops being an independent measurement and functions as a crowd-sourced prediction market.

What to Do With This

Audit your internal product and sales surveys this week. Look at the variance across your team's customer interviews: if every summary report reaches the exact same conclusion about pricing or feature demand, your team is adjusting their notes to match leadership consensus. Separate raw interview transcripts from customer success summaries to strip out social herding before making your next roadmap decision.