# Herding in Polling Herding is what happens when pollsters, having produced a result far from the consensus, decline to publish it — or adjust it toward the consensus through defensible-looking methodological choices — because being wrong alone is much more damaging than being wrong with everyone else. ## The statistical fingerprint Herding is detectable without any inside knowledge, because **honest polls should scatter.** If a race is genuinely at 48–46 and pollsters are drawing samples of a thousand or so, sampling error alone guarantees a substantial spread of published results. You should see 51–43 and 45–49 among them. When the published polls in the final weeks of a race cluster far more tightly than sampling theory allows, that clustering is itself the evidence. The distribution is *too narrow*. Analysts test this by comparing the observed variance of late polls against the variance implied by their stated sample sizes; a large shortfall is the herding signature. This is a rare and elegant case where **a methodological failure leaves a specific mathematical trace that cannot be argued away** — the polls are either scattering as much as their own reported margins of error require, or they are not. ## Why the incentive exists A pollster's asset is reputation. Consider the payoff matrix at the end of a close race: - Publish an outlier, and the outlier is right: modest reputational gain. - Publish an outlier, and the outlier is wrong: severe reputational damage, cited for years. - Publish near consensus, and consensus is right: no change. - Publish near consensus, and consensus is wrong: no individual damage — "everyone missed it." The asymmetry is stark and the rational response is to hug the average. Nobody has to be dishonest for this to happen. The adjustments are made through choices that are individually defensible: a different likely-voter screen (see [[Likely-Voter Models]]), a different weighting target, a decision that a particular sample looked unrepresentative. Each choice is arguable. The *distribution* of choices is not. This is a clean instance of Goodhart's Law operating on a measurement system: once pollsters are evaluated on proximity to the final result, and the consensus is the best available guess at that result, proximity to consensus becomes the target — and the poll stops being an independent measurement. ## Why it breaks aggregation [[Poll Aggregation]] assumes polls are independent draws. Herding destroys that assumption in the most damaging possible way, because it does not merely correlate the errors — **it removes the spread that aggregators use to estimate uncertainty.** The result is a compounding failure: the aggregate looks unusually confident precisely when the underlying polls have stopped carrying independent information. Tight clustering gets read as convergent evidence when it is actually evidence that the pollsters are looking at each other. Forecast intervals narrow at exactly the moment they should widen. ## The political consequence Herding tends to bias toward whatever the consensus already is, which means it suppresses evidence of late movement. A race genuinely breaking in one direction in the final ten days is the situation where herding does the most damage, because that is when an outlier poll is most likely to be correct and most likely to be withheld. It also means **published polling understates its own uncertainty at the exact moment the stakes are highest**, feeding overconfident coverage and, downstream, the "the polls were rigged" narrative that follows any large miss — see [[Election Denialism]]. The irony is that the real methodological failure is not the one the fraud claims allege, and it is discoverable in the published data by anyone willing to compute a variance. ## Related [[Poll Aggregation]] · [[Likely-Voter Models]] · [[Reading Probabilistic Forecasts]] · Polling · Goodhart's Law · Margin of Error · [[Mathematics and Power Hub]]