# Poll Aggregation Averaging polls reduces random sampling noise. If each poll is an independent noisy estimate of the same quantity, combining *n* of them shrinks the standard error by a factor of roughly √n — the Signal vs Noise logic that makes aggregation obviously worth doing. The word carrying all the weight in that sentence is **independent**. ## What aggregators actually do Serious aggregation is more than a mean: - **Recency weighting.** Older polls get discounted, because opinion moves. - **Quality weighting.** Pollsters are scored on historical accuracy and methodological transparency. - **House-effect adjustment.** Pollsters with a consistent lean relative to the average get corrected toward it. - **Trend-line estimation.** Fitting a curve through the polls rather than averaging a window, so the estimate tracks movement rather than lagging it. Each of these is a modeling decision, and they interact. Two aggregators using the same polls produce different averages. ## The correlated-error ceiling Here is the limitation that matters, and it is not a technicality. **Polling error is not primarily sampling error.** Sampling error is the part the ±3 describes, and it is the part aggregation kills. The larger part is *systematic* — nonresponse patterns, likely-voter modeling assumptions (see [[Likely-Voter Models]]), weighting targets, mode effects. These errors are **shared across pollsters**, because pollsters use similar methods, similar sampling frames, and similar assumptions about who votes. When errors are correlated, averaging does not help. If every pollster is understating support for a candidate by three points because the same demographic is systematically declining to answer surveys, the average of fifty such polls is also off by three points — and it now carries the false authority of a large effective sample. **This is the single most important thing to understand about poll aggregation: it converts many noisy estimates into one precise estimate of the wrong number, whenever the bias is common.** The precision is real. The accuracy is not improved at all. Recent American cycles have shown exactly this signature: aggregate errors substantially larger than the aggregate's stated uncertainty, in a consistent direction, across essentially all public pollsters. ## Why the good models handle it and the naive ones don't Well-built forecasting models account for correlated error explicitly — by inflating uncertainty well beyond what independence would imply, and by correlating errors across states so that a miss in Pennsylvania implies a likely miss in Michigan. This is why a responsible forecast shows a much wider distribution than the polls alone would suggest, and why its state outcomes move together. A naive average shows a tight interval and treats states as independent. It will look more precise and be more wrong, and in a close election the difference between those two treatments is the difference between "toss-up" and "safe." ## The political consequence Aggregation produces a **single authoritative number** where before there was visible disagreement. That is genuinely useful and also genuinely distorting: the visible spread across polls was information about uncertainty, and collapsing it into an average discards that signal in favor of a figure that gets reported as fact. The 2016 cycle is the canonical case. Aggregates showed a modest, stable lead. The polls were not catastrophically wrong by historical standards; the *interpretation* of aggregate precision was. Coverage, resource allocation, and public expectation all ran on a number whose uncertainty had been understated by the aggregation process itself. ## Related Polling · [[Likely-Voter Models]] · [[Herding in Polling]] · [[Reading Probabilistic Forecasts]] · Signal vs Noise · Margin of Error · [[Mathematics and Power Hub]]