Methodology

Polling averages: method cr-average-1.0

Caldwell averages summarize compatible published surveys. They are not election forecasts, winning probabilities, or substitutes for the underlying findings.

Which polls qualify

A reviewed definition specifies exact question wording, all recorded response choices, question base, survey population, accepted interview modes, geography and, for elections, the race. Only capitalization and whitespace are normalized; unlike wording, populations or response choices are not automatically combined. Only published, reviewed surveys with published source references qualify. Each question must have a known sample of at least 300 respondents. Multiple-response questions are excluded.

Window and weights

For each as-of date, the window includes fieldwork end dates from that date back through the preceding 29 days. A survey must also have been released by the as-of date. Only the latest survey from each pollster contributes: ties use the later release date, then the higher record ID. At least three different pollsters must qualify, with at least one survey ending within the preceding 14 days. Otherwise the estimate is withheld as insufficient or stale.

Each included poll receives an unnormalized weight of 2 raised to the power of −age/14, where age is the number of days since fieldwork ended. Divide each weight by their total, multiply each poll value by its normalized weight, and add the results. A poll 14 days old has half the raw weight of a poll ending on the as-of date. Sample sizes above the minimum do not change the weight. Percentages are rounded to one decimal only for display; exports retain the calculation precision.

Shares and election margins

A share totals the response labels named in the definition. An election margin subtracts the Republican candidate response total from the Democratic candidate response total within each poll, then averages those differences. Blue indicates a positive Democratic margin; red a negative Republican margin; purple is within two points. Color intensity changes at 2, 5 and 10 points. Neutral gray means no eligible estimate or multiple races needing selection; it does not mean a tied race.

Dates, comparisons and boundaries

Observation year and election cycle are separate. Year overlays align calendar month and day for the same opinion definition. An unmatched February 29 remains a gap. Missing and stale observations are never converted to zero or connected with interpolated lines. Race definitions are tied to an election cycle; comparisons across different candidates or race definitions require separate editorial review.

State shapes are a geographic overview. District and local shapes require a reviewed boundary with source, vintage, GEOID and matching election cycle. A statewide value is never substituted for a district estimate.

Reproduction and limits

Each dated calculation exposes its included record IDs and versions, source links, fieldwork dates, release dates, values and weights. Saved snapshots retain the definition and inputs. Historical charts are recomputed from current verified record versions, restricted to original release dates; they are not an archive of what Caldwell knew or published at that time. Source corrections can therefore change a historical series. Superseded snapshots are withheld from the public snapshot endpoint.

This initial method does not estimate house effects, pollster quality, design effects, correlated error, or an uncertainty interval. The minimum counts and weighting rule are publication standards, not a claim of validated predictive accuracy. Averages can still share systematic bias. Coverage gaps and exclusions remain visible in the evidence table.