BR

Pollster Performance

NOTE: the numbers below only compare the election result with what each poll saw in its last fieldwork before each election. They are not a measure of poll quality, and they should not be compared from one election to another or from one office to another. To understand the numbers, see the calculation and the full explanation of the project.

Election

Presidente

How the weighted error is calculated: the measurement is based on two factors and two adjustments.

Decisive margins: the difference between poll and result in the margin between 1st and 2nd place, and between 2nd and 3rd when the runoff spot was at stake, adjusted for how contested that spot actually was.

Misplaced electorate: how many voters would have to be moved within the poll for it to match the final result. This measure weighs more when the winner is close to 50% of the valid vote, because that directly affects whether there is a runoff. It is not needed in a runoff, where it is proportional to the error in the decisive margin.

Wrong side: adds a penalty when the poll places the electorate past the decision lines (wrong leader, wrong runoff pair, whether there would be a runoff or not), weighted by how close the result actually was.

For presidential polls, we give a discount to pollsters that published their last numbers four or more days before the election, calculated from other polls with fieldwork in the same period.

In years with two Senate seats, the decisive margins are between the two winners and the third-placed candidate.

We always use the valid votes as published by the pollsters when available; otherwise, we assume that voters who did not state a vote are distributed proportionally.

Precision: the weighted error measures closeness to the result, not the quality of the poll. There is therefore plenty of noise: even a perfectly random sample misses by chance. That is why the numbers appear as whole numbers, each race shows a "chance band" and the polls appear in three levels instead of a ranking.

Effective sample size: each poll's weighting and design reduce its effective sample size. Unfortunately this cannot be measured pollster by pollster with the published data. We use the same effective size for all, 970 interviews, estimated from the fluctuation between consecutive polls of the same pollster in the same race, net of the others' movement.

Levels: each error is compared with that of the final polls for the same office in the other elections, relative to what chance would produce in each race.