Predicting is difficult, but verifying the prediction afterwards is easy. Weather models are a higher science, but after a day it is possible to verify whether the expected midday temperature is correct or not by taking a look at a thermometer. By keeping a daily log of how wrong the weather forecast for a while, you get an indication of the margin of error in the model.
This is more difficult with polls. You cannot wait for dozens of elections to estimate the margin of error. If we look at the evidence for the March 16 referendum, we see that the result is within the margin of error, ten parties and seven outside it. With polling evidence, there is a 95 percent chance that the true result is within the margin of error. A score of 10 out of 17 times within the margin of error is not good.
Did the polls perform poorly? not nessacary. The D66 has been on the rise last week, so it’s not surprising that the game from day sixteenth to day seventeenth is going up a bit. This In 1Editors gave Kaag an additional program the day before the election, only due to opinion polls, giving an enhanced effect.
Another reason it’s hard to say that polls were wrong is that polls pose a hypothetical question. (Not usually hypothetical like the poll that showed that Omtzigt’s non-existent list would get 23 seats, but still.) Polls ask the question, “If there were elections today, how would you vote?” Because there are no “today” elections, as a pillar you always have a battle in the event of an incorrect election result. There are ways to indicate the quality of polls, but they are much more complicated than just putting the result next to the forecast.
These margins of error are always misinterpreted. The first poll gave D66 27 seats with a margin of 2. In the election broadcast it was made clear that the result should be between 25 and 29 seats. This is incorrect: according to the arithmetic model, things went well in 19 out of 20 cases (and only if the model was correct), but 24 seats, for example, were also a possibility (and became a reality).
It is also illogical to talk about a margin of error of two for parties that get one seat on the polling day. A negative number of seats in the House of Representatives will be unique. Additionally, smaller fractions have smaller margins of error than larger fractions. The narrative, “A poll can be up to two seats” is not an easy way to explain something statistically complex, but it is simply incorrect.
With the political chaos in The Hague, I don’t dare to predict when the next elections will be, but I hope that the ballot will be used more accurately by then.
Casper Albers Professor of Statistics at the University of Groningen.