Tuesday, April 11, 2017

The problem with the BC Polls

In the last BC Election, the polls were, simply, wrong.

Look at this Forum poll for example. Page 5. You may note that only 64 people who responded claimed to have not voted; this implies an unheard of 95% turnout. This is what originally caught my eye when I went hunting for ways to find polls that are wrong.

Take a look at this poll in particular. Page 4 gives us a nice breakdown we can use.

This is where we get into some math. In the last election the BC Liberals took 44.14% of the vote, so, I will assign them 4414 points. According to this, 81% (or 0.81) are going to back them again. That means the party gets 3575.34 points from these voters. The 9% that say they are voting NDP earns that party 397.26, and so on. The NDP itself gets 3335.64 points for voters who say they voted NDP last time, and will vote NDP again. When you total these points you end up with the BC Liberals taking a 41% share; as such, the Liberals should be actually polling closer to 41% than the reported 38%. The NDP would be at 43%, while the Greens would be at 13%. Contrast this last number with the 17% the Greens are reported to be at.

Put simply: people who plan to vote Green are much more likely to answer the pollster and take the poll.

If you apply this to Federal results, you get 44% for the Liberals, 42% for the NDP, 10% for the Greens, 3% for the Tories, and 1% for other parties.

When looking elsewhere, this pattern holds.
Take a look at the last Forum poll in Alberta in 2012. Notice the raw sample size for each party in the "past vote" 1170 for the PCs and 92 for Wildrose, as well as 211 for the NDP, and 268 for the Liberals. The total sample was 1949. Now there will always be some distortion, but consider that there *should* be nearly two times as many Liberals. In fact if you apply the same math I did above, you'll find this poll *actually* says the PC Party is at 36%, but Wildrose is only at 35% A close race, but it corrects the 3 points Wildrose had been assigned incorrectly. While incorrect about the Greens, doing this also gives us the correct 1% total the Alberta Party would take. In fact, if one assumed a flat 2% of voters, from each of the 3 left-of-centre parties switched to the PC Party to stop Wildrose, as well as the usual jitter in polling, you get your end result, 44%-34%

This is, of course, not an exact way to find the "real" poll results. It is simply a way to find out if the poll in question is close to being accurate or not.

1 comment:

  1. Interesting but I think you are reaching the wrong conclusions. If only 41% of this sample says they voted Liberals last time around, it means the sample is skewed against the Liberals. On the other hand, 43% of NDP means there are too many NDP voters sampled.

    So we should increase the poll numbers of the Liberals and decrease the one of the NDP.

    See what I mean?