2 ms·
The idea of a probability based poll is that if you asked 3500 people about something that was a good random sample, you could extrapolate to the whole populat
by PaulHoule 2y ago
The idea of a probability based poll is that if you asked 3500 people about something that was a good random sample, you could extrapolate to the whole population with pretty good accuracy.
The traditional way to do that is call phone numbers at random, but you run into biases such if a man and a woman are at home probably the woman answers the phone -- so you have to weight the results based on demographic characteristics.
One problem is that people are less and less inclined to answer the phone when a stranger calls so it might be that the people who answer the phone are not representative of the general population, overall that hasn't seemed to be a big problem until Donald Trump entered the scene because whenever he's being asked about probability based polls seem to underestimate his level of support. Looking at the tracker here
https://www.natesilver.net/p/trump-approval-ratings-nate-silver-bulletin https://www.natesilver.net/p/trump-approval-ratings-nate-sil...
Trump does a few points better than the tracker here
https://ropercenter.cornell.edu/presidential-approval https://ropercenter.cornell.edu/presidential-approval [1]
The second tracker is simply reporting the results of probability-based polls whereas the first one is applying some weighting to (mainly) non-probability-based polls that use different methods but generally are starting from a sample which is less likely to be representative to begin with but doing more work with weighting to try to make up for it. It's not a huge difference (maybe that $74B is really $69B or $78B or something) but for an election in a particular state the difference between 49.5% and 50.5% is all the difference.
[1] Full disclosure: I helped code this one.