That 75,000 is my estimate for the counties with <= 95% counted. For the counties > 95%, I tried out 96%, 98%, and 100% counted.
The current "No" fraction, 62.0%, is a bit less than my prediction from 75% to 94% counted, 62.1% to 62.3%.
The final prediction of 538 was No 57.3% and of RCP was 58.3%.
I couldn't find any discussion of why this victory was so good. Democrats opposing Larry Elder? Republicans being discouraged because of their claims of election fraud?
-
I must discuss ways in which polls can be in error.
One way is sampling bias, reaching an unrepresentative selection of would-be voters. That was the great flaw of the Literary Digest's 1936 poll of the US Presidential election. It was a massive effort, but it focused on such things as car registration lists, things that were biased to more affluent people. By not correcting for this sampling bias, they got an embarrassingly wrong prediction of who would win.
Another way is statistical error. If one repeats random samples from a population, one will get different samples each time, and their statistics will be different. How much different can be calculated for some statistical distribution.
Considering the case of two choices, I set them to 1 and 0, and I find the statistics on the average value and how much it varies.
The choices have a probability p of being 1 and (1-p) of being 0. Their average is p, and the "standard error of the mean", the standard deviation of different samples' value of the mean is 1/sqrt
* sqrt(p*(1-p)) for n samples. That factor of 1/sqrt
is a very general one, and it is very useful for making rough estimates of sampling error.
For 1000 samples and p = 0.62, the SEM is 1.5%
That's for one standard deviation. Using a normal distribution for being less than 1, 2, and 3 standard deviations, the probability is 0.68, 0.95, and 0.9973.
The current result is about 3 stdevs greater than 538's final result, so it's barely more than what one would expect from sampling error.