2016 Postmortem
Related: About this forumClinton - Trump statistically tied .... in Alaska
It's 37-34 with Trump ahead by 3% points but that is within the poll's MoE of + 4%
link - http://www.realclearpolitics.com/docs/2016/Moore_AK_Oct_2016.pdf
Seems amazing that AK is at all close. But I like it
Loki Liesmith
(4,602 posts)The larger number is the number that is more probable to be ahead
Foggyhill
(1,060 posts)That's certainly not the case for just one poll
Especialy if you add all the other
Possible methodological failings of 90% of those polls like various types of sampling biases
If you got 10 polls and one party is always ahead it may mean something if the poll is well constructed Even If it is within the margin for just one poll
Polling is not just about stats and most are pretty bad
Imperialism Inc.
(2,495 posts)So, it's just as likely to be -6 as it is to be tied, and both are far less likely than it is for the real number to be close to -3.
Loki Liesmith
(4,602 posts)The mathematics underlying confidence intervals require that at a 95% confidence, about 1/20 polls will be from a unrepresentative sample. Which means a lead outside the CI can happen about 1 in 40 times. A lead switch inside the CI will happen more frequently, So a smaller confidence interval (90% CI) you may get a lead switch 1 in 20 times. 19 out of 20 times the party ahead will be the true leader. A lead of 1% point will usually correspond to a ~60% CI, so 4 times out of ten you will get unrepresentative sample, and 1 time out of 5 you will get a lead switch. Most of the time the party in the lead will be the true leader.
Foggyhill
(1,060 posts)That's only true if yoûre doing a random poll of the underlying pop you want polled
Most polls can't claim to even come close to that because of methodological issues
It's expensive and hard to get this random sample and most poll fail
If your sample is biased, the margin of error or the 1/20 means nothing at all
Got a hell of a lot of experience in stats and some in polling too
That's why you have so many polls all over the place, it's not just about the margin of error there are many problems with the sampling itself
If you got more than one decently run poll saying the same thing, it does say something though collating them together is pretty iffy. Doesn't stop people from doing it though