While that is an interesting article, I believe it is wrong, and the polls, therefore, were wrong.
Here's why I think this way.
RCP averages the polls by averaging the Trump poll portion, and then averages the Clinton/Biden portion, and calls this the average poll.
However, this average drops the margin of error within each poll in the average. That's where their failure lies.
As you know, I've been modeling polls going back to the 2010 House races. I use the MOE to convert the poll spread into a probability of the leader actually winning (Nate Silver does the same thing). RCP's article only uses the high/low spread to declare the poll's success or failure (they called it a success).
Consider some real data:
I have 10 polls for Florida throughout September 2016. Simple averaging of the polls gives me Clinton 44.5 Trump 44.2.
Now, take a deeper look at these three polls:
- NYT/Siena: Clinton 41 Trump 40 MOE 3.3%
- Fox13/Opinion Survey: Clinton 47 Trump 46 MOE 4.0%
- Suffolk University: Trump 45 Clinton 44 MOE 4.4%
All three of these polls has the leader ahead by 1%. However, when the MOE is factored in, these polls become very different. Looking at the probability of winning:
- NYT/Siena: Clinton 64 Trump 36
- Fox13/Opinion Survey: Clinton 61.7 Trump 38.3
- Suffolk University: Trump 60.7 Clinton 39.3
The last Florida poll in September was from Mason-Dixon on 9/27-9/29. It showed Clinton 46 Trump 42 with an MOE of 3.5% This converts to a probability of Clinton 90.8 Trump 9.2. The Fox13/Opinion Survey poll above was from 9/27-9/28, and showed a lower, but still leading, probability of winning.
Just taking the last 5 polls from the second half of September, the average (and probability of winning) is:
- Clinton 45.6 (71%)
- Trump 43.4 (29%)
The September poll spread was +2.2 for Clinton. President Trump won Florida 49 - 47.8, or +1.2 for Trump.
RCP can say that the result was within a polling "mean absolute error" of +2.7%, but a +2.2 lead in Florida in September was a 71% probability of winning.
It's a harder argument to make that a poll that predicts a 71% probability of a candidate winning who ultimately loses was a correct poll, when all the others polled the same way. One outlier poll, maybe, but not all of them.
-PJ
Thank you .
I ignore these RCP nonsense polls analyze .
There no logic in averaging corrupt and slanted and biased data from sources like BY Slimes , ABC , NBC , CNN , CBS , Bezos newsletter ,
People read headlines and not study the internals which show how the
skewed results were cooked up .
We are basically we are flying threw heavy fog with no instruments .
I fear the Pres is taking advice only from his clueless learjet lib daughter and son in law who appears to be protecting his Cheap labor express lobby pals and hanging with Val Jones the Commie CNN fraudster.
You cited Nate Silver so I thought his views on the matter would be of interest. Incidentally, I do not pretend to understand the calculations converting margin of error into level of confidence, which Nate Silver considers as well in this article. I am however struck by his conclusions:
it didn’t take any sort of extraordinary, unprecedented polling error for Trump to defeat Clinton. An ordinary, average polling error would do — one where Trump beat his polls by just a few points in just a couple of states — and that’s the polling error we got.
You read that right. Polls of the November 2016 presidential election were about as accurate as polls of presidential elections have been on average since 1972.
I am confident that you are fully familiar with Nate Silver's thinking but I believe others on this thread who reflexively reject polls when they are unfavorable would find his article instructive.
Finally, here is Silver's bottom line The media narrative that polling accuracy has taken a nosedive is mostly bullshit
