Tuesday, July 26, 2016

State polls and date relevance - by Vox Day

DH has been resolutely predicting a Clinton win on the basis of the state polls, which correctly predicted Obama wins in the last two elections.
The state by state projections as of today including all most recent polling still indicate a Clinton win with around 312EV. Trump has not altered the road map at this point. Including polling changes that could happen between now and election day, Sec. Clinton is cruising towards victory. 
Although I respect DH's acumen and take him very seriously with regards to anything that involves data analysis, I am nevertheless predicting a Trumpslide, a win of even bigger proportions than 312 Electoral College votes for Trump. How is it possible for me to do that considering the supposedly reliable evidence of the most recent state polling that DH is citing?

The reason is pretty straightforward. While the state polls have been pretty good predictor of the election results, my suspicion was that this is only true of state polls taken in the last week prior to the election. Before that, they tend to bounce all over the place. Unlike the national polls, they don't always tend to favor the Democratic candidate, then fall more in line as the election approaches; the state polls appear to be less corrupt than the national ones.

Allow me to demonstrate. I looked at the results of the McCain-Obama race, since that one was more similar to the current race given that it also lacked an incumbent, in all seven of the states identified as key "battleground" states. In each case, I listed the following:
1.    The earliest date that any state poll got the correct result.
2.    The latest date that any state poll had either a) the wrong candidate winning or b) a tie
3.    The RCP average of the final state polls from the last week prior to the election
4.    The actual results.
Rasmussen 2/14 - 2/14 Obama +10
FOX News/Rasmussen 9/14 - 9/14 TIE
RCP Average: Obama +7.3
Final Results: Obama +10.3

SurveyUSA 2/15 - 2/17 Obama +6
Mason-Dixon 9/29 - 10/1 McCain +3
RCP Average: Obama +4.4
Final Result: Obama +6.3

PPP  9/27 - 9/28 Obama +3
FOX News/Rasmussen 11/2 - 11/2 McCain +1
RCP Average: Obama +1.8
Final Results: Obama +2.8

Quinnipiac 9/5 - 9/9 Obama +5
Mason-Dixon 10/29 - 10/30 McCain +2
RCP Average: Obama +2.5
Final Results: Obama +4.6

SurveyUSA 2/26 - 2/28 Obama +9
Denver Post/Mason-Dixon 9/29 - 10/1 TIE
RCP Average: Obama +5.5
Final Results: Obama +9.0

Rasmussen 10/8 - 10/8 Obama +1
Reuters/Zogby 10/31 - 11/3 McCain +1
RCP Average:  McCain +0.4
Final Results: Obama +0.3

Associated Press 10/22 - 10/26 Obama +12
Politico/InAdv 10/19 - 10/19 TIE
RCP Average:  Obama +6.5
Final Results:  Obama +12.5

So, as early as FEBRUARY there were three battleground polls that correctly predicted the result, but in four other battleground states, there was not a single poll among the dozens that were taken that correctly predicted the result until October, or in two cases, November. Since none of the three correct polls were performed by the same company, and since in one case, that same poll went on to incorrectly predict a result that was off by 10 points, it's pretty clear that these results were random and therefore unable to serve as the basis for a predictive 2016 model.

Not only that, but there are no February state polls comparing Trump to Clinton, because back in February, Scott Adams, Mike Cernovich, Helmut Norpoth and I were about the only individuals publicly going on the record and stating that Trump would be the Republican nominee.

Now let's look at the latest date that a state poll incorrectly predicted the winner. Even in a state that Obama won by 12.5 percentage points, there were polls in October indicating a dead heat. Mid-September is the earliest date of an incorrect poll; in North Carolina, which was close, even the RCP average had McCain winning right up until the election took place.

Taken in sum, this means that it makes no sense to pay much attention to the state polls until September. What we can before then, however, is the general trend from one candidate to the other; in many of these battleground states, the gradual shift from McCain to Obama, or from leaning Obama to strong Obama, is apparent.

And what do the state polls show in this regard? At the moment, they are too much in flux to clearly read a trend, but they appear to be gradually following the shift from Clinton to Trump already seen in the national polls. So, I see no reason to revise my prediction of a Trumpslide.