## Friday, February 26, 2016

### Stop the presses!

I was part way through a post on another topic when I heard a story about how Donald Trump was predicted to have a 97% chance of beating Hillary Clinton in November  and a 99% chance of beating Bernie Sanders.  The prediction, by Helmut Norpoth of Stony Brook University, is described in a lot of stories, but this one gives the most details.  He's been developing the model for a number of years, and this paper gives a prediction for the 2012 election.   At that time, the formula was:

y=48.2+.445PI-.138PO+.366y1-.338y2

y is the Democratic share of the two-party vote, while y1 and y2 are the Democratic shares in the previous two elections.  PI is the the share of the vote among the top two finishers in the New Hampshire primary for the candidate of the incumbent party, in this case the Democrats.  PO is the share for the candidate of the opposition (Republican) party.  The values of the primary variables are limited to 35-65--if you get more than 65% it's counted as 65; less than 35% is 35.  So the key idea is that candidates who do well in the NH primaries do better in the general election, which seems reasonable.

Trump is predicted to get 54.7% of the two-party vote against Clinton.  The other potential Republican candidates are predicted to get just under 50%.  Trump would get a 65 on PO, while any other would get 35 (Trump got 35%, and second-place finisher Kasich got 16%, meaning Trump got about 68% of the top two share).  So a 30 unit difference in PO gives about a 5 unit difference in y, which is consistent with the formula above.

But Sanders is predicted to do slightly worse than Clinton, losing narrowly to a non-Trump candidate, even though he won the NH primary easily, with about 61% of the top two vote.  With the formula above, he would be predicted to run almost 10 points better than Clinton, meaning an easy victory against Trump and a historic landslide against any of the others.

It sounds like the current predictions are based on a formula using the first few primaries, where Clinton and Sanders are running about even.  But the paper I linked to above investigated the issue and concluded that the New Hampshire primary was the only one with any predictive power.  So if you accept the Norpoth formula from 2012, you should not believe the Norpoth predictions for 2016.  Personally, I wouldn't put much faith in either one, but that's a subject for another post.