Predictions: Kentucky and Oregon

Here are my predictions for today's Democratic primaries in Kentucky and Oregon.

In the national polls Sanders has been losing 0.5-1%/week versus Clinton. The Google Search Trends reflect this. Previously Sanders was getting around 65% of the combined Sanders + Clinton search trend, but now this has fallen to around 56%. For instance, in the past 7 days even though it is one of the few states which has an active campaign left -- Sanders has only got 53.5% of the searches in Kentucky.

So while Sanders is campaigning hard in Kentucky (as is Clinton), will that be enough to go against the decline in enthusiasm of his supporters as they realize that his odds of winning are very small?

Both Oregon and Kentucky have a lack of polling making predictions very uncertain. Kentucky has a single poll from March 1 - 2, with Clinton 43 and Sanders 38.

Oregon has a May 6-9 poll by DHM that has Clinton 48% to Sanders 33% - which is especially strange considering the same polling company had Clinton only ahead by 44 to 39 in July 2015. Sanders support has doubled (or more) since July 2015. Someone in the PredictIt comments wrote that the 2015 poll used bad methodology and over-sampled Sanders supporters - which could easily be true. Oregon also has a second poll from March 28-April 1 with Clinton 37 % to Sanders 36% which I'm ignoring because it is old (though in retrospect giving it equal weight to the DHM poll is probably more likely to give an accurate prediction - but that is also cherry picking based on the fact that I don't like the DHM poll).

By contrast to the weak polling, the FB likes and Google Search trends for Sanders are very strong in Oregon.

I decided to run my models with and without polling. There is a strange effect where the model with the Oregon poll gives Sanders a higher prediction, even while the poll strongly favors Clinton.

Note I am predicting the candidate's share of the Clinton + Sanders vote. I'm ignoring other candidates (and they might get several percent in Kentucky).

KY forecast (with the single latest poll): Clinton 56.9%, Sanders 43.1%
KY forecast (without polls): Clinton 57.8%, Sanders: 42.2%

OR forecast (with the single latest poll): Sanders: 62.2%, Clinton 37.8%
OR forecast (without polls): Sanders 60.1%, Clinton 39.9%

My Kentucky forecast is an outlier among the models that I have seen. Other people are predicting a smaller lead for Clinton or even a small Sanders win. One unknown in Kentucky is how many Trump supporters are registered as Democrats (as the primary is closed). Trump supporters and other conservative Democrats typically vote for Sanders (as they did in WV).

Both of these primaries are closed. My model has found the "closed" variable to be statistically not significant. This is a bit strange considering that independents vote overwhelmingly for Sanders in states where Sanders has moderate to strong support (eg he doesn't win a majority of independents in areas of the South where he only gets 20% of the votes).

KY final result was Clinton

KY final result was Clinton 46.8, Sanders 46.3 (or normalized to 50.2 / 49.8)
OR result so far is Sanders 56.0, Clinton 44.0

KY poll model error: 6.7%
KY no poll model error: 7.6%

OR poll model error: 6.2%
OR no poll model error: 4.1%

So the no poll model did slightly better (11.7% vs 12.9%).

Several people were arguing that Facebook and Google Search Trends would be less significant for Kentucky due to less broadband internet access, however this looks like they were less significant for both Kentucky (where Clinton did well in those variables) and Oregon (where Sanders did well). Or it could be that they did poorly in Kentucky and Sanders did poorly in Oregon due to other reasons - notably the growing realization that the campaign is over (and growing desperation among some - see the crazy fight over 1-2 delegates at the Nevada convention).

My model actually didn't include the WV results, so that could have adversely affected it as Sanders has strong support in WV and a dozen or two other Appalachian counties.

My Oregon no poll model was one of the most accurate (Benchmark had Sanders at 51, Tyler had Sanders at 70).

Result Update

KY result stays the same
OR: Sanders increases vote share from 56% to 57.2%, and my error decreases by 1.2%.

More predictions