538.com map

[Alcon]

Thanks for doing this.

[Alcon]

538 doesn't give Michigan to McCain, at least I don't think so.

Their regression model (weight 0.69) does.

[Alcon]

Good to know.  Have never really sorted out what exactly 538 claims to do .

So, what exactly do these regression maps mean?

What exactly is a regression model?

Regression Q&A 

[Alcon]

Good to know.  Have never really sorted out what exactly 538 claims to do .

So, what exactly do these regression maps mean?

What exactly is a regression model?

Regression Q&A 

So, just to clear this up....given the polling data he gets (national & statewide), he fits this model of his to that data...and the 'regression map' gives the prediction of that regression model?  Pretty cool.

Yep, you got it.  It is pretty cool.

Just to clarify:  The regression map is actually just created (and not used on the site) from his listed regression model.  The regression data is weighted to 0.69, as you can see in the state data area on the 538 page.  Its composition of the total depends on the number of polls, their size, their age and the quality of the pollster.  In D.C., it's 100% of the prediction, because there are no reasonably recent polls.  In Ohio, his model bows heavily to poll data; the regression is weighted to 6% of the prediction.  So, the regression is only (an often-small) part of his projection model.

Poll data is also partially trend-adjusted (not sure at what weight) relative to national polls.  If anything, it's an overly conservative system -- it takes about a week, by my estimates, to fully appreciate a trend and consider it "real."

It's basically all the stuff we tried to account for in our heads, quantified, except it obviously doesn't check for whack internals.

[Alcon]

The Snapshot includes the regression, but also includes the polling averages, which factor into the regression.  The regression is just another data point, not a prediction of any kind.  Just for clarity's sake

[Alcon]

Today's update brings Obama's win chance to 90.5%.

[Alcon]

Nate assumes a poll is good until it's proven bad, not vice-versa.  I disagree, but whatever.

[Alcon]

WHY THE HELL DID NATE SILVER WEIGHT THE LA TIMES POLL SO DAMN HIGH?!

I'm thinking of taking the site off of my bookmarks.  It's becoming freaking aggravating.

Nate's methodology has its screwy points, one of which is probably over-emphasizing polls with limited track records.  But he can't weigh based on how much he likes a poll.  That being said, I wish he included more past record stuff from Presidential elections so as to down-weigh that sort of thing.

But he can't just go, "LA Times, kerplunk," or anything.

[Alcon]

How would you suggest he adjusts for internals?  That's pretty damn complicated stuff -- although I agree using primary performance as the exclusive baseline is a bad idea.

[Alcon]

Image Link