Friday, October 12, 2012

Nate Silvering It Up

A couple of weeks ago, I made predictions about the upcoming November election for the U.S. Senate.  In the section regarding Virginia, I made an assertion that I could more easily picture a Tim Kaine/Mitt Romney voter than I could a Barack Obama/George Allen voter.  But then I got to thinking about whether or not that was really the case, that there were some combinations of Republican senate candidates that go better with Obama and some Democratic senate candidates that go better with Romney.  So I started keeping track of poll numbers from to see if there was anything that I could glean from poll data from states with senate contests this year.  Short answer: nothing interesting.  But in the process of gathering data, I started making charts for the polls, and I made this way of displaying poll data that I don't think I've seen much of before.

It makes no sense to me to compare one day's poll releases with the next day's poll releases.  Polls done by different firms often use different methodologies to come up with different numbers that represent different viewpoints on how the electorate will vote.  However it does make sense to me to compare one company's poll of a state with the same company's poll from several weeks earlier.  The difference from that previous poll should be a better indicator of movement of opinion in the electorate rather than difference in polling methodology.  

So I took the poll numbers I had gathered, plotted them up, and connected each polling firm's dots.  The aggregate sticks and dots looked interesting, and sometimes one could see a clear trend in the polls, but just as often the graph looked like a lot of noise.  So I plotted another line, a thick line representing the median value of active polls, defined as the most recent polls from each polling firm.  This method has its advantages and disadvantages.  One advantage is that it is less volatile: the median value doesn't move around as much as the average poll values of other sites, and it will not be thrown off as much by one outlier poll.  But this means that it doesn't capture the reactions of the electorate to events in the campaign, like debates and conventions, unless the poll movement is sustained for awhile.

So here are some graphs of some swing states including the many polls released today, October 11.

Florida: Obama leading by 1 (though this poll doesn't include the poll released late in the afternoon from Mason-Dixon showing Romney up 7)

Ohio: Obama by 6

Virginia: Obama by 2

Wisconsin: Obama by 4.5

These are all a couple of points more friendly to Obama than most others are predicting, I think mostly because I use many polls that are a month old if they're the most recent poll produced by a firm.  I include Wisconsin in these graphs because the graph looks like a really abstract version of a horse or a turtle.

Here are some of the more interesting Senate graphs.
Arizona: Flake up by 2 over Carmona, but Carmona's support level skyrocketing in September. (note: needs more polls)

Connecticut: Chris Murphy up 3.5 over Linda McMahon, but Linda McMahon's support softening since late August.  I don't understand why she was trending up in August, so I couldn't possibly explain why she is trending down now.

Florida: Bill Nelson up 9 over Cornelius McGillicuddy IV.  The gap is wide and keeps getting wider.  Notice how the same polling firms that produce very precise presidential numbers are all over the place when it comes to Senate numbers.  
Massachusetts: Elizabeth Warren by 3. It took until mid-September, but Warren is finally polling consistently well in that very blue state.

Missouri: McCaskill up by 1 over Akin.  Akin's infamous legitimate rape gaffe occurred on August 19th, and a week and a half later the median poll value showed a legitimate polling swing from Akin +4 to McCaskill +5. 

Nevada: Dean Heller up by 5
 Ohio: Sherrod Brown by 8.

Virginia: Kaine up by 3, much less poll variance than Ohio or Florida.

Wisconsin: Tammy Baldwin up by 3.5.  This is another one where I don't know what happened to make Thompson's support fall in September.  There were a ton of polls in mid August showing Thompson up by a decent margin, and then there was a polling drought for a month, and by the time mid September rolled around, Baldwin was on top for some reason.

I'm going to be interested to see how these simple median poll predictions stack up to the actual election and all the other much more complicated algorithms.  In the meantime, I will update these graphs and add new ones in future posts as new polls come out (right now I only have states with close-ish senate races, but I may or may not add more).

No comments: