As you almost certainly know, hotly contested primary and special elections are being held today in what NBC's Chuck Todd has dubbed Super Senate Tuesday because of closely watched Senate primaries in Arkansas, Kentucky and Pennsylvania. I'm most most intrigued with how the polls perform in the two Democratic primaries involving incumbent senators (Blanche Lincoln in Arkansas and Arlen Specter in Pennsylvania) because of the potential lessons they may teach about what to look for in polling on incumbents for the rest of the year (Chris Cillizza's readers apparently share my interest, as they prefer the moniker "Incumbent Armagedon" for today's contests).
Our final trend estimates show Senator Blanche Lincoln leading challenger Bill Halter by nine percentage points (45.4% to 35.9%) in Arkansas and virtual dead-heat in Pennsylvania between Senator Arlen Specter (42.5%) and Joe Sestak (43.1%). If you use our chart's smoothing tool to pick the more sensitive estimate, the chart gives Sestak a roughly three point lead (44.7% to 41.4%), mostly because it gives greater weight to the one "outlier poll" from Suffolk University that has Sestak leading by 9 points. Three others all have the Pennsylvania race well within their respective margins of error.
In the last three general elections, polls have collectively been mostly unbiased, and as such, polling averages and trend estimates that minimize sampling error have been very accurate predictors of the final margins. Primary elections, however, have often been a very different story. Individual polls tend to be more variable and there are many examples of contests where the final polls have collectively missed the final margins by a mile -- most recently in the Democratic presidential primaries in 2008, especially New Hampshire, South Carolina and the Virginia's Democratic primary for Senate in 2009.
There are many reasons why primary election polling is more error prone -- see AAPOR's investigation of the 2008 primaries for a complete catalogue -- but I'll highlight three: Smaller electorates are harder to sample and model, final snapshots sometimes miss late trends, and the lack of party preferences leave more voters feeling truly uncertain about their choice until confronted with a ballot.
All of that is a long way of saying: I wouldn't be at all surprised if the actual results in Arkansas, Pennsylvania and Kentucky turn out differently than final polls suggest.
And that brings me to my interest in the two Democratic primaries featuring incumbent Senators. Polls in both contests feature relatively large percentages of the vote that are still undecided -- 11% to 15% in Arkansas and 11% to 16% in Pennsylvania. The safest interpretation is that many of these uncertain voters are unlikely to vote, but some see the potential for what some pollsters used to call the "incumbent rule," the idea that incumbents "get what they get" on the final poll an undecided voters "break" decisively for challengers. Such a pattern would put Sestak comfortably ahead in Pennsylvania and would have Lincoln falling short of the 50% necessary to avoid a run-off in Arkansas.
The most concise answer to that comes in the form of a "tweet" last night from our intern, Harry Enten:
Can we PLEASE STOP with undecideds break for challenger! THEY HAVEN'T IN 10 yrs. THEY DIDN'T IN TEXAS GOP OR ILLINOIS DEM GOV PRIM EITHER!
He is right that the pattern we saw in polling in the 1980s and 1990s largely disappeared over the last ten years. However, we did see something of an incumbent effect in the New Jersey Governor's race last year, especially in polls conducted with live interviewers. The fact that Republican Chris Christie did consistently better on automated survey, and that the automated surveys came much closer to the final margins, suggests that the undecided respondents on live interviewer surveys "broke" decisively in Christie's favor.
As such, note that all of the final week polls in Pennsylvania and Arkansas used live interviewers, and the anti-incumbent sentiment evident last fall has only intensified over the last six months.
Of course, all of this just speculation, though there are clues in the final surveys themselves. Feeling especially curious about Pennsylvania, I asked the pollsters at Quinnipiac and Muhlenberg Universities to run run cross-tabulations of the favorable ratings of the two candidates among the voters who were undecided on their final polls. These tabs yield roughly 150 undecided respondents in the Quinnipiac poll conducted last week and exactly 111 undecided respondents interviewed by Muhlenberg University from May 6 to May 16.
The results, as shown in the table below, are more or less consistent: Both show most of these voters to be disengaged. Roughly two-thirds don't know or can't rate Sestak and nearly half are unable to rate Specter. Only 18% of the undecideds on the Muhlenberg poll and 19% on the Quinnipiac poll rate Specter unfavorably. Put simply: it's hard to see a big break to Sestak coming from the undecided vote.
Of course, undecided voters are just one potential source of error. As always, we'll just have to wait for the actual results later tonight.