A front-page post on myDD today suggested that favorability ratings were pretty useless in determining the winner of this month's Senate elections.
In a one-off comment, I suggested that once factoring the PVI of a state into the equation, much better results emerge, which come very close to nailing the election results in most cases.
After I made that post, I did an analysis of the 10 widely watched Senate races with incumbents (MI, MO, MT, NJ, NY, OH, PA, RI, VA, and WA), and came up with a pretty good and sane methodology for working favorability ratings into election results.
The results are on the flip.
For each state, I came up with an incumbent's PVI by averaging the amount by which the candidate of their party exceeded or underachieved the national percentage of the vote in their state.
For example, in 2000, Gore got 48% of the vote nationally. He got 51% of the vote in Michigan. In 2004, Kerry got 48% of the vote nationally. He got 51% of the vote in Michigan. So Stabenow's PVI would be +3.
In addition, I assumed that the extent by which the national PVI would affect a candidate was related to how much they were in line with their national party. I used the National Review's ratings (not because they are necessarily the most accurate, but because they seem to be in line with public perception, which is what matters here).
I multiplied the PVI number by the rating (for incumbent Democrats, I multiplied by the liberal rating; for incumbent Republicans, I multiplied by the conservative rating), and added it to their last SUSA favorability rating to get their expected vote total.
But there are two more factors in play. The first is the MoE of the actual poll (around 4% across the board). The second is election-day factors, like GOTV and other field operations (responsible for around 2%, says the experts).
So I took the expected total, and gave it +- 6% in either direction. The results follow:
Michigan (incumbent Democrat Stabenow):
PVI: +3% Democrat
Liberal Rating: 86%
Last Favorability Rating: 49%
Expected Vote Total: 51.6%
Actual Vote Total: 57%
MoE: 4.1%
Spread: 45.5% - 57.7%
WITHIN SPREAD
Stabenow ended up pretty close to high end of the spread, probably due to the fact that leaners broke decidedly for Stabenow.
Missouri (incumbent Republican Talent):
PVI: +2.5% Republican
Conservative Rating: 66%
Last Favorability Rating: 44%
Expected Vote Total: 45.7%
Actual Vote Total: 47%
MoE: 4.1%
Spread: 39.5% - 51.8%
WITHIN SPREAD
Talent slightly overperformed his expected vote total, but well within the MoE.
Montana (incumbent Republican Burns):
PVI: +9% Republican
Conservative Rating: 73%
Last Favorability Rating: 40%
Expected Vote Total: 46.6%
Actual Vote Total: 48%
MoE: 4.1%
Spread: 40.5% - 52.7%
WITHIN SPREAD
Burns slightly overperformed his expected vote total, but well within the MoE.
New Jersey (incumbent Democrat Menendez):
PVI: +7.5% Democrat
Liberal Rating: 92% (Progressive Punch)
Last Favorability Rating: 38%
Expected Vote Total: 45%
Actual Vote Total: 53%
MoE: 4.1%
Spread: 38.8% - 51%
OUTSIDE SPREAD
Menendez well outperformed his expected vote total. It's probably too much to ask a state with a PVI of +7.5D to vote for a Republican.
New York (incumbent Democrat Clinton):
PVI: +16.5% Democrat
Liberal Rating: 80%
Last Favorability Rating: 67%
Expected Vote Total: 80%
Actual Vote Total: 67%
MoE: 3.8%
Spread: 74.4% - 86%
OUTSIDE SPREAD
Hillary had very high favorables and a very high PVI. They combined in this case to produce unrealistically high targets. In this case, her favorables and the state PVI were highly correlated. Hillary's actual vote total was actually identical to her favorables.
Ohio (incumbent Republican Dewine):
PVI: +1% Republican
Conservative Rating: 55%
Last Favorability Rating: 38%
Expected Vote Total: 38.6%
Actual Vote Total: 44%
MoE: 4.1%
Spread: 32.5% - 44.7%
WITHIN SPREAD
DeWine actually well outperformed his expected vote, but was still within the spread. This probably has to do with the fact that late-breaking voters broke much better for Dewine than early deciders (Dewine got 48% of voters who decided in the last three days), which suggests that the race dynamic changed after the last favorability poll (Oct. 16). As further proof of this, Dewine got only 42% of the vote from those who decided more than a week before the election, which is in line with out expected numbers.
Pennsylvania (incumbent Republican Santorum):
PVI: +2.5% Democrat (-2.5% Republican)
Conservative Rating: 70%
Last Favorability Rating: 38%
Expected Vote Total: 36.3%
Actual Vote Total: 41%
MoE: 4%
Spread: 30.3% - 42.3%
WITHIN SPREAD
This is another case of a change in the race dynamics toward the end of the race. Voters who made up their mind more than a week before the election (i.e. voters who made up their mind at the time of the poll on October 16) gave Santorum 38% of the vote, pretty much as expected.
Rhode Island (incumbent Republican Chafee)
PVI: -13%
Conservative Rating: 40%
Last Favorability Rating: 46%
Expected Vote Total: 41%
Actual Vote Total: 47%
MoE: 4.1%
Spread: 35% - 47%
BARELY WITHIN SPREAD
Chafee outperformed the expectation here probably for two reasons: (1) he probably wasn't even considered a 40% conservative in RI; (2) His Republican-ness was factored into his low favorabilities here (the last poll before the election was the only time in the entire race that his favorables dipped below his unfavorables, and they promptly shot right back up after the election was over). Late-breaking voters voted 49% for Chafee.
Virginia (incumbent Republican Allen):
PVI: +3.5% Republican
Conservative Rating: 86%
Last Favorability Rating: 47%
Expected Vote Total: 50%
Actual Vote Total 49%
MoE: 4.1%
Spread: 43.9% - 56.1%
NAILED IT
Washington (incumbent Democrat Cantwell):
PVI: +3% Democrat
Liberal Rating: 77%
Last Favorability Rating: 50%
Expected Vote Total: 52.3%
Actual Vote Total: 57%
MoE: 4.1%
Spread: 46.2% - 58.4%
WITHIN SPREAD
Cantwell slightly outperformed her expected vote total, probably due to the fact that McGavick wasn't a great candidate.
Conclusions:
I believe that the polls would be even better predictors of performance had they been taken closer to the elections.
In states with very high PVIs for or against an incumbent (NY, RI), the PVI seems to have been factored into the favorability ratings.
In many races, late-breaking voters broke for incumbents, according to the exit polls published on CNN's website. This accounts for a number of the overperformance of incumbents like Stabenow, Dewine and Chafee (and justifies my padding of the MoE by an additional 2 points).
The bottom line is that once PVI is adequately factored into the equation, favorability ratings are actually quite useful as predictors. The reality, though, is that so much changes in the last week or so of the campaign that any remotely close race will remain to close to call regardless of favorabilities.