72 hours from election day! Who's going to win?
Nate Silver says that "For Romney to Win, State Polls Must Be Statistically Biased." I think this is correct. Silver's model has Obama at a 83.7% favorite, and that is likely to turn into a 90-95% favorite by election day the way his model works. Silver's argument for Obama being a favorite: he's ahead in Ohio, and if he's ahead in Ohio, he's substantially more likely to reach 270 electoral votes. This is also correct.
That said, I do believe that there is a substantial chance that the state polls are a biased prediction of where election day results turn out. Because the 538 (and Princeton) model rely so heavily on the assumption that the state polls are unbiased, if the state polls are biased, then the 538 and Princeton models will also be biased. (Note that "bias" is a neutral statistical term; it does not mean that the pollsters are conspiring to favor Obama, just that their likely-voter model is producing results that overestimate his standing.)
1) Polls have historically had a bias against Republicans relative to how election day turns out. Silver disputes this; I've explained why his analysis on that question might be wrong before. On the other hand, the 2012 polls may have fixed this problem.
2) Polls have historically had a bias overestimating the performance of sitting presidential incumbents. There's certainly a counterargument against this. On the other hand, perhaps Kerry's failure of a 2004 bounce reflects the offsetting bias against Republicans. On the other other other hand, perhaps the 1980 and 1996 collapses by Democratic incumbents from their polling positions reflect the distortion caused by a relatively strong third-party candidate, plus Clinton's 1996 decline might simply reflect a reversion to the mean. But if undecideds break 75-25 against Obama, that's a 1- or 2- point shift relative to the polls.
3) The state poll toplines, which is all that the 538 and Princeton models use in their calculations, show enormous advantages for Democratic turnout in state races, advantages greater than even Obama realized in 2008. I think that this is very likely an error that, if corrected for, will show material change. There are many many reasons to think that the Democratic advantage of 2008 will be weaker or nonexistent in 2012, when Obama won 53-46 nationally (and Ohio by 4.6 points).
(a) Obama leads Obama voters only 84-13. Now, some 2008 McCain voters have died or won't vote for Romney; some 2012 Obama voters will be new voters; some 2012 Obama voters will be 2008 McCain voters; that 84-13 may well turn into 90-10 on election day. But it's hard to see where the improvement in D turnout is going to come from when so many Obama voters are disaffected.
(b) Independents overwhelmingly favor Romney, by 20 points in some polls. It's implausible that independents have swung so wildly, yet Democrats are more enthused and more likely to turn out than Republicans. Some, such as Josh Marshall, have posited that the makeup of independents has been changed because disaffected Perot/Tea-Party/Paulite conservatives have left the Republican party. Dan McLaughlin's counter-argument strikes me as stronger. Even under Josh Marshall's calculation of voter ID, D+7 among "all adults" should not translate into a 2012 that is better than 2008 for Democrats: Republicans are more likely to vote; Republicans are more likely to vote for Romney than Democrats for Obama (especially given the shift in Catholic opinion since 2008); and Marshall necessarily concedes that the independent vote will be less pro-Obama than in 2008. And I don't believe Josh Marshall's claim that the current makeup of all adults is D+7.
(c) McCain intentionally hamstrung his campaign in 2008 with unilaterally disarming self-abnegating campaign-finance limitations, next to no on-the-ground spending for get-out-the-vote efforts ("GOTV"), and wasting money on such tactics as a national ad congratulating Obama on winning the nomination. This undoubtedly made McCain feel better about himself, but meant that the GOP had a historically unprecedented disadvantage in GOTV. Even with the supposed 2012 Obama advantage in field offices (which neglects the fact that GOP turnout efforts are largely volunteer), Romney is doing comparatively better in 2012 than McCain did in 2008, which is the relevant comparison for whether the state polls are accurately predicting that Democratic turnout advantage will be higher in 2012 than in 2008.
(d) Some polls show Republicans more excited about this election than Democrats; others show the excitement level about even. But even taking the latter polls as an accurate metric, the relevant comparison is again versus 2008, where the Democrats had a huge excitement advantage. If the Democrats are going to outperform 2008, which they have to achieve the slim lead they achieved in 2008 given the shift in independents, they need to be more relatively excited than in 2008. They're not.
(e) The Obama campaign is assuming turnout will be 72% white; polls are assuming turnout will be 74% white. But there's a strong argument that turnout will be more like 75% white. Whites favor Romney by about 61-39; non-whites Obama about 80-20. This would be a shift of 0.8% relative to the polls.
(f) Reports about early voting are mixed, with National Review and Battleground Watch reporting relative advantages for Romney versus 2008, and Democrats claiming otherwise. This election will likely have a record percentage of early voters, and it's hard to say to what extent that reflects cannibalization or improved turnout. We'll disregard this as ambiguous.
The Princeton model takes the state polls at face value; Nate Silver has made a decision to have the 538 model disregard each of these factors and assume that the polls are as likely to be biased for Obama as for Romney. Thus, there is a substantial chance that the 538 and Princeton models are reflecting a house effect. GIGO: garbage in, garbage out. McLaughlin's post on the subject is a must-read; so is Kevin Holtsberry on Ohio; Michael Barone makes the aggressive case for Romney.
All that said, if I had a gun to my head the morning of November 3, I would rate Obama a slight favorite, with about 60% to 40% chances, a movement from the tossup I suggested on Twitter. (I'm not going to succumb to the fake precision of Nate Silver's model and suggest I can forecast with three significant digits.) Why?
(1) Each of the effects I mention above are overlapping to some extent. You can't add them up to say that the polls are biased by seven points. And Obama has some lead in the state polls, plus has more favorable electoral-college scenarios than Romney in a close election: the so-called "firewall." Thus, I could be right about three, four, or even all seven of those points about state-poll bias, and it still might not add up to enough of an effect to put Romney over the top.
(2) I have to recognize that there's some chance that I'm suffering from confirmation bias, looking only at the data that favors my preferred result, and that I'm wrong. Nate Silver is spending a lot more time thinking about the election and polls than I am, and it's more likely that he's seeing things that I'm not than I'm seeing things he's not; maybe if I were to try to duplicate his efforts, I would end up with a model that was closer to his than my current thoughts are. I thus have to discount my thinking that Romney is doing better than Silver thinks. I think partisans on both sides have been silly about Silver, with some Democrats treating him as infallible to the point that a challenge of Silver is a challenge of scientific truth, and some on the right being simply innumerate in their attacks on Silver. Still, occasionally Silver lets the mask slip and makes a statement that's pretty close to mine about the unpredictability of the election and the chances his model is wrong.
(3) There are pollsters who do take the effects I mention into consideration somewhat; in particular, Rasmussen is very aggressive in doing so. But even Rasmussen shows an election that is close to a tossup, with Obama having more "outs" in the electoral college than Romney. And I think there is a reasonable argument that Rasmussen is too aggressive in its likely-voter model.
(4) A certain number of intangible factors favor Obama. Hurricane Sandy changed news coverage dramatically in the last week of the election: a disaster brings Americans together, lets a president seem presidential, and, in this particular case, provides a conscious or subconscious reminder of Bush administration incompetence. Moreover, every minute spent on dramatic Sandy stories is a minute the news isn't spending on Obama administration failures in the economy, in energy policy, in deficit reduction, in Libya. It's hard to think that the double-standard of "fact-checking" that more harshly (and often incorrectly) scrutinizes Romney claims than Obama claims doesn't have some effect on undecided voters.
(5) What am I supposed to think about Romney throwing $12 million into Pennsylvania at the last minute? Pennsylvania has consistently voted more Democratic than the rest of the nation. It's hard to imagine a scenario where Romney wins Pennsylvania but doesn't win Ohio and/or 270 non-Pennsylvania electoral college votes. The Pennsylvania U.S. Senate race doesn't seem particularly close. There are four possibilities: (a) the Romney campaign sees a real "undertow" election, and legitimately thinks it can get meaningful coattails in Pennsylvania; (b) the Romney campaign is panicking, and throwing a Hail-Mary in Pennsylvania because it doesn't think it can get to 270 without it; (c) the Romney campaign has more money than it can spend efficiently, and is throwing money at Pennsylvania as an expensive feint to divert Obama campaign resources; (d) the Romney campaign doesn't know what it is doing or inefficiently hoarded money. If one assigns Bayesian probabilities to each of those scenarios, that doesn't look good for Romney.
(6) Finally, the way that Romney (and American Crossroads) are running campaigns suggests some problems as well. The Romney campaign feels that it has to play defense on things where Republicans are correct (such as the Lilly Ledbetter Act); it is affirmatively trumpeting protectionism instead of defending free markets. American Crossroads is blanketing Virginia with advertising attacking Democrats from the left for Medicare and education spending cuts. Romney hasn't made the Supreme Court an issue at all in the campaign. I'm clearly not the swing voter who is being targeted by these ads, and perhaps they reflect an attempt to dampen Democratic turnout. But it's hard not to think that Romney doesn't think he can win the election by selling conservative and free-market ideas. That has longer-term consequences even beyond what it says about the 2012 election.
We'll find out on Tuesday night or late Wednesday morning. That said, it's important to realize that this election is only one data point, and in the absence of an unexpected 55-45 landslide one direction or the other, the election will not be proof or disproof of any particular model of forecasting elections.
My swing-state predictions:
Likely Romney: FL, NC
Lean Romney: CO, VA
Too close to call, slight lean Obama if I have to choose: OH
Lean Obama: IA, WI
Likely Obama: MI, NH, NV, PA
Solid Romney: The other McCain 2008 states, IN, Nebraska 2nd District
Solid Obama: The other Obama 2008 states
Thus: Either 281-257 Obama or 275-263 Romney. Note, though, that the electoral-college scenarios generally favor Obama. Obama can lose WI or IA if he wins OH; if Obama wins IA and WI, he only has to win one out of CO, OH, and VA to get to 270. Note further the small chance of a Romney popular victory and Obama electoral college victory, especially if New York/New Jersey turnout is depressed.
Because of this electoral-college firewall, if the leaning states were really independent toss-ups, Obama would have a much better than a 50-50 or 60-40 chance of winning. In fact the errors are going to be correlated, it's unlikely that Obama wins Virginia and Colorado but loses Wisconsin and Ohio, and my 60-40 estimate reflects a certain chance that Silver/Wang is completely right, a certain chance that I'm right or partially right, a small chance that Rasmussen is completely right and I'm being overconservative, and a very small chance that the state polls are completely wrong and we're seeing an undertow election like Barone and Domenech think.
(Updated to add links, fix typos, and add specific swing-state predictions.)
Update, November 4: One more reason to be a bit more skeptical of the polls.