Undecided voters probably sometimes break sharply for the challenger. But I can find no evidence that this rule is useful in "allocating" reported undecided voters in presidential elections. In fact, overall, reported undecideds seem to break about evenly.
In support of the claim that undecideds break for the challenger, TruthIsAll has indirectly* cited an essay by Chris Bowers at MyDD. Bowers canvassed 28 presidential polls going back to 1976, and concluded that "undecideds break overwhelmingly -- better than 6 to 1 -- in favor of the challenger in a Presidential race." But a close reading reveals that Bowers did not examine declared "undecided" voters at all!
*[1/6/07: TIA has objected that he had never "cited" Bowers, which is, technically, true. On his Election Model page, he linked to a blog post by Ed Fitzgerald that is solely devoted to Bowers' argument. Fitzgerald did update the post, primarily to quote -- TruthIsAll! I had assumed that TIA intended to direct readers to Bowers' evidence. More below.]
Bowers basically found that on average, incumbents did about 0.3 points better than the final polls indicated, while challengers did about two points better -- indeed a ratio of more than 6 to 1. This result mostly had nothing to do with undecided voters. Nine of the 28 surveys were from 1996, a year in which the pre-election polls overstated incumbent Bill Clinton's winning margin by over 4 points on average. Five of those surveys reported no undecided voters whatsoever -- and their errors were not systematically different from the other four surveys'. (The second most accurate survey that year reported 11% undecided, equalling the other eight surveys combined.) Clearly Bowers's analysis cannot be used as a guide to allocate reported undecided voters. At most it might imply that challengers tend to do well when some polls report a lot of undecideds, but I see little support for even this inference.
I examined a total of 31 presidential polls going back to 1972, using the same National Council on Public Polls dataset used by Bowers. (1972 was the earliest year with multiple polls, at least one of which reported undecided voters.) In some elections, the polls with higher reported undecided rates indicated less net support for the incumbent than the other polls; in others, they indicated greater incumbent support. Overall, the results indicate no statistically significant relationship between percentage undecided and incumbent support. Thus, there is no statistical support for altering the margin to account for reported undecideds. Another way to say that is that reported undecideds seem to break about evenly (since allocating them doesn't appreciably alter the margin), perhaps slightly toward the challenger on average. But every election and every poll is different, so it is probably best just to say that we don't know.
[1/6/07: TIA notes that he has "cited a 155- election survey in which challengers won the undecided vote in 127 and incumbents won the undecided vote in just 19." The study, by Nick Panagakis, is summarized here. As Panagakis writes, the polls "are predominantly from statewide races, with a few U.S. House, mayoral and countywide contests thrown in. Most are from the 1986 and 1988 elections, although a few stretch back to the 1970s." As far as I can tell, no presidential polls are included.
[TIA also reiterates that "world-class pollsters Zogby and Harris... have stated that challengers typically win over 2/3 of the undecided vote." Conceivably in their polls that is true. (Incidentally, TIA attributes to "Lou Harris" an unattributed quotation in a 2004 Harris Poll report, although Lou Harris left Lou Harris & Associates in 1992.) I have here presented empirical evidence that, among all available presidential polls going back to 1972, the undecided vote did not break 2/3 to the challenger. TIA often ignores discrepant evidence, but not always so obviously.]
On average, it is true that incumbents don't do better -- or, rather, much better -- than their predicted shares in the final polls. As noted above, in the polls examined by Bowers (and by me), on average, incumbents did about 0.3 points better than the final polls indicated.
Averages can be misleading, and this one is. Remember the old joke about the hapless target shooter who grazes a tree with one shot, kills a bystander with the next, but brags that on average he nailed the bullseye? Well, in two of the five elections in the Bowers analysis (the five being 1996, 1992, 1984, 1980, and 1976), the incumbent improved his vote share by about 2 points over the polls. In 1984, Ronald Reagan averaged 57% over six polls and got 58.8% in the official returns. In 1976, Gerald Ford averaged 45.7% over three polls (although that average is influenced by one poll that is much lower than the others) and got 48.0% in the official returns.
These elections are partly offset by 1996, not only because Clinton actually earned a lower vote share than predicted in the polls (50.3% in the polls, 49.2% in the vote), but because there were more polls in that year. So, on average, the incumbent poll figures are close to the official percentages.
George W. Bush's 2004 share across the 14 trial heats in the pollingreport.com roundup was about 48.5%; his official vote share was 50.7%. Thus, an improvement of about two points, as in 1976 and 1984 -- hardly a wild exception to a well-established rule.
[1/6/07: TIA complains that the two cases in which the incumbent improved his vote share are unrepresentative, because "Reagan won a landslide" and there were so few polls for Ford. There is nothing to prevent TIA from arguing that his rule is true except when it isn't, in which case(s) it doesn't count -- but then I don't see how it is much of a rule. He also claims that I assume that "Bush's TRUE vote was 50.7%," although I did not assume any such thing. I noted that Bush's official vote share was about two points larger than his average poll share in this dataset.]
TIA has stated that Bush's approval rating on November 1 was 48.5% based on the "average of 11 polls." I'm not sure which polls are included in that figure; different surveys typically yield slightly higher or lower approval figures. During October 2004, the Gallup Poll reported six approval figures for Bush: two above 50%, three below 50%, and one exactly 50%. The last pre-election survey put him at 48% approval, 47% disapproval, and 5% "not sure." (Gallup's two November polls gave Bush 52% and 55% approval.) So, Bush's final Gallup approval rating was below 50%, although not very far below.
So, is a final Gallup approval rating (slightly) below 50% a sure sign of imminent defeat? Starting in 1972, TIA cites three incumbents with approval ratings below 50, and two with approval ratings in the 50s. In order of approval rating (as cited by TIA), they are:
| Year | Incumbent | Approval rating | Vote share |
|---|---|---|---|
| 1992 | Bush (I) | 46.5% | |
| 1980 | Carter | 44.7% | |
| 1976 | Ford | 48.9% | |
| 1996 | Clinton | 54.7% | |
| 1972 | Nixon | 61.8% |
(By the way, TIA's 1992 figure appears to be from late September. A Gallup/Newsweek poll in late October gave Bush a 37% approval rating, which does not much alter the analysis. Similarly, TIA's 1980 figure for Carter appears to be from after the election, not before it; a mid-September poll showed Carter at 38% approval (which may well have fallen by Election Day, although we have no way of knowing how far). I do not know why TIA excluded 1984; a Gallup poll in late October showed Reagan at 58% approval, and he won 59.2% of the vote. I've used what I believe to be the latest pre-election Gallup figures in the chart below.)
Notice that prior to 2004, there is no incumbent with an approval rating between 46 and 55. So, one might say it is a rule that "incumbents with approval ratings under 50 never win." But one might just as accurately say that "incumbents with approval ratings over 46 always win." Does one of these statements make more sense than the other? For what it is worth, let's examine a plot of final approval ratings vis-a-vis incumbent two-party vote shares from 1956 on.
In the plot above, the dashed blue line is the best-fit line for all these elections. The solid purple line is the best-fit line excluding the 1956 and 1964 elections, which appear to be out of trend. (The 1956 and 1964 approval ratings are from summer polls -- as are the 1972 and 1976 ratings.) Either way, it appears that incumbents with approval ratings over 45 are likely to win! The thinner, vertical purple line depicts this break-even point. (Notice that Carter in 1980 and Bush in 1992 each were in the mid-40s of the two-party vote despite approval ratings in the 30s.) Note also that 2004 lies below the best-fit lines: if anything, Bush might have been expected to do slightly better in the election based on his approval rating. With so few data points, it would be fairest to say that the election could have gone either way.
[12/14/06: Parts of this discussion used TIA's numbers, while parts used the best numbers I could find; I've tried to clarify which is which. It seems to me that TIA's numbers are actually less friendly to his case, but it doesn't matter much.]
[1/6/07: TIA says, "You can have your own theories, but not your own facts. And the FACT is that ALL incumbents with final approval ratings under 50% LOST the election – and that includes Bush. Your attempt to twist this FACT is patently obvious. Your scatter plot is just a diversion. Why don't you just accept it and move on?" This response is not well reasoned. As I noted, it is equally a FACT that ALL incumbents with final approval ratings over 46% WON the election (and Ford came pretty close). TIA regards his fact as true and my fact as twisted, but on what basis? Either one of these "facts" omits most of the information in the table. The scatterplot presents all the information. I have to wonder: why does TIA seem so afraid of this scatterplot? Perhaps because no one who looks at it is likely to agree that obviously 2004 should fall below the horizontal line that separates a Bush win from a Kerry win. In short, because it sinks his case. Sorry about that.]
Previous: (1) The Pre-Election Polls
Next: (3) Describing the Exit Poll Discrepancies
Up to Table of Contents