« The Newsweek Poll | Main | New Format & URL »

October 03, 2004

The Incumbent Rule

In discussing recent poll results, I said last week that I tend to focus most on President Bush's job rating and percentage of the vote (which tend to track closely with each other), adding, "that both are hovering just at or above 50 suggest an ultimately close contest, with Bush receiving just about the support he needs to win." Many of you asked me to elaborate, and I had been meaning to write up an explanation of what pollsters often refer to as the "Incumbent Rule," that explains why the undecided vote often breaks toward challengers in races featuring an incumbent.

Then, just as I started writing this up in earnest, another Democratic pollster named Guy Molyneux, a partner at Peter D. Hart Research Associates, beat me to the punch. His excellent article in the American Prospect, "The Big Five-Oh," which appeared online on Friday, makes more or less the same argument I had been promising. It is worth reading in full, but let me first give you the gist and offer some additional supporting evidence.

The basic idea is that voters make their decisions differently in races involving an incumbent. When newcomers vie to fill an open office, voters tend to compare and contrast the candidates' qualifications, issues positions and personal characteristics in a relatively straightforward way. Elections featuring an incumbent, on the other hand, are as Molyneux puts it, "fundamentally a referendum on the incumbent." Voters will first grapple with the record of the incumbent. Only if they decide to "fire" the incumbent do they begin to evaluate whether the challenger is an acceptable alternative.

Voters typically know incumbents well and have strong opinions about their performance. Challengers are less familiar and invariably fall short on straightforward comparisons of experience and (in the presidential arena) command of foreign policy. Some voters find themselves conflicted -- dissatisfied with the incumbent yet also wary of the challenger -- and may carry that uncertainty through the final days of the campaign and sometimes right into the voting booth. Among the perpetually conflicted, the attitudes about the incumbent are usually more predictive of these conflicted voters' final decision than their lingering doubts about the challenger. Thus, in the campaign's last hours, we tend to see "undecided" voters "break" for the challenger.

That's the theory. Does it have any empirical support?

· In 1989, Nick Panagakis, president of Market Shares Corporation (the firm that polls for the Chicago Tribune) analyzed results from 155 surveys, most from the late 1980s, all conducted during the last week before an election. In a famous article in The Polling Report, Panagakis found that in 82% of the cases, the undecideds "broke" mostly to the challenger.

His conclusion? "Incumbent races should not be characterized in terms of point spread. [Suppose] a poll shows one candidate leading 50% to 40%, with 10% undecided…Since most of the 10 points in the undecided category are likely to go to the challenger, polls are a lot closer than they look – 50% to 40% is likely to become 52% to 48%, on election day" (emphasis added).

· Just last month, Chris Bowers of MyDD updated Panagakis' work. Though he found some signs that the incumbent rule might be weakening in state and local races, he found even stronger support for it in presidential elections. In 28 surveys involving presidential elections, 86% showed undecideds breaking mostly to the incumbentchallenger.

· Finally, Guy Molyneux averaged "the final surveys conducted by the three major networks and their partners" in the last four presidential elections featuring an incumbent (see the table here), and found that,

In three of these the incumbent fell short of or merely matched his final poll number, while exceeding it only once, and then by just a single point (Ronald Reagan). On average, the incumbent comes in half a point below his final poll result…In every case, the challenger(s) -- I include Ross Perot in 1992 and 1996 -- exceed their final poll result by at least 2 points, and the average gain is 4 points. In 1980, Ronald Reagan received 51 percent, fully 6 percentage points above his final poll results. (emphasis added)

I can confirm Molynuex's numbers and suggest some additional evidence.

· Clinton's support in 1996 was even closer to the final result if the average also includes the other final national polls done in 1996: Gallup, Pew, Harris, Hotline/Battleground and Zogby (using their last report that included results for undecided). Clinton's average on the eight polls was 49% -- exactly what he got on Election Day.

· Finally, we can examine Gallup's near 60-year database of pre-election polls. Unfortunately, on their last poll, Gallup always releases a projection that allocates the undecided to the candidates (and I have not been able to find documentation on how they apportion the undecided). Until 1996, Gallup never reported their final results with undecideds included. Also, prior to the 1980s, the next-to-last Gallup survey each year typically fell as much as a month before the election, when its predictive value was weaker.

Incumbent Final Gallup Projection Actual Result Differ- ence
1956 Eisenhower 60%59.5% 57.8% 1.7%
1964 Johnson 64% 61.3% 2.7%
1972 Nixon 62% 61.8% 0.2%
1976 CarterFord 49% 48.1% 0.9%
1980 Carter 44% 41.0% 3.0%
1984 Reagan 59% 59.2% -0.2%
1992 Bush 37% 37.7% -0.7%
1996 Clinton 52% 49.2% 2.8%
Average 1.3%
Source: Gallup

However, the final Gallup projections (sans undecided) show an intriguing pattern: In the presidential elections since 1956 that featured an incumbent, Gallup's final projection of the incumbent's vote exceeded the incumbent's actual vote six of eight times. The only exceptions were Ronald Reagan in 1984 and George H.W. Bush in 1992, and then by only 0.2% and 0.7% respectively. On average, Gallup's projection of the incumbent's vote has averaged 1.3 percentage points greater than the actual result. Obviously, without seeing the raw results we can only speculate, but this pattern suggests that Gallup has allocated too many of the undecided over the years to incumbents.

Now, consider the incumbent rule in relation to the recent polls. Bush has received 49-50% of the vote on the RealClearPolitics' rolling seven-day average since the Republican convention. Similarly, as Molyneux also notes, Bush's percentage of the vote in key battleground states like Florida, Ohio, Missouri, Iowa, New Hampsire, Nevada and West Virginia now ranges between 48-49% in polls done just before the first debate. In Wisconsin and Colorado, Bush does a bit better (50-51%); in Pennsylvania, Michigan, Minnesota, Maine and Oregon he runs a few points worse (45-47%).

None of this implies that the current standings will determine the final outcome. Where the race ends up a month from now could obviously be different. However, the incumbent rule tells us that, at any given moment, the President's percentage of the vote relative to 50% is a better indicator of where the race stands than the margin separating Bush and Kerry. It also suggests the appropriate way to read the final polls just before the election (and these are my ranges – others may differ): If the average result of all the final polls (including undecided) puts Bush's percentage at 50% or higher, the President will likely win. If Bush's percentage is 48%-49%, the race is headed for a photo finish. At 47% or lower, the President will likely lose (add 1% to these ranges in any state where Ralph Nader is not on the ballot)

The main point: The incumbent's level of support is more important than the margin.

[Continue with More on the Incumbent Rule]

Note: Interpretation of political survey data is even more subjective than methodology. Others may disagree, and opposing viewpoints are always welcome in the comment section.

Correction: In the comments, Dwight Meredith correctly noted an apparent error in the values for Eisenhower in the table above, since 60%-57.8% certainly does not equal 1.7%. The correct Gallup estimate for Eisenhower in 1956 was 59.5% (Gallup reported their results to once decimal that year only), so the 1.7% difference was correct. My spreadsheet displayed the rounded value (60%) even though it used 59.5% for the calculation. I've corrected the table.

Related Entries - Incumbent Rule, Interpreting Polls, The 2004 Race

Posted by Mark Blumenthal on October 3, 2004 at 11:34 PM in Incumbent Rule, Interpreting Polls, The 2004 Race | Permalink


Very nice piece. Well done. Thank you.

Posted by: Rory | Oct 4, 2004 1:43:04 AM

Thanks for the report.

However, Pat Caddell has been saying over and over that, in presidential elections, undecideds usually break for the INCUMBENT, not the challenger. He said 1980 is among the few exceptions to this - largely because of the debate, which was held in the last week of the campaign.

In 2002, many low polling sub-50% incumbents, such as Wayne Allard, Tim Johnson, Mary Landrieux Gray Davis, etc., etc., etc., were re-elected.

I think the "old rules" may go out the window Post-911.

Posted by: Dr. X | Oct 4, 2004 3:16:24 AM

Thanks for the explanation. I have a question though: polls suggest that while Bush’s support is quite solid, about half of Kerry supporters are voting against Bush, not for Kerry. Based on my own personal anecdotal evidence, I might suspect that a sizeable portion of those later Kerry supporters may decline to vote and a smaller number may change their minds near election day. [Two of my acquaintances recently did so, though that is not sole what my supposition is based on.] That would cause Kerry's numbers to drop below his poll numbers in the election, possibly by as much as 4-5%, if say 10% of his supporters failed to vote or switched sides [relative to Bush voters].

I can’t remember another election where such a large part of a candidates support was of this negative character. How do pollsters address this issue, if they do so at all?

Posted by: C. Owen Johnson | Oct 4, 2004 6:30:58 AM

Nice post. Please check your math for the Eisenhower reelection in the table. 60% - 57.8% equals 2.2% not 1.7%. I think that brings the average difference to 1.3625%.

Posted by: dwight meredith | Oct 4, 2004 8:36:44 AM

C. Owen Johnson makes the comment: "Kerry supporters may decline to vote and a smaller number may change their minds near election day." I disagree with Mr. Johnson. I suspect that those voting for Kerry because they are "ABB" (anybody but Bush) are more likely to cast a vote than those who merely favor Karry over Bush. There is always more emotion involved in being radically against someone than being mildly for his/her opponent.

Posted by: Glynn Harper | Oct 4, 2004 10:15:58 AM

Excellent work. This has fast become *the* must-read blog for anyone interested in interpreting political polls.

Posted by: Tom Miller | Oct 4, 2004 10:47:48 AM

I offer a different view.


We are a month away from the election. Those who are undecided now are not more likely to break for the challenger. At least, that is what the history of Presidential elections says (lower ticket races may be different).

Perhaps those undecided the day or two before the election are, but we are not there yet.

Posted by: Gerry | Oct 4, 2004 11:04:02 AM

One note to add to your excellent piece. The percentages I reported in my study were not the number of times the majority of undecideds broke for the incumbent or the challenger. Instead, the numbers I produced were based on the total number of undecideds in all of the polls and time periods that I studied. That is, 86% of the total nunber of undecideds in all 28 Presidential polls I examined broke for the challenger, while only 14% broke for the incumbent.

As good news as that may seem for a challenger, it actually only accounts for about a two point gain. This is because the number of undecideds in final Presidential election polls is much lower than in other elections. Bush needs to be at least at 47% to have any chance on Election Day, but Kerry probably needs to be at least at 46%.

Posted by: Chris Bowers | Oct 4, 2004 11:11:43 AM

I would argue that this factor may be significantly less pronounced in this election because incumbents usually run on their records, and President Bush has instead run on a platform of saying that Kerry is an unacceptable choice -- he's made the election a referendum on Kerry more than on himself. I therefore think undecideds will break more evenly this time than in the recent past.

Posted by: Ryan Booth | Oct 4, 2004 11:17:18 AM

I just went and checked some of the data I have.

In 1996, the last Rutgers poll gave the incumbent, Clinton, an 11 point lead over his main challenger, Dole. He won the state by 17.

For New Jersey, in 1992 the last Rutgers poll gave the challenger, Clinton, a 12 point lead over the incumbent, Bush. Clinton carried the state by 2 points.

In 1984, the Rutgers poll gave the incumbent, Reagan, a 16 point lead over the challenger, Mondale. Reagan won by 21 points.

In 1980, Reagan proved the exception to the trend showing that in NJ the incumbent gets the late breakers, as the last Rutgers poll had it tied at 36. Carter got 2% more than that, but Reagan won by 13.

In 1976, Rutgers final poll had Ford, the incumbent, by one over Carter, the challenger. Ford carried the state by 2.

It is safe to say that the incumbent rule does not apply to the Rutgers poll for New Jersey Presidential races.

I was only able to find the 1992 Columbus Dispatch poll for Ohio. It had Clinton up 1 1/2 points over Bush. The final margin was 1.83. So there was no challenger surge there. While I do not have the numbers for 1996, their last survey's writeup said that Dole remained within striking distance but Clinton was poised to win the state by the largest margin for a Democrat in 32 years. That means it had the margin somewhere bigger than 2 points but still 'within striking distance'. That sounds like they had it pretty much on target, as Clinton won by 7.

I sure would love to see some data backing up the idea that various state polls have shown that the challenger makes up ground on the incumbent in Presidential races. While I have only a handful of data points (in two states) they do say the opposite.

Posted by: Gerry | Oct 4, 2004 11:58:25 AM

I left off a datapoint. In 1972, Rutgers final survey had the incumbent, Nixon, winning by 23 points over the challenger, McGovern. The final result in New Jersey was Nixon by 25 points.

Posted by: Gerry | Oct 4, 2004 12:01:19 PM

Last election, seems to me Gore represented the incumbency, yet 80% of undecideds borke for him. So, the "exceptions" are in '84, '92 and 2000. Getting hard to tell which is the rule and which is the exception.

Posted by: Larry | Oct 4, 2004 12:49:55 PM

Inre: Gore. He was not the incumbent. He may have been considered the closest thing to one when pundits attempted to round-hole this square peg, but he was never the incumbent except when he ran for re-election in Tennessee.

This same incumbency-rule application was unsuccessfully applied to the 2000 NY Senate race where HRC was given the incumbenet moniker. The undecideds broke markedly to her.

Posted by: Eric Hinz | Oct 4, 2004 1:21:10 PM

re Gore:

I believe undecideds broke for Gore because of the last minute DUI revelation.

Posted by: Dr. X | Oct 4, 2004 2:05:23 PM

I notice that the last three Republican incumbents (Nixon, Reagan, Bush) polled at or above their % in final poll avg. while Clinton and Carter fell short of forecast. I have found that the generic congressional preference polling avg. have underestimated GOP share of house vote each year 1998-2002. Has anyone tried to isolate a potential systemic underestimation in GOP voters in polls? I realize this would be small, on the order, absent other factors, of 1-1 1/2%.

Posted by: khurst | Oct 4, 2004 3:00:02 PM

When comparing these polls over 50 years, isn't there a timeliness factor? I remember on election eve in 1980 the networks were calling it $0% Carter, 40% Reagan, 10% Anderson and 10% undecided. Reason they missed so bad was that their polling stopped 10 days before the election....it took that long to assemble the data. Only a truly wealthy polling effort (like Pat Caddell) could tell Carter the night before that he was gonna get whipped.

Now, computing power lets relatively modest outfits do overnight polling by state, district, etc. So it seems that there is a less of a swing towards election day.


Posted by: michael | Oct 4, 2004 3:01:28 PM

This election is pretty unique, in my opinion, given the Iraq war and Kerry's back-and-forth positions on it. The interesting thing about "the Undecideds" is that they don't like Bush, but they don't seem to be interested in voting for Kerry either. I can not imaging Kerry getting a majority of the undecideds given his "flip-flop" reputation on Iraq - the most important issue to many of them. His debate performance was good, but there's simply no way for him to shake his wishy-washy Iraq reputation. I think the real surprise in the election will be the Nader vote - even given the fact that he's not on a number of state ballots. Roughly 25% of Nader's votes in 2000 came from self-described Republicans. Many of the undecideds are Republicans who just can't hold there noses and vote for Kerry. They will either stay home or vote for Nader or some other third party. In a way, the '92 Perot suprise (he got 19% when he was polling at only 9%) is a powerful example of people who did not like either the incumbent or the establishment challenger.

Posted by: AndrewNJ | Oct 4, 2004 5:15:47 PM

In your table, you have Carter as the incumbent in 1976. The line should read Ford.

Posted by: Elliot | Oct 4, 2004 6:56:53 PM

In the behavioral sciences, the past is the best predictor of the future. I am more persuaded by what voters have actually done in past elections than speculations why this time will be different (so I'll be right every time until I'm wrong, then). MP's arguments seem very persuasive to me.

Several wild cards do occur to me, however.

Will Nader also pick up undecideds, moving him from a 1+% range to a 2+% range?

Are there states or regions which are consistently less predictable in their incumbent/challenger break? NH, NM, and NJ have all been counterintuitive at times. Is this just the normal amount of outliers?

Are there oft-reported polls which are known by professionals to be less useful?

Posted by: Assistant Village Idiot | Oct 4, 2004 7:44:11 PM

Will Nader also pick up undecideds, moving him from a 1+% range to a 2+% range?

Doubtful. Nader polled in the high 3's, almost 4% in the final 2000 polls, and ended up around 2.5. Since he's not uniformly on ballots across the country this time around (swing states included), his national number should be quite irrelevant, and it would be shocking if it ended up much more than 1%.

Posted by: Steve F | Oct 4, 2004 8:14:51 PM

Nader's vote total throughout the whole US will undoubtedly be hampered by ballot access issues. For example, he's not on the California Ballot. It's a pretty interesting fact that in 2000 he only received half as many votes as polling anticipated but this election is very different than 2000. There is one factor that will really help Nader this time. He is the only anti-Iraq-war candidate in the race. A lot of people (Dems and some Repubs) who are fervently against the war completely will vote for Nader since they don't feel right for voting for either Bush or Kerry. Remember, Perot's anti-deficit position gave him the moral high ground - which is what many people vote for. What did Nader represent in 2000 - a hodgepodge of issues. This time, Nader is a vote against the War, clear and simple. As Kerry moves to the center on the Iraq issue many on the Howard Dean far-left will defect to Nader (regardless of Dean's support of Kerry). In states where Nader is on the Ballot he will outperform his polling as many people who are marginally pro-Kerry will vote there conscience - for Nader.

Posted by: AndrewNJ | Oct 4, 2004 9:22:53 PM

Andrew: I think Nader's performance will depend on the perception of the state of the race on November 1 and 2. If it's generally perceived that Bush will win by a comfortable margin, some significant number of left wing Dems (and by "significant" I mean "less than 5%") may decide to vote their consciences and go for Nader. But if the election is perceived to be 2000 redux, I would guess that Nader will get very few of their votes, as they will decide to believe the anti-war iteration of Kerry.

Posted by: Brandon | Oct 5, 2004 8:05:01 AM

Interesting discussion. The conclusion seems to be that Bush may be in trouble now and that he'll certainly be in trouble if the final round of polls show him tied or with less than a 2 point lead.

Could Mystery Poller further expand on two topics of some interest?

First, what effect does the increased frequency of polling have on his thesis. Specifically it was pointed out that the last public poll on the 1980 election ended 10 days before election day. Persumably the oldest data in the last pre-election polls these days is considerably fresher and therefore we might well expect the latest poll to be more accurate than in 1980 and previous. The results shown in your Gallup table don't seem to support that conclusion very well.

Second, is there a party effect involved. Judging by the very limited data in that table, the Republicans seem to suffer less than the Democrats from soft support. Anything to that idea?

Posted by: Don | Oct 5, 2004 10:42:35 AM

Are the undecides breaking to the challenger, or are the pollster's faulty weighting and likely screens causing the errors to creep into the polls? 1996 and 2002 seem like good examples of this. The discussions above are good examples why polls are more art than science.

Posted by: Cableguy | Oct 5, 2004 2:20:42 PM

Is the undecided theory equally valid for individual states?

Posted by: Mark | Oct 5, 2004 5:31:27 PM

The comments to this entry are closed.