The Undecideds, Take Two

We are on the homestretch and I will be posting shorter posts quite frequently leading up to Tuesday’s night Election Eve coverage, which I will try to blog while doing double duty with election night coverage at the Grille here at the College.   In this post I want to revisit an issue I discussed in an earlier blog, and during my talk yesterday at the College: who are the undecideds, and for which candidate will they vote?

In the earlier post, I drew on data presented by a Pew poll to show that the undecideds are slightly more white, less affluent, less educated and less politically aware than voters as a whole, and therefore were more likely to break for McCain than for Obama.  I also suggested, however, that with only 6-8% of voters still undecided, that they would all need to break for McCain to make a difference in this race.  However, using data from a study discussed by Mark Blumenthal at, I estimated that the undecideds might only break for McCain by a slight margin, say 53-47, which would net him only about 1% in the overall vote.  That might be enough to push him over the top in some tossup states, but probably not enough to change the outcome of the race.

However, using tracking poll numbers provided by my crack research assistant Kaitlynn Saldahana, I decided to take a second look at the estimates of the number of undecideds to see if they jibe with Pew’s numbers.  Remember, each pollster has a different policy for dealing with undecideds – some “push” them to make a choice, thus folding more uncertain voters into the overall column for a particular candidate.  Others allow the undecideds to stay undecided.  In looking at the data from the various tracking polls, we find that estimates of the number of undecideds range from a low of 2% in Rasmussen and Zogby to 9% in the IBD poll.  Here’s the latest data on the race, including undecideds, for each tracking poll, (first number McCain, second Obama, and third is the undecideds. Note that because Rasmussen doesn’t list undecideds, I categorize the unaccounted vote as undecided):

Rasmussen       47        50        2

Zogby              43        50        2

Gallup              45        50        4

Hotline             42        48        6

Battleground     45        49        6

IBD/TIPP:        43.9     46.9     9.2

Additionally, here’s data from the major news surveys which come out more intermittently (same format – McCain, Obama, Undecideds)

ABC                42        51        6

FOX                44        47        6

CBS                 41        52        5

What does this suggest?   The average for pollsters showing 5-6% undecideds is McCain 42.8%, and Obama 49.4.  The average for pollsters with 2-4% undecided is McCain 45% and Obama 50%.  What this suggests is that when pollsters push the undecideds to make a choice, they appear to break for McCain by a roughly 5 to 1.  This is a very crude estimate, of course, based on a small number of polls with a lot of uncertainty, but it is entirely consistent with my read of the demographics of these voters.  So, let’s assume there are still about 3% undecideds out there, and they do break for McCain at 5 to 1.  This gives McCain approximately 2.4% while Obama gains maybe a half percent. This will make the final tally closer to  51.5 Obama to 47.5 McCain, with about 1% going to other candidates.   This is all back of the envelope calculations, of course, and is subject to revision as more polling data comes in.  Most importantly, it doesn’t tell us much about the key battleground states.  But, if my reasoning is correct, it suggests that we should see the race tighten at the national level in the next few days, if the undecideds become decideds.  However, on Election Day last year, 3% of those surveyed still claimed to be undecided.  If that holds true this year, we won’t see the McCain gain until the votes are counted.



  1. -What will the overall popular vote breakdown look like if we follow the (unlikely) plan that McCain wins the toss-up states and flips a few Obama states to go ahead or break even in electoral votes? i.e what is the minimum increase in popular vote McCain needs to go ahead in the Electoral College?

  2. I’d be very surprised if last-minute deciders broke so decisively for McCain. Gopoian and a co-author found that late deciders allocate their votes almost randomly at the end (Political Behavior 16:1, 1994), which is more what you’d expect for less attentive, less informed voters. Looking at past exit polls, the biggest advantage among late deciders in a national presidential campaign recently has been John Kerry’s 7 percentage point advantage in 2004. Otherwise, the two candidates have roughly tied among this group.

  3. My colleague Bert Johnson raises a good point regarding previous research in this area and provides a nice hypothesis for us to test in this election cycle. The article he cites examines late deciders (those making up their minds in the last 2 weeks) in the five presidential elections from 1972-1988. The authors conclude that late deciders are a unique subset of voters – lower information, lower interest in politics – and that their vote cannot be predicted based on the usual voting models that apply to earlier voters. On this much Bert and I agree. Where we disagree, I think, is in the implications of the findings. I read Gopoian to conclude that there is no systematic explanation for how late deciders vote; each election provides its own unique context that determines how undeciders break that particular year. In other words, one can NOT use prior elections to predict how undecideds will vote in the current election. This is particularly the case this year, I think, because for the first time since 1952 there is no “establishment” candidate (no incumbent or vice president) running. Note that in the elections Gopoian studies that is not the case. Given this, I think one needs to look only at the current electoral context to determine how the late deciders break. And as I suggested in my earlier post, I see the polling data so far indicating that they will break in greater numbers for McCain, given their demographics. But this will be one thing Bert and I will certainly discuss on Election Night at the Grille!

  4. There is also a good article about undecided voters in the Washington Post today:

    The gist of the story is that the McCain camp is making an argument very similar to the one Dickinson has made above: the demographics of the undecided voters seem to suggest that they will be more likely to vote for McCain. Nevertheless, the authors, Robert Barnes and Jon Cohen, also point out that other pieces of information complicate the picture:

    “Among those voters who say they are still undecided or open to persuasion, more are currently McCain supporters than Obama backers. So McCain has to fight to hang on to those voters as well as persuade those leaning toward the Democrat. And Obama’s supporters seem more enthusiastic. Among all likely voters in the Post-ABC poll, 49 percent say they will “definitely” vote for Obama heading into the final weekend, compared with 40 percent who say the same about McCain.”

    It may be true that undecided voters are more likely to vote for McCain, but the very fact that more of his supporters are “undecided” complicates the issue. The truth is probably not as simple as the McCain camp suggests: while the demographics of undecided voters suggest that they are more likely to vote for McCain, there is no guarantee that they actually will. It is Obama who has the larger number of strong supporters; the question is whether McCain can close the gap on election day.

    To be perfectly frank, I think it is hard to trust any comments about the polls coming from either candidate with only a few days left in the race. McCain’s camp could easily be seeing this trend because they want to (or even because they have to). Obama’s campaign statements on the matter could also be equally biased because they also know that perceptions can easily become a reality on the campaign trail. Alternatively, Obama could use closer poll numbers to keep his base as mobilized as possible. The good news is that we’ll know the accuracy of the polls in a few days.

    Perhaps it is a question better left for November 5, but I think it’s worth considering before the results come in:

    One of the arguments against giving Obama the democratic nomination was that he won a lot of delegates in states that “wouldn’t matter” in the general election. While external factors (such as the economy) have certainly played a role in shaping the map, hasn’t Obama’s broad electoral strategy already been vindicated? Four days before the election, he is competing in traditional red states like Montana, North Dakota, North Carolina, and Indiana and holds significant margins in Nevada, New Mexico, Colorado, Virginia, and New Hampshire. Add to that Obama’s influence on down-ticket races in the South, and you have a pretty decent justification for what he has done.

    I wonder what the race would look like now if Clinton were the candidate. My guess is that it would be significantly closer than it is right now.

  5. Bert has a reply that somehow was deleted before I could approve it. Here it is:

    “I’d believe Matt’s argument here if, in the article cited, the authors had attempted to predict late deciders’ votes solely on the basis of factors that occur in multiple elections — party ID, economic evaluations, and other issues. But the authors run separate models for each election and incorporate campaign-specific factors such as each voter’s response to questions like (for 1972) “Can Nixon be trusted?”, “Can McGovern be trusted?”, (for 1976) does Carter have a “presidential personality?” and (for 1988) how would you characterize Dukakis’s leadership qualities?

    The fact is, none of these campaign-specific factors did a better job of predicting late deciders’ votes than the “fundamentals” (economy, party ID, etc.) or policy positions on foreign affairs, race relations, or social issues. Nothing explained late deciders’ choices very well. This group of voters, for example, is much more likely to cross party lines than other voters.

    I buy the argument that there are some late deciders who are disgruntled Republicans who are less enthusiastic about McCain than Obama voters are about Obama. Interestingly, though, the Gopoian piece suggests that these people cannot be expected to support McCain based on party ID alone (see p. 66). I wouldn’t be surprised if McCain had a slight edge among those who make up their minds at the last minute, but I don’t expect the margin to be very large.”

  6. Bert and I agree that as a voting bloc, “undecideds'” voting decisions are not driven to the same degree by the factors that explain early deciders’ votes. Because the previous study by Gopoian found no statistically-significant campaign-based explanations for how late deciders’ vote, however, Bert concludes that their choice is simply random, that is, that late deciding voters in effect flip a coin when deciding for whom to vote. In contrast, I believe that undecideds are moved systematically, but by campaign-specific factors that are unique to each election. The fact that Gopoian did not identify these factors in earlier elections does not mean they don’t exist in this one (or for that matter that they didn’t exist in earlier elections.) Note that in theory we can test our respective claims in a couple possible ways. We might look at Tuesday’s exit polls which usually identify who made their choice in the last 2 weeks and see if one candidate received a significantly different level of support from this group than did the other. Less reliably (since we can’t actually document individual movement), we might track each daily tracking poll during the last two weeks to see which candidate gains votes as the number of undecideds declines (keeping in mind that we might actually be observing early deciders changing their mind). Again, I expect that Bert and I will revisit this discussion in Tuesday, but I invite others to join in. How do you think the undecideds will break, and why?

  7. So in looking at the various breakdowns by groups that have been polled, the only categories that McCain is drawing above 55% are identified Republicans, and weekly white Church-goers. If we’re to think that undecideds (which are far more demographically mixed) are going to break significantly for McCain, why would they exceed his performance in every other demographic group? I certainly buy that McCain’s likely to draw a majority of undecideds, but numbers like 5 to 1 seem completely off the map compared to other polling indicators.

  8. OK, being the nerd that I am, I spent some time on a Saturday trying to roughly replicate the Gopoian study for the 2004 election, using data from the National Election Studies. Here’s what I’ve found. I ran four different logistic regression models. In the first two, I aimed to predict a Bush vote on the basis of a series of demographic and issue characteristics: liberalism, income, age, gender, party ID, and attitudes on the economy, the Bush tax cut, and the Iraq war. In one logit model I focused on the vote choice of early deciders, and in the second, I focused on those that made their decision in “the last few weeks” or sooner (restricting the “late deciders” to the last few days results in too few cases for analysis). Using these variables, I come up with results very similar to Gopoian’s: for early deciders, their votes are very predictable. Partisanship and issue attitudes on taxes and Iraq come out as very significant, and 94 percent of cases are correctly categorized.

    In the second model, focusing on late deciders, the only variable that comes out as significant is partisanship, and only 77 percent of cases are correctly categorized. This suggests that there may be some effect involving people “coming home” to their natural party, but otherwise, the votes of late deciders are unpredictable.

    In the second two models, I included the variables listed above, plus two “feeling thermometer” variables that reflect voters’ general affect towards the two candidates on a 0 to 100 “hot” or “cold” scale. My theory was that these variables might pick up any “miscellaneous” campaign effects that could exist. When these variables are included, we do much better at predicting the votes of late deciders. The feeling thermometer variables become very significant, partisanship drops off in importance, and the model correctly predicts 89 percent of late deciders’ vote choices.

    So does this mean campaign effects do exist for the last-minute undecideds? Possibly, but there’s another explanation for the feeling thermometers’ significance: They may reflect post-hoc rationalizations of late deciders’ votes. If you’ve voted for a candidate, you may end up deciding that the candidate is an all right guy. I’m still more comfortable with predicting a mild effect based on previous party ID, but not much else.

Leave a Reply

Your email address will not be published. Required fields are marked *