How Undecideds Split, 1948-2004

11/03/2008 03:30 am ET | Updated May 25, 2011

Another look at how undecided voters have ultimately voted.  There is much speculation about this, and I've offered a bit of empirical evidence.  Here is a historical look.

The National Election Study (NES) is the leading academic study of electoral behavior, originally developed at the University of Michigan but now managed by a broad board from many universities. 

The NES got it's start in 1948 when, by a lucky break, the Michigan scholars conducted a foreign policy survey in the fall during the campaign. The survey was not directed to the election, but did include a vote choice item. And then Truman won, and the Michigan survey got it right. So they went back and reinterviewed everyone from the pre-election survey to try to throw light on how the vote came to and what it could help explain about the other polls that got it wrong.

In 1952, the Michigan group developed a new, specifically election oriented, survey. Once more they interviewed before the election and then reinterviewed the same respondents after the election.

Landmark books followed, most prominently The American Voter (1960) by Angus Campbell, Philip E. Converse, Warren E. Miller and Donald E. Stokes. (May I add that it was my honor and pleasure to serve as a research assistant to each of Campbell, Converse and Miller during my graduate career, though a good while after The American Voter!)  And for our purposes, a series of National Election Studies was born, which in every presidential election since 1948 has conducted a pre-post survey, allowing us to glimpse how votes shift from Sept-Oct intentions to November action. The NES is now supported by the National Science Foundation as a public resource for the study of elections. (Google for NES for details.)

So, thanks to these pioneers in electoral research, we can see how undecided voters have divided over the years.  The one caveat is that in most years the sample size of the undecided is modest, so the sampling error is large. But the data at least offer some useful lessons.

The break has ranged from 50-50 to a maximum of 23-66 (the rest going to third parties in 2000).  For incumbent parties the median is 42 and for challengers 53. (Means are less different-- 43 for incumbents to 50 for challengers.) 

What is not accounted for here is dropping out. I've not calculated the percent who don't vote, which is sometimes substantial. Here I wanted to focus on those who actually voted and how that split. Voters count more than non-voters in this case.

I think the most important result for this Tuesday is that the last 60 years give little support for a massively lopsided vote among undecided. At most, a 2-1 split is as good as it gets.