Univac: What went wrong?

by Cliopatria

In 1952, a Univac computer, with some relatively primitive algorithms and less less raw computing power than most modern appliances, correctly predicted the results of the election hours before final returns were known and in stark contrast to the predictions of human experts.

Somewhere along the way, we lost that ability: polling in the last few election cycles has been pretty weak as a predictor, and we've got millions of times more computing power and immensely denser and better data to work with. Even the most direct comparison, election-night exit polling, failed miserably last time, and not just in Florida.

Are the races really tighter? Is that the problem? Or are we outthinking ourselves: has the Heisenberg effect (I know, physicists hate it when social scientists use this shorthand, but it's not going away) overwhelmed our statistical models so that the ubiquity of polling is its own downfall?

comments powered by Disqus

More Comments:

Jonathan Dresner - 10/31/2004

I hadn't run across that story (I'm pretty well versed in Asimov and Clarke, too; but they both wrote so much), but I've always loved the story in which the "most qualified" citizen is chosen by computer based on a voluntary testing program: economics, military affairs, political theory, sociology, etc., were all required. Unfortunately, the problems of the presidency became so complex that there was no one qualified individual anymore..... so a group of people had to take the test for a single, pretty well qualified, figurehead.

I agree with your comments about KC's comments about the gay marriage amendments: he's the only person I've seen talking about their strategic value to the Republicans, so I strongly suspect that the statistical models haven't been adjusted for it.

Lloyd Kilford - 10/31/2004

There's a story by one of the Golden Age science ficton writers (either Asimov or Clarke) where the US elections are decided by choosing one (representative and statistically significant) voter and interviewing him to find out what the country thinks.

I think that there's going to be some interesting randomness brought in by the nature of the USA - I think KC Johnson's post above talks about how the presence of an initiative on banning gay marriage on the ballot could drive turnout up and help the Republicans in some states. The polling model's going to have to be really good to cope with that (and possibly people who are going to vote for this amendment might not want to tell pollsters that).

Ralph E. Luker - 10/31/2004

Derek, Your comment may be correct on the facts. I'm no expert, so I don't know. Is it not sufficient, however, to be correct on the facts? Your last two sentences don't serve the purpose of interpretation of the facts. They are, simply put, mean spirited. A Red Sox fan can afford to be generous and gracious these days.

Jonathan Dresner - 10/31/2004

Unfortunately, the corrective algorithms in polling are based on guesses about what went wrong in the past, and don't seem to keep up terribly well with what's happening in the present.

"Reasonably accurate" would be fine, if we just did them now and then and didn't give them much weight.....

Jonathan Dresner - 10/31/2004

There's really two different issues, poll-wise. Election-day polls are particularly insidious, and the mostly voluntary suppression of that data until polls have closed is a trend to be encouraged. But the failure in Florida is going to be replicated elsewhere, if the current polling data means anything.....

which leads us to the other issue: after decades of campaign season polling, why can't aren't the models more stable and reliable? There have been several examples, in recent years, of election day results that were widely divergent from poll results in the final week of the campaign (a friend of mine pointed out that three examples she was aware of from the last cycle all happened in Diebold voting machine states and all broke in the Republicans' favor..... I'm just saying), which calls into question the utility and the effect of doing these polls.

If we want to know how things are going (and who is 'we' anyway: I'd much rather know how the policy is being made than how the rhetoric is polling) it seems that these polls are not serving that purpose, either.

Thomas Joseph Sawyer - 10/30/2004

The 1952 model of projecting a final winner based on early results evolved almost to perfection. It became so valid that we "Banned It" claiming it was affecting outcomes. Actually, the networks voluntarily quit releasing their early results (Until the polls closed)just ahead of a congressional ban.

Of course, as a cost savings measure, the networks all pooled their resources to have only one set of sample precincts used by all. The problems with calling the vote in Florida based on this sample was re-hashed adnauseum in November-December of 2000.

Derek Charles Catsam - 10/30/2004

Jonathan --
I actually have an issue with your premise that polling has been wrong in recent election cycles. Of what elections are you speaking? In the last election, the polls actually had it very nearly right. That they could not get the fine points of the scant margin in Florida that turned the Electoral College, and only that with an arbitrary deadline set by the Supreme Court (which is where the Supreme Court went wrong and that no one bothers to notice -- thay actually said a full recount would be ok, but that it had to be done by a date that would have been impossible, as if giving another week would somehow have been impossible to do), does not make the pre-election polls wriong. Most all had an incredibly close election. Most had Gore winning slightly. How were the polls wrong? The polls were similarly accurate, even with the presence of a whack-job third part megalomaniac, in 1996 and even by election eve 1992. The polls were right in 1988 and 1984. In other words, the factual root of your entry is, well, completely, totally, absolutely wrong.
Good post otherwise.
dc

Lloyd Kilford - 10/30/2004

The polling in 1992 was spectacularly off (predicting a hung parliament or a Labour victory) when in fact the actual result was a solid although quite narrow Conservative victory. I think that one of the problems then was that many Conservative supporters were unwilling to admit this to the pollsters.

As far as I know, there is now a correction term in the polling to allow for this. Polls here seem *reasonably* accurate.

Ralph E. Luker - 10/30/2004

Aren't there some values in tension here? There's our desire to know how things are going in the campaign, which your post addresses, but there are also privacy rights which the pollsters invade and I suspect that voters in Hawai'i would be among the first to complain if pollsters told us _all too clearly_ well in advance of the closing of the polls in the islands who was the certain winner in the presidential contest.