The Election and the Census
The False Analogy of Exit Polls
Some pundits are citing exit pollsters' double flip-flop in predicting Florida election results as
evidence that we shouldn't use modern statistical methods in other arenas, most notably in the
national census. The real lesson is just the opposite.
"You just had the most vivid example of why trying to rely on mathematical models to
predict real humans is dangerous," said former House Speaker Newt Gingrich. The pollsters'
failure in Florida, he told a Fox News interviewer, is "why some of us are concerned about
sampling for the census."
Gingrich has it backwards. If the presidential election were organized the way the census
is, we wouldn't be in the situation we are now. The Census Bureau would have gotten the
election right the first time.
First, the Census Bureau would have started a decade ago to design this year's presidential
ballots, making them look the same for every voter in every precinct in the country. Every line
item would have been checked and rechecked by design experts, content experts and copy editors.
They would have considered readability and meaning and the technical requirements for
registering an understandable response. The ballots would have been tested and retested on pools
of typical voters. They would have been submitted to Congress for examination and approval, and
the technical aspects of distribution, voting and collection would have been planned long ahead of
time and made uniform nationwide. Voters would have received advance notice of the coming
election and formal reminders to cast their ballots.
When the polls closed on election day, the ballots would have been counted nationally,
using state-of-the-art tabulating equipment, by federal officials whose full-time jobs require them
to evaluate the quality and uniformity of the procedures and to improve them constantly. We
would have known the final tally last Wednesday morning, if not sooner even if the election were
as close as it appears to be. And many observers are already recommending procedural reforms
along these lines in order to improve the accuracy of all future vote counts.
Accuracy is critical, for elections as well as for the census. That is why some of the Census
Bureau's methods would work well for an election. But the goal of an election is very different
from the goal of the census. One seeks an accurate count of the choices made by some Americans
called voters. The other seeks an accurate portrait of all Americans themselves. That is why
sampling, the use of modern statistical methods, essential for an accurate census count, didn't
work so well in predicting the choices of actual voters.
Voters are by far a messier lot than citizens in general, hard to pinpoint before election
day. Voters are not selected by anyone, nor do they qualify to vote by any income characteristic
or any demographic other than age. Before and after voting day, they can and often do refuse to
answer questions about their preferences, change their minds or provide misleading responses.
They are self-selected. If they don't turn out, that's that. The government doesn't go after
the sick, the unaware, the uninterested or the hostile to demand that they make a choice among
the candidates. That means that a U.S. election is not a valid statistical sample of citizens' wants.
It is a winner-take-all horse race that anyone can enter and that a wildly unrepresentative and even
tiny cross-section of Americans might decide.
In an election, therefore, it would be ludicrous, not to mention unconstitutional, to deduce
voters' choices for their leaders from the tiny samples used by exit pollsters without an actual
one-by-one count, as Florida is proving.
The census, however, seeks a snapshot of the entire population at a given moment and in
precise geographic positions, not a decision among several options by whomever may be
interested in participating.
The census first tries to count as many Americans as humanly possible, nose by nose. But
it inevitably fails to hear from everyone. But the Census doesn't stop if you don't send in your
Census form. Unlike voting, the Census sends canvassers out to ring doorbells and prod non-responders for answers. It begs for answers over many weeks not just on one day. And
apparently unlike the election officials in Florida, the Census uses state of the art optic scanners to
read Census forms that have an error rate of less than 1%, two to five times below the reported
error rate of machine punch card readers in Florida. And unlike in Florida, when the Census
Bureau's machines can't read the forms, carefully trained people enter them by hand.
And finally because it is a different process with a different goal, the Census can, after
exhausting all other effective methods for getting an accurate count, make use modern statistical
methods to correct the historic undercount of millions of Blacks, Hispanics, Asians and American
Long ago, the census was as locally diverse, uncoordinated and optional as voting is now,
but that was back in the 19th century. Using modern statistical methods, along with traditional
counting methods the census should provide an amazingly accurate picture of the U.S.
population. The national election system can profit by studying the Census Bureau example to
achieve its own very different goals.