The census will be wrong. We could fix it.

I have an op/ed in tomorrow’s Washington Post about statistical sampling and the census.  It boils down to the claim that by failing to use the best statistical techniques we have to enumerate the population accurately, we’re getting the answer wrong on purpose in order to avoid getting it wrong by accident, and possibly violating the Constitution as a result.  And that estimating an unmeasured quantity to be zero is a really bad estimate.

The book Who Counts?  The Politics of Census-Taking In Contemporary America, by Margo Anderson and Stephen Fienberg, was an invaluable resource for the piece — highly recommended.

One argument I cut for space involves Kyllo vs. US, in which the Supreme Court ruled, in an opinion written by Antonin Scalia, that the use of a thermal imaging device to detect heat coming off the exterior wall of a house, and thus to infer the presence of a drug operation inside, can constitute a “search” for Fourth Amendment purposes.  On the other hand, Scalia questions the constitutionality of statistical adjustment of the census, expressing doubt that such a procedure would still be an “actual enumeration” as required by the Constitution.  So, for Scalia:

• “Search,” in 2010, includes a scenario in which something of interest inside the house is not seen or otherwise sensed by any person or people, but is inferred by means of a scientific instrument that didn’t exist in Constitutional times.
• But “enumeration,” in 2010, does NOT include a scenario in which the population is not counted one by one by any person or people, but is inferred by means of a statistical instrument that didn’t exist in Constitutional times.

Is that a problem?

Update (4 May):  It turns out I’m not the first arithmetic geometer to weigh in on census adjustment.  Brian Conrad in the New York Times, August 1998 does in three sentences what took me 1000 words:

Human intelligence plus a little brute force is often far more efficient and accurate than brute force alone. This is why statistical sampling is the superior way to carry out an ”actual enumeration” of a large population. Just ask any Republican who relies on a poll or who takes a blood test rather than drain every drop from his body.

Is the Washington Post congenitally incapable of doing math?

From today’s WaPo opinion page, Charlotte Allen tells her fellow women they stink at stuff:

Women really are worse drivers than men, for example. A study published in 1998 by the Johns Hopkins schools of medicine and public health revealed that women clocked 5.7 auto accidents per million miles driven, in contrast to men’s 5.1, even though men drive about 74 percent more miles a year than women.

Note that the figure given is accidents per miles driven. Which means that “men drive more miles a year than women” has nothing to do with the point the author is attempting to make. It would make as much sense to write “even though dark-horse Juno failed to win the Best Picture Oscar.”

But the version as published contains a percentage! A deftly inserted percentage never fails to give a soothing impression that somebody, just off-stage, is industriously doing some science.

Allen goes on to observe that

The only good news was that women tended to take fewer driving risks than men, so their crashes were only a third as likely to be fatal.

“Don’t drive with her, drive with me — I’m a much better driver, because I’m 10% less likely to get into a fender-bender, though I suppose I ought to mention that I am three times more likely to kill us both.”

The article also remarks that “[n]o man contracts nebulous diseases whose existence is disputed by many if not all doctors” (Gulf War Syndrome, call your agent!) and that men don’t miss work because they’re depressed (you too, William Styron and Pete Harnisch!) Oh, and that women are congenital failures at analytic thinking in general, and math in particular. Is it possible the article is some kind of bizarre performance piece meant to illustrate this latter point?