Some people will not return the forms. In 1990, only 65 percent of households did, down from 75 percent in 1980. The falloff is usually laid to many factors: a weariness with all surveys; distrust or fear of government; a scarcity of time; less civic virtue, and confusion of the census forms with junk mail. Hundreds of thousands of door-to-door canvassers will then try to fill in the gaps. They, too, will fail, because some people won’t open their doors, some people don’t want to be found, some people give bad information and some canvassers make mistakes. What to do?
For the 2000 census, the Clinton administration proposed–for the first time in U.S. history–not to try to count every American directly. Instead, the Census Bureau would stop at 90 percent of households and estimate the last 10 percent by statistical sampling. It was this plan that the Supreme Court last week struck down as illegal. No one should have been surprised, because the law (the Census Act as amended in 1976) is clear. The relevant section explicitly prohibits sampling, ““for the determination of population for purposes of apportionment of Representatives in Congress among the several States . . .''
Although the administration was skirting the law, it has enjoyed the public-relations advantage in the feud over the census. The conventional wisdom goes like this. A population ““undercount’’ mostly affects the poor, minorities and the cities where they live. Because congressional seats and federal funds are based on population figures, these groups are shortchanged. Almost all scientific experts think the undercount could be cured by statistical sampling. But congressional Republicans oppose sampling for purely partisan reasons: it would cost them seats.
There’s some truth here–and much exaggeration. Start with the undercount. It’s fairly small. For 1990, the Census Bureau estimates (and this, too, is inexact) that it missed 1.8 percent of the population. Among blacks the figure was 5.7 percent. Both figures were slightly higher than in 1980 and halted four decades of improvement. Even so, few social and economic statistics achieve the 98 percent accuracy of the overall population figure.
Next, the undercount’s effects. They’re also modest. In 1996, about $180 billion–including highway and Medicaid funds–was distributed by formulas that included population. Perhaps that’s now $200 billion. But possible population errors are so tiny that the distribution would barely change. The National Research Council cites a study that puts the shift at three tenths of 1 percent. On $200 billion, that’s $600 million.
How about congressional reapportionment? Well, in 1991 the Census Bureau estimated (based on a post-census survey) how much state populations might be adjusted for the undercount. With these numbers, perhaps three states would have gained another House seat (California, Georgia and Montana) and three would have lost (Oklahoma, Pennsylvania and Wisconsin). Normal swings in states’ populations–mainly away from the Northeast and Midwest toward the South and West–produce much larger reapportionment changes.
Still, sampling would be desirable if it clearly improved the population counts, and many statisticians think it would. Unfortunately, that’s not certain. Sampling is already widely used for parts of the census. One in six households gets the ““long’’ form that asks questions about income, education and housing, among other things. The results from these questions reflect sampling. But all samples have margins of error.
The plan to use sampling to offset an undercount would work something like this: after the census, you do a large sample survey of a cross section of Americans; then you compare each response in the sample with the same individual’s response for the census; some people show up in the sample and not the census; if 2 percent of the sample wasn’t captured by the census, that measures the undercount; the results would then be projected nationally. Sounds simple.
It isn’t. One obvious question is: if the census missed Joe–or Joe didn’t want to be counted–why would the sample survey find him? There is no obvious answer. And ordinary errors may mean that some people caught by the census are reported missing by the sample. For these and other reasons, some experts think that the sample could reduce accuracy.
The trouble now is that the Supreme Court hasn’t settled the census dispute. The White House’s view of the court’s ruling ensures otherwise. The court, says the administration, bars the use of sampling only for the apportioning of congressional seats among states; but the ruling permits sampling-adjusted population data for the ““redistricting’’ of seats within states by legislatures and governors. So the Census Bureau is considering one set of figures for national congressional reapportionment and another (with adjustments) for the states to use for internal redistricting. Naturally, the Republicans say this flouts the court’s ruling and are threatening to withhold funds for sampling. What we can expect is more congressional strife and more lawsuits.
Democrats’ passion for sampling is no less political than the Republicans’ aversion to it. Each party seeks the greatest advantage in crafting districts that favor its candidates. But adhering to sampling after the court’s decision would create at least a problem of perceptions: issuing two sets of numbers for uses that are primarily political would look corrupt, even if it weren’t. It suggests the customizing of statistics for political ends.
There’s no way to count the population exactly. The question is not whether sampling would produce slightly better or worse numbers. This is a close call on which reasonable people can disagree. But until the case for sampling is overwhelming, we ought to stick with the traditional head count. Doing otherwise threatens the integrity of a system that, until now, has enjoyed public confidence.
