Down for the Count?

Every census of a vast country like the United States is an estimate. Millions don't respond to the mailed census forms, and every front door can't be visited by follow-up head-counters - particularly in tightly packed urban areas.

The count came up so short in 1990 (at least 10 million) that the Census Bureau devised a plan for using sampling methods to arrive at a more accurate estimate next time around, in 2000. Sampling is an almost universally accepted statistical tool. But Republicans in Congress have dug their heels in - no sampling!

Why? Sampling's critics may say it's because the Constitution specifies an "actual enumeration." But the Constitution also says that the counting shall be done "in such manner" as Congress directs. There's nothing barring techniques like sampling. The real issue here is political, not constitutional. Some in the GOP don't really want a more accurate count of the hardest-to-find Americans, the poor and new immigrants who typically vote Democratic. Larger numbers in those categories could affect the political character of congressional districts allotted to states after 2000, when the new census becomes the basis for reapportionment. Specifically, it might become harder to create "safe" Republican House seats.

But the effects of an undercount go beyond representation. They can skew the distribution of a range of federal assistance programs, since localities partake according to their populations. Beyond governmental concerns, businesses assessing markets and researchers analyzing society rely on census numbers.

After 1990, the calls for improvement were loud. The sampling procedures drawn up by the Census Bureau are a far cry from "guessing," as some charge. The counting process would begin with the traditional mailed census questionnaire, sent to every dwelling on a master address list for the country. In 1990, about 65 percent of households responded. Follow-up interviewers will contact a large number of those who don't respond, with an emphasis on areas with high rates of non-response. The bureau hopes this will boost the total contacted to 90 percent.

But that leaves 10 percent uncounted, and now the going gets tougher. This is where sampling would have its biggest impact. A sample of 25,000 census "blocks" would be chosen for a second close, physical canvassing of every residence - a step that wouldn't be practical for the whole country. The results of this canvass would be compared to the earlier head count. "Estimation factors" would emerge that could be used to correct counts in all blocks, with a close eye to corresponding demographic features like homeownership, race, and age of residents.

This spring, the bureau will conduct some dress rehearsals of this system in geographically varied parts of the country. Congress allowed for that much. But a full-scale gearing up for 2000 remains problematic.

Preparations for the dress rehearsals have underscored another problem facing the census: It's difficult to find workers to conduct the count. With today's very low unemployment, few jump at the short-term, no-benefits census jobs. This problem will be exacerbated if Congress orders a labor-intensive, no-sampling national head count.

Meanwhile, the Census Bureau is having to split its management - one part moving ahead with the sampling plan, another working on contingency plans in case Congress flatly rules out sampling. Congress's own General Accounting Office just issued a report warning that continuing indecision over census methods could imperil the 2000 count.

One other note: If the GOP leadership in Congress has its way and demands an "actual" count, the price could be at least $1 billion higher than the sampling approach.

For a more sensible, and accurate, census, Washington's politicians should back off and let the experts in the Census Bureau apply their apolitical expertise.