"Cellphones Challenge Poll Sampling" (NYT)

Interesting article on how survey sampling in the US is being affected by the fact that many people no longer keep land-line telephones and rather have only cell phones. The article states that "the issue came up [during polling for the US elections] in 2004, but cellphone-only households in 2003 were 3 percent of the total. They now run 16 percent, according to Mediamark Research." More from the article:

According to data from the Centers for Disease Control and Prevention’s National Health Interview Survey, adults with cellphones and no land lines are more likely to be young — half of exclusively wireless users are younger than 30 — male, Hispanic, living in poverty, renting a residence and living in metropolitan regions.

The Pew Research Center conducted four studies last year on the differences between cellphone and land line respondents. The studies said the differences were not significant enough to influence surveys properly weighted to census data. With the increase in cellphone-only households, that may not be the case next year. Researchers, including the New York Times/CBS News poll will test that by incorporating cellphones in samples.

The estimates in the Health Interview Study suggest that cellphone-only households are steadily increasing.

“If the percentage of adults living in cell-only households continues to grow at the rate it has been growing for the past four years, I have projected that it will exceed 25 percent by the end of 2008,” Stephen J. Blumberg, a senior scientist at the National Center for Health Statistics, wrote in an e-mail message.

The American Association for Public Opinion Research has been examining the question and formed a group to study it. The association says it will issue its report early next year.

Clearly the limited impact for election polling does not apply to surveys that attempt to measure other population parameters. The weighting fix proposed by Pew is adequate when estimating simple population parameters (e.g. proportions), but things get much more dicey when we move to a regression framework. (See this article by Andrew Gelman for a recent take.)