Recount: Census Changes How It Estimates The Uninsured

Oops.

Last year, the U.S. Census Bureau estimated there were 50.7 million uninsured Americans in 2009. Today, it revised the figure for 2009 down to 49 million after adjusting the way it counts. What gives?

For years, researchers have complained that the Census Bureau overestimates the number of uninsured because of the way it accounts for missing data, a process that statisticians call imputation. A 2007 study in the peer-reviewed journal Health Services Research by researchers at the University of Minnesota found the Census in 2004 overestimated the number of uninsured by 1 percent or 2.5 million people.

Missing data as a result of non-responses are a common problem in survey research. About 11 percent of people who respond to the Census survey do not respond to all the questions on the survey. Statisticians employ various techniques to “guess” what the missing data is are. The Census Bureau uses an imputation technique called “hotdeck” to help fill in missing data by “borrowing” answers from a respondent with similar demographic characteristics.

Researchers studying the Census procedures said the bureau was filling in the missing data with a higher-than expected rate of uninsured people. Part of the reason was that the Census effort was trying to account for a possible bias in the survey. That bias counts more people as having dependent coverage than may be valid, and the researchers said the Census effort underestimated the number of people who had dependent care coverage.

To fix this problem, the Census Bureau said today it has changed how it accounts for the missing data by revising how it categorizes families and households. That makes it easier for the bureau to consider people who have not filled in the data as having dependent coverage.

Census put out a paper today explaining the change: Census revised its uninsured estimates for all years going back to 2000 as a result of the change.

Julie Sonier, deputy director of the State Health Access Data Assistance Center, a nonprofit group based at the University of Minnesota, said the change follows previous changes by Census to make the count of uninsured more accurate. She said the Census change was a direct response to her colleagues’ research at the University of Minnesota in 2007.

“Improving the estimates and the accuracy is an ongoing process she said,” she said. The center worked on the survey change with the Census Bureau.

A Census official cited lack of research dollars as reason the change took nearly five years to make.