Interstate Migration Has Fallen Less Than You Think: Consequences of Hot Deck Imputation in the Current Population Survey

Estimates were retrieved online (http://​www.​census.​gov/​population/​www/​socdemo/​migrate.​html).

After we noticed the discrepancy between imputed and nonimputed data, we corresponded with Census Bureau staff. They told us that the imputation procedure was changed in 2006 in such a way as to reduce the interstate migration rate. In response to this article, and after its circulation, the Census Bureau added a note on its website (http://​www.​census.​gov/​population/​www/​socdemo/​CPSnote.​pdf) that explains the change in imputation procedure and cautions users about its effects on measured migration rates. The change is also mentioned in a footnote in an unpublished Census Bureau working paper that compares the ASEC with the American Community Survey (Koerber 2007:14).

The 2009 ASEC tells us the fraction of Americans who moved between February–April 2008 and February–April 2009, while the 2009 ACS provides only an average of one-year migration rates for intervals from January 2008–January 2009 to December 2008–December 2009.

The ASEC measures mobility with retrospective questions: “Did this person live in this house or apartment one year ago?” and “Where did (reference person’s name/you) live one year ago?”

See Brown (1984) for an application of bounds to missing earnings data in the CPS.

Personal communication from David K. Ihrke, Journey-to-Work and Migration Statistics Branch, Housing and Household Economic Statistics Division, U.S. Census Bureau, Oct. 21, 2010. See also Koerber (2007).

In Kaplan and Schulhofer-Wohl (2011), we discussed and ruled out alternative explanations for the 1999 and 2000 changes, including changes in the CPS weights associated with the 2000 census and the expansion of the ASEC sample in 2001.

In the 2010 ASEC, the standard error of the interstate migration rate, calculated using the replicate weights that accompany the public-use data file, is 0.05 percentage point for the full sample and for observations with original migration data, and 0.18 percentage point for observations with imputed migration data.

Specifically, we categorize a respondent as having imputed migration data if migration status (whether the respondent lived in the same home one year ago) was hot deck allocated; if the person is coded as a migrant and state of residence one year ago was hot deck allocated; if the person is coded as a within-state or within-county migrant and the county of residence one year ago was hot deck allocated; or if the person’s migration data were inferred from a householder, parent, or spouse whose migration data in turn were hot deck allocated or inferred from yet another respondent whose data were hot deck allocated. For 1.6% of weighted observations, the imputation flags show that the migration data were inferred from another respondent, but this other respondent does not exist. This problem mainly occurs when an unmarried person has an imputation flag indicating “assigned from spouse.” We treat these unlinkable observations as having nonimputed data because there is no information to indicate that their data came from hot deck allocation. However, the migration rates in unlinkable observations follow the same pattern as the rates in imputed data, so some researchers may prefer to treat unlinkable observations as imputed. Classifying unlinkable observations as hot deck allocated does not significantly change our findings.

The change in imputation procedures had little effect on rates of migration from abroad and between counties in the same state. These results are available on request.

The same sort order must be used for imputing all variables because some imputations are functions of other imputations.

The ACS spends up to three months attempting to collect data from a given address, while the ASEC collects data only in a specific week. Thus, if the address is vacant in a certain week but occupied in any of the next 12 weeks, the ACS will find a migrant where the ASEC would find a vacancy. The ACS and ASEC survey universes also differ in some ways.

Before 2005, the Census Bureau’s source and accuracy statements for the ACS say the survey is designed to represent only the counties where data were collected.

The summary tables that the Census Bureau publishes on its American FactFinder web site (http://​factfinder2.​census.​gov), which are based on a somewhat larger data file, show a virtually identical overall interstate migration rate but do not allow us to distinguish imputed from nonimputed data or group-quarters residents from others.

Koerber (2007) reported that the ACS uses a hot deck allocation procedure in which donors and recipients always have the same current state of residence. Therefore, the ACS procedure does not generate the spurious interstate moves that the pre-2006 ASEC method produced when donors came from different states than recipients.

“SOI Tax Stats—Free Migration Data Downloads,” Internal Revenue Service (http://​www.​irs.​gov/​taxstats/​article/​0,,id=​212718,00.​html).

Although we hesitate to overemphasize small fluctuations in the data, we conjecture that the slight increase relative to trend in 2006 and drop relative to trend in 2007 may be related either to fluctuations in the housing market or to migration induced by Hurricane Katrina in August 2005.