Tuesday, March 22, 2011

Lawler: On Census Housing Stock/Household Data

by Bill McBride on 3/22/2011 05:12:00 PM

CR Note: I'm trying to use the Census 2010 housing stock and vacancy data to estimate the excess vacant housing supply. Here are few "data nerdy" words of caution from economist Tom Lawler for anyone trying to use this data (including me):

CalculatedRisk is one of the few entities (people?) besides me who is tracking the release of Census 2010 housing stock data (total, occupied, and vacant) that have been released so far, and yesterday it put up a spreadsheet showing state data from Census 2010, Census 2000, and Census 1990. It is available here. As of this morning Census had released data for 42 states for 2010.

A word of caution (as I’ve noted before): the “official” Census data from both 2000 and 1990 appear to “undercount overall housing units, with the “undercount” being larger for vacant homes than for occupied homes. As a result, overall (or gross) vacancy rates are understated in the “official” Census data – though by how much is a subject of debate among analysts.

For both Census 1990 and Census 2000, Census analysts conducted a “Housing Units Coverage Study,” which uses data from the “Accuracy and Coverage Evaluation” (A.C.E.) to estimate the “accuracy” of the Census housing counts. The 2000 HUCS was released in February, 2003, and is available here (PDF). As with estimated population over/under counts, Census has not used these “post-Census” studies to change the “official” Census figures.

Here are some different data for occupied and vacant homes as of April 1, 2000. Note that the ACS data are based on a sample, and ACS assumes that the “official” housing unit estimate is “correct.”

 Official Census 2000Including Reinstated UnitsDual System Estimate ACS
Total Housing Units115,904,641115,877,639116,586,458115,904,641
Occupied105,480,101105,463,423105,808,904104,819,002
Vacant10,424,54010,414,21610,777,55311,085,639
Gross Vacancy Rate8.99%8.99%9.24%9.56%


The HUCS notes that the occupied plus vacant housing unit dual system estimates do not quite add up to total housing units because of “rounding.”

As I’ve noted before, the higher vacancy rate from the ACS C2SS was surprising to some given its “2-month” vs. “usual” residence rule. Some believe that since the main purpose of the decennial Census is to count people and not necessarily housing units, there may not have been enough “emphasis” placed on “finding” all vacant housing units – though not everyone agrees with that view.

The Census 1990 HUCS also estimated that the official Census 1990 housing unit counts missed more vacant than occupied units, and also “undercounted” total housing units by about 0.96%. The official “gross” vacancy rate for 1990 was 10.09%, while the HUCS suggested that a more “accurate” GVR was 10.46%. (There are other significant problems with the Census 1990 housing stock data by “type of structure,”)

Neither the 2000 nor the 1990 HUCS provided estimates of housing unit count “misses” by state, though they did provide some data on misses by broad Census region. In 1990 the housing unit undercounts were greatest in the South and the West, as were the gross vacancy rates. The Census 2000 HUCS “showed” less variability in estimated “misses” and vacancy rates across regions, though the gross vacancy rates were estimated to have been understated the most in the Northeast and the West.

The other “source” of housing vacancy data from the Census Bureau, of course, is the quarterly Housing Vacancy Survey, which compiles quarterly averages of monthly data based on a much smaller sample than either the ACS or (of course) the decennial Census. As I’ve noted many times, the HVS is NOT designed to estimate the housing stock; rather, the HVS uses “official” Census estimates of the housing stock (including the decennial Census, as well as subsequent estimates which “build up” off the decennial Census data, building permits, and “crude” estimates of the annual loss in the housing stock based on suspect data from the AHS). The HVS then assumes that its vacancy RATES by type are correct, and based on these rates and the “exogenous” (to the HVS) housing stock estimates, the HVS derives occupied and vacant housing units. If in fact the decennial Census focused on counting “people” and not on counting vacant housing units, however, then it actually doesn’t make “sense” for the HVS to assume that Census’ housing unit estimates are correct.

Instead, IF the HVS vacancy rates were correct, then in all likelihood that would imply the Census estimates of the housing stock were too low. If the Census housing stock estimates were correct, that would imply the HVS vacancy rates were too high. Truth may lie in between, but based on some of the HVS’ state and MSA vacancy rate data, it appears extremely likely that the HVS vacancy rates are way too high – mainly in the “rental” and the “other” category.

As is clear from various Census publications, not just counting all housing units but also determining whether a given housing unit is vacant or not can actually be tricky, and getting an accurate tally often involves having to do multiple “calls,” trips, or follow-ups. Many analysts believe that the HVS, which is a voluntary survey, doesn’t produce very accurate vacancy rates in any given quarter – in any words, there is high degree of “non-sampling” error, with a bias toward a vacancy rate that is too high (and, by the way, a homeownership rate that is too high.) In addition, the HVS’ sample size for some areas is extremely small, and since housing markets across the country are not even remotely close to homogenous, there is a pretty high (and not as easy to calculate as one thinks) sampling error as well.

I’m only putting all this in because it’s a lot trickier than it should be to look at available government information/data, including time series that for the most part are not consistently derived over time, to estimate the current (much less last year’s) “excess” supply of housing.