Thursday, February 02, 2017

Lawler: The Household Conundrum, Part II: The American Community Survey (ACS) Data

by Bill McBride on 2/02/2017 11:02:00 AM

Earlier:
• Lawler: The Household Conundrum, Part I: The CPS/ASEC Data

• Lawler: The Household Conundrum, Part I Continued: The CPS/ASEC Data for 18-29 Year Olds

From housing economist Tom Lawler: The Household Conundrum, Part II: The American Community Survey (ACS) Data

Not including the Decennial Census, the American Community Survey is the largest survey that the Census Bureau administers, and it collects detailed socioeconomic information previously collected in the so-called “long form” of the Decennial Census from about one in every six households. The ACS is also a “mandatory” survey, as opposed to other Census surveys such as the CPS making “non-response” issues less of (though still somewhat of) an issue. Given its substantially larger sample (and in the past more timely sampling frame) than other periodic surveys, the ACS is widely viewed as the most reliable (though far from perfect) source of information on socioeconomic trends in the United States (and more especially for regional data).

From the perspective of certain types of housing analysis, however, the ACS data is of limited usefulness. For one, the ACS has only been fully up and running since 2005, and as such there is not a lot of historical data.

For another, ACS data are only available annually, and annual estimates (which are yearly averages) are based on population and housing units estimates available at the time of that year’s survey, that is, prior year estimates not updated to reflect revisions in historical population and housing unit counts. Stated a different way, the “time series” of ACS estimates is not consistent with the latest historical estimates either of the US population or of the US housing stock. (Time series estimates of households from the CPS/ASEC are also not consistent with revised estimates of population counts.)

There are also a few “technical” issues: e.g., the ACS’ “residence” rule (has the householder lived in the surveyed unit for over two months) is different from that of the Decennial Census (is this home the “usual residence” of the householder). While there is no clear evidence that this “residence rule” results in materially different household estimates, it’s still worth noting.

Having said that, below are comparisons of various household estimates from the ACS, the Decennial Census, and the CPS/ASEC for 2010. I have adjusted the CPS/ASEC estimates to reflect by “best guess” of what the estimates would have been if the CPS/ASEC had used Census 2010 counts. I have also adjusted the ACS estimates to reflect the estimated undercount of the housing unit count form the Decennial Census from the post Census Coverage Measurement Study (almost all of that undercount was in vacant housing units). The Census 2010 household counts are also adjusted to reflect the post-Census Coverage Measurement Study, although that undercount was extremely small. Also shown are homeownership rates estimates.

Table 1: Various US Household Estimates by Age Group, 2010 (000's)
  Census
(Apr. 1)
ACS
(Average)
CPS/ASEC
(Mar.)
Total116,752115,255118,760
15-245,4034,9316,325
25-3417,96317,90119,173
35-4421,29821,38721,729
45-5424,91524,79325,113
55-6421,34721,22020,777
65-7413,50913,29613,400
75+12,31911,72712,243

Table 2: Various US Homeownership Rates by Age Group, 2010 (000's)
  Census
(Apr. 1)
ACS
(Average)
CPS/ASEC
(Mar.)
Total65.1%65.4%67.1%
15-2416.1%14.7%22.8%
25-3442.0%41.3%45.0%
35-4462.3%61.9%65.5%
45-5471.5%71.7%73.5%
55-6477.3%77.9%78.6%
65-7480.2%81.1%82.0%
75+74.5%75.7%78.5%

The ACS household counts are derived via a multi-stage process, but in effect the number is “controlled” to estimates of the US housing stock. Since the ACS vacancy rate estimate was higher than the “actuals” from the Decennial Census, the “occupied” housing unit count (or “households) was lower. The CPS/ASEC household estimates, in contrast, are controlled to population count estimates and do not take into account housing unit estimates. Thus even though CPS-based estimates of housing vacancy rates in 2010 were much higher than both Census and ACS estimates, the CPS/ASEC household estimate is higher. (The CPS-based household estimate from the Housing Vacancy Survey, which is controlled to the housing unit estimates, is materially lower than the ACS estimate). It is worth noting that if ACS estimates were controlled solely to population estimates and ignored housing unit count estimates, the household estimate would probably be higher (as was the case in 2005.)

Focusing on Table 1, ACS estimates by age group on balance match those of the Decennial Census better that those from the CPS/ASEC, though the ACS estimates for under 25-year old householders is a bit low (the CPS/ASEC estimates for this age group are way too high), and the same is true for householders 75 years or older.. Interestingly, household estimates for 35-64 year old householders are very similar in all three columns.

On the homeownership front, ACS estimates are significantly closer to Decennial Census results than CPS/ASEC estimates not just in aggregate, but across all age groups.

From the standpoint not just of aggregate household estimates but also household estimates by age and homeownership estimates, the ACS appears to be superior to the CPS/ASEC. It is still true, however, that ACS household estimates vary significantly from Decennial Census estimates.

The major reason the latter is true is that the ACS estimate of the housing vacancy rate in 2010 (13.07%) was well above the Decennial Census estimate adjusted to reflect post-Census coverage measurement (11.88%). (The CPS/HVS housing vacancy rate for all of 2010 was 14.34%).

It appears as if part of the ACS’s higher vacancy rate reflected misclassification of occupancy status (a “matching” of ACS to Decennial Census showed more units misclassified as vacant that units misclassified as occupied), though part may also be related to the different sampling frame (the 2010 ACS for the most part did not use the updated Master Address File compiled as part of the Decennial Census).

(more later)