Monday, July 13, 2015

Reconciling Different Household Counts from Census Bureau Surveys

by Dan McCue
Senior Research Associate
One of the major challenges faced by housing analysts and demographers is the lack of consistency among various Census Bureau surveys.  Particularly troublesome is the persistently wide range of difference reported by surveys in the number of households in the US, a key measure of housing demand.  In 2013, for instance, household counts reported by various Census Bureau surveys ranged from a low of 114.6 million in the Housing Vacancy Survey (HVS), to 116.3 million in the American Community Survey (ACS), to a high of 122.5 million in the Current Population Survey’s March ASEC (CPS/ASEC) - a span of fully 7.8 million households (Figure 1).  Annual surveys also differ widely in their measures of growth in the number of households, confounding efforts to gauge recent trends.  Indeed, household growth measures for 2013 ranged from 0.3 to 1.4 million depending on the survey, leaving data observers unsure whether growth in that year was historically weak or incredibly strong.

Source: JCHS tabulations of US Census Bureau data.

Given their importance to much of our work, the Joint Center has recently released a research note to highlight some of our current thinking on Census Bureau survey counts including: the differences in household counts among annual Census Bureau surveys and their causes; which surveys we believe to have more accurate counts than others and why; how we use different household count data for different analyses at the Joint Center; and how much of a difference using alternative headship rates based on different household counts would have on the current JCHS household projections.

The research note finds that the major source of difference among survey counts of households appears to be whether or not the surveys are person-based or stock-based.  Person-based estimates, which count the number of people who report being heads of households, consistently result in higher numbers of households than stock-based estimates, which count the number of housing units that are occupied.  We don’t exactly know why there is such variance in person-based and stock-based survey results, but the magnitude of difference between these two approaches is big and can be roughly approximated by the difference in household counts of the HVS and CPS/ASEC, because they are essentially stock-based and person-based versions of the same underlying survey sample. 

Among the annual surveys, the person-weighted CPS/ASEC has an advantage in that its household counts have come in closest to decennial census counts, which we take to be the benchmark. It is also a relatively timely survey and has the longest track record of matching census growth over the decades.  However, the CPS/ASEC does have a number of other shortcomings. One of the its major shortcomings, year-to-year volatility, can be reduced by smoothing over the data using rolling averages, but that approach also reduces the timeliness of the survey for measuring short-term trends in household growth. 

The second major annual household survey from which to get household counts, the stock-based HVS, is the most timely measure, providing quarterly results in addition to annual counts, and it also offers a series of annual household counts that use a consistent weighting vintage, which provides a more stable framework for measuring annual household growth trends than surveys that adjust their underlying survey controls year to year.  However, those vintage weighting controls do not eliminate a high amount of annual volatility.  Recent results underscore this fact, for it is highly unlikely that household counts in the HVS jumped fully 1.3 million between the third and fourth quarters of 2014 as reported in the HVS.

Finally, lack of timeliness is also a major disadvantage of using the third major annual census survey, the ACS, for analyzing annual counts of households: while other annual survey results have been out for months, the 2014 ACS is not due out until late 2015.  The ACS also has much lower household counts and appears to be essentially a stock-based survey with slightly higher household counts than HVS that is most likely a result of its more inclusive rules for determining occupancy of a unit.  Still, even with its lower base of counts, as a large and detailed survey the ACS may prove to provide a reliable measure of the growth trend in households, but so far this survey has had only a few years under a consistent weighting methodology in which to judge its reliability.

Overall, there is no satisfying conclusion to which annual survey is best for measuring annual household growth trends and none is perfect. CPS/ASEC is volatile but can be smoothed over at the expense of timeliness; HVS is timely and attractive in that its counts are pinned to stable consistent weighting vintages across years but it is still volatile, and possibly the vintage controls bias growth too low; and ACS is not timely enough to be helpful in measuring recent trends and is a survey without much history in which to judge its accuracy, though it has promise as a large and detailed survey that receives a relatively high amount of resources from the Census Bureau.  In terms of the number of households, however, there is reason to believe that higher counts of the CPS/ASEC, obtained from person-based weighting approaches, do appear to be preferable to lower counts of the HVS and ACS in offering counts closest to that of the official decennial censuses.  It is largely for these reasons we have used household counts from CPS/ASEC as a key input in Joint Center household projections, which apply current headship rates (the ratio of households to people) to population projections to produce household projections that form a baseline for estimating future housing demand.




No comments:

Post a Comment