The goals of the research described in this paper were to investigate and better understand the effects of GQ population on the variances of estimates of the total resident population and the mechanisms driving these estimates of variances. To accomplish these goals we examined the direct effects of GQ estimates in our descriptive analyses. We also examined two alternative weightings to understand how certain steps of the GQ weighting affect the variances of the GQ estimates.

Since the GQ population differs from the household population, the inclusion of the GQ population has noticeable effects on the estimates of total persons exhibiting given characteristics and estimates of proportions of the total resident population. The reasons for this effect are clear and expected. Of greater interest were the effects on the variances of the estimates of the total resident population, which are not as well understood, and, we suspect, will be larger for certain characteristics for small geographies, such as tracts and block groups.

We found that for state-level estimates for all nine of the characteristics we studied that the variances of estimates of the total resident population were close to those of the household population. This result is not surprising because the GQ sample design and weighting are designed to produce state-level estimates. For the county-level estimates, we found that for characteristics to which we do not control, the SEs and CVs change little. In contrast, for the characteristics to which we do control, we saw that the SEs and CVs of the total resident population were less than those of the household population. Controlling the total resident population at the county level prevented the variances of their estimates from being adversely affected when the GQ population was included in the total resident population. This suggests pursuing future research for tract-level estimates when the 2005-2009 5-year estimates are ready.

The experimental results examined the effects of two key weighting steps in the GQ weighting. When we did not account for the difference between the expected and observed populations, we saw that the experimental weighting generally produced lower CVs at the state and county level. This indicates that the second phase adjustment increases the variances and suggests that a better frame will produce more accurate results.

When we omitted the GQ controls, we saw that the experimental weighting had higher CVs at the state level but tended to produce lower CVs at the county level. These better state level estimates are consistent with our design goals. The lower CVs at the county level are likely attributable to the low coverage rates within certain major GQ types. We speculate that if we oversample those GQ facilities with low coverage rates, such as other long-term care facilities and other non-institutional facilities, and undersample those with higher coverage rates, such as juvenile facilities, we can expect to reduce the variances while maintaining the overall GQ sample sizes and costs.

Lastly, we note that prior research had shown large variances for the estimates of total GQ population for counties (Beaghen and Stern, 2009). This phenomena arises because typically all sample persons in a given GQ yield the same GQ population and thus instead of a sample of ten the effective sample is one. Another way of looking at it is that the intracluster correlation is one. Estimates of characteristics do not face this limitation.