Questionnaire Dwelling Unit-Level and Person Pair-Level Sampling ...

outwinsor percentages represent the total amount of residual weight when the weights are
trimmed to the critical values (used for ev definition) relative to the total sample weight. For
evaluation purposes, the outwinsor percentage is considered the most important of the three
percentages, as this gave a measure of the impact of winsorization (or trimming) of ev weights
(if we performed this treatment). See Sections 5.1 and 5.2 for the domains that were used to
define extreme values.
8.3 Slippage Rates
The slippage rate for a given domain is defined as the relative percentage difference
between the sampling weights and the external control totals, both before and after ps. The
control totals for QDU and person pair ps are derived from the screener dwelling unit (SDU)
weights, which were poststratified to U.S. Bureau of the Census population estimates (Chen et
al., 2008). Table F.1 displays QDU national domain-specific weight sums for both before and
after ps, as well as the desired totals to be met through ps. Table K.1 shows the same for the pair
sample. These tables also show the relative percentage difference, or the amount of adjustment
necessary (positive or negative) to meet the desired totals. The first relative difference is used
explicitly during the ps modeling procedure to identify potential problems for convergence.
Large differences in domains with relatively small sample sizes are indicative of potential large
adjustment factors, which may cause problems in convergence while satisfying bound
constraints. The reason is that adjustments required for one domain may have an adverse effect
on another domain when a unit belongs to both.
As an example, consider that Table F.1, for the 2006 QDU domain household size of one,
indicates a sample size of 5,641 with a total design-based weight of 30,195,407 and a census
total of 30,104,368 with an initial slippage rate of 0.30 percent, which would imply a common
weight adjustment of ≈ 0.996985, if this were the only calibration control. Similarly, looking at
pair data in Table K.1, the pair domain category of pair race other has a sample size of 956, a
design-based weight of 12,778,059, and a census total of 13,369,146, showing an initial slippage
of -4.42 percent. The resultant required adjustment would be ≈ 1.046258, if this were the only
control. However, in the generalized exponential model (GEM), all controls are simultaneously
satisfied under a complex algorithm that allows for different adjustment factors for different
units.
8.4 Weight Adjustment Summary Statistics
Tables G.1, G.2, and L.1 through L.3 display summary statistics on the product of weight
components before and after all stages of adjustment for the QDU and person pair, respectively.
The summary statistics include sample size (n), minimum (min), maximum (max), median
(med), 25 th percentile (Q1), 75 th percentile (Q3), and the unequal weighting effect (UWE). Note
that in Tables L.2 and L.3 the sample size for pair age group, pair race/ethnicity, and pair gender
are slightly different. This is because those variables were defined using screening demographic
information in the nonresponse adjustment of respondent pairs, while in the poststratification of
respondent pairs, they were defined from questionnaire demographic information. Because UWE
is directly affected by weight adjustment factors and extreme weights, these values—along with
the percentage of extreme weights as noted in Section 8.2—were used as guidelines for
determining model adequacy.
74