요약
Weight adjustment for non-response and, to a lesser degree, coverage error are applied to
the initial weights to produce the final analytic weights. Frequently the adjustment is applied
using post-stratification. As Cochran suggested, the effectiveness of the post-stratification
adjustment completely depends on available true population parameters, or the control totals.
In practice, however, the population parameters are frequently estimated directly or indirectly
through modeling. This paper demonstrates potential adverse effects of using incorrect
population parameters in post-stratification adjustments, by applying log-linear models to a
cross-tabulation of sex, race/ethnicity, and age from 2010 U.S. Census data. In particular, a
caution is advised when we apply raking ratio adjustments to the initial weights with estimated
control totals - in general, the last step for well-designed surveys but frequently the only step
for not-so-well-designed surveys.