|A Comparison of Taylor Linearization and Balanced Repeated Replication Methods for Variance Estimation in Medical Expenditure Panel Survey
|For computing sampling variances of the Medical Expenditure Panel Survey (MEPS) Household Component estimates, the Taylor linearization method is generally used.
The MEPS public use files include variance strata and cluster identifiers to facilitate variance computation using this method.
Also a file containing a BRR replication structure (in the form of a set of half sample indicators) is also made available so that the users can form BRR replicate weights from the final MEPS weight to compute variances of MEPS estimates using either BRR or Fay's modified BRR (Fay 1989) methods.
These replicate weights are useful to compute variances of complex non-linear estimators for which a Taylor linear form is not easy to derive and not available in commonly used software.
However, the BRR replicates derived from the final weight represent a shortcut approach because the replicates are not produced starting with the base weight and all adjustments made in different stages of weighting are not applied independently in each replicate.
So the variances computed using the one-step BRR do not capture the effects of all weighting adjustments.
The Taylor approach, as implemented in most software, also does not fully capture the effects of different weighting adjustments.
Of particular interest here is the effect of adjustments using external control totals which are expected to reduce the variance.
|Rockville (MD): Agency for Healthcare Research and Quality