| |
Sample weights are
needed to derive population estimates from the survey sample.
In complex survey designs, such as that used in MEPS, sample data
must be multiplied by the appropriate weights to obtain unbiased
estimates for the U.S. civilian non-institutionalized population.
Survey weights typically reflect the probability of being sampled,
adjustments for nonresponse, as well as post-stratification adjustments
(Cohen et al., 1999). Post-stratification
is a method used to match subgroup distributions with that of
population estimates from another data source such as the census
or Current Population Survey. The investigator needs to apply
the correct weights for the planned analyses, as well as to understand
what the weights reflect.
In a simple random
sample, every unit in the population has the same probability
of being drawn. The fraction of the population that is sampled
is the sample size divided by population size. This is not the
case for complex survey designs. Understanding how weights would
be constructed for simple random samples, however, can help explain
how they are constructed for more complex samples. To calculate
the weight of each sampled member, multiply each member by 1/fraction.
If the sample size was 100 and the population was 100,000 then
the weight of each sampled member would be 1,000. This means that
any sampled member's response is taken to represent 1,000 identical
responses in the population.
All surveys based on
stratified and/or cluster samples will supply subject-specific
weights for cross-sectional analysis of one year of data. In MEPS,
when the household is the unit of analysis the researcher will
need to use the family level estimation weights. Likewise, if
the unit of analysis is the individual, person-level estimation
weights are selected. Surveys, such as MCBS and MEPS, that employ
rotating panel surveys will supply separate sets of longitudinal
weights for subjects who have been followed for two years or three
years. Longitudinal weights for two years of follow-up cannot
be used for a one-year study or three years of follow-up. Weights
may be stored in separate "weights" files along with
subject identifier, strata and/or cluster identifiers. Weights
and strata variables can be merged, as needed, with files containing
questionnaire responses. Statistical analysis software that is
designed to analyze complex survey data will indicate when to
specify the names of variables for strata and weights that come
from the "weights" file. The survey documentation serves
as the guide for selecting the appropriate weight.
|