Creating a 2000 IES-LFS Database in Stata

PROVIDE Project Technical Paper 2005:1

• Frequency weights (fweight): These are integer weights that indicate duplicated
observations. If for a given observation fweight = n it implies that there are n other
identical observations in the population that are represented by this single sampled
observation.

• Sampling weights or inverse probability weights (pweight): These weights denote the
inverse of the probability that a certain observation is selected to enter into the sample.
Sampling weights are typically associated with survey data. Although, as discussed in the
previous section, survey designs are often quite complicated, a simple example shows
how inverse probability weights are related to normal frequency weights. Suppose there
are N = 200 households in the population from which the sample is drawn. If 10
households are selected randomly, the probability of selection is P = 10/N = 0.05. The
inverse of this is 1/P = 20, which is simply a frequency weight. Usually the interpretation
of sampling weights is not as straightforward. The calculation of the probability of
selection is more complicated when the sampling design involves clustering and
stratification.

• Analytic weights (aweight): These weights are inversely proportional to the variance of an
observation. Thus, if the variance of the j^th observation were σ²1w , w_j∙ would represent
the weight attached to that observation. Typically the observations represent averages, and
the weight is the number of elements that gave rise to these averages.

• Importance weights (iweight): This type of weight has no formal statistical definition.
Each observation’s weight indicates the relative ‘importance’ of that observation. Since
they are rarely used with survey data it will not be discussed any further.

The weights provided with the IES 2000 are inverse probability weights and are based on
the Census of 1996. This in itself is problematic since the Census of 2000 (SSA, 2003a)
revealed some biases in the Census of 1996.⁹ A family of commands specifically designed to
handle the complexity of such sample designs exists in Stata (svy-commands). The following
section gives a more detailed overview of survey estimation.

⁹ A number of other sets of weights are also available, although at this stage none of these have been officially
approved by Statistics South Africa, and hence the original household weights of IES 2000 and the
person weights of LFS 2000:2 are used at this stage by the PROVIDE Project.

7

© PROVIDE Project

<<   9   10   11   12   13   14   15   >>

More intriguing information