A tool for weighting sample survey data and variance estimation
In order to correct for sampling error and non-response bias, often some kind of adjustment weighting procedure is carried out. For each record in the sample data file an adjustment weight is computed.
The computation of adjustment weights requires auxiliary information. This relates to variables that are measured in the survey, and for which the population totals are known from another source, e.g., a registration.
Bascula is a software package for calculating weights for all units in a sample using auxiliary information. The auxiliary information is gathered in a weighting model, which forms the basis for the weighting procedure. Several weighting methods are supported, most of them based on the general regression estimator. Bascula can also use the computed weights to estimate population totals, means and ratios as well as variances based on Taylor linearization and/or balanced repeated replication (BRR). For the purpose of variance estimation several sampling designs are supported.
Bascula is part of the Blaise System for computer-assisted survey processing. Bascula can be used either as a menu-driven interactive program or as a software component suitable for developing custom weighting/estimation applications.
Main features
Bascula offers the following methods for weighting and estimation:
- Post-stratification
- Ratio estimation
- Linear weighting
- Multiplicative weighting
The first three are based on the general regression estimator whereas the fourth method is also known as raking ratio estimation or iterative proportional fitting.
Further weighting options:
- consistent linear weighting between persons and households
- graphical diagnosis of the weights
- bounding of weights
Two methods of variance estimation are implemented in Bascula. The first is Balanced Repeated Replication (BRR), a resampling method. The second is the Taylor linearization method which is new to Bascula 4. Both methods can produce tables of variance estimates for total, mean and ratio estimates based on the adjustment weights. The sampling designs supported are :
- stratified simple random sampling without replacement
- stratified two-stage simple random sampling without replacement
- stratified multi-stage sampling with replacement in the first stage
System requirements
Windows 95/98, Windows NT 4.0 or higher, a Pentium II or faster processor, and a minimum of 16Mb of system memory (32Mb recommended).
