Use of survey weights
From ICE Primer: A Tobacco Control Research Methodology Primer
- what are survey weights
- examples
- relationships with inclusion probabilities
- how are survey weights used in estimation
- pitfalls in the use of weights
- misunderstood by some software
- a-weights and p-weights
- pooling data sets with very different weights
- use of survey weights in analysis
- suggest fitting a model both with and without weights
- if parameter estimates differ, should explore alternative models
- difference tells you that the model does not fit part of population
Construction of survey weights
In construction of survey weights, the basic weights are usually taken to be the reciprocals of the inclusion probabilities of the sample members. These weights will then be adjusted in various ways to try to bring them closer to the ideal, that wi should be the number of population members represented by sample member i. The main adjustments, in order, are (i) inflating the weights to correct for non-response; (ii) modifying the weights (as little as possible) to provide calibration or benchmarking to known totals. For example, we could require that if we used the sample to estimate the number of population members who are male aged 20-24, we would get the officially correct population number:
and similarly for other age-sex groups. Typically, weights are calibrated to age-sex group sizes within large geographic areas.
When the weights are highly variable, particularly within a stratum, the estimators constructed from the weights can have quite high variances, and hence high standard deviations. In official health surveys efforts are made to produce a sampling design which is close to self-weighting within strata at the household level. There may still be a high variation in individual weights, requiring the use of various techniques for smoothing the weights to keep variation under control while maintaining calibration.
