This note on using survey weights in modelling has been provided as a response to questions that have been asked by ABS Remote Access Data Laboratory (RADL) clients regarding the use of weights in their models. It should be useful for those contemplating the use of survey weights in their models, and who would like to know how this might be accomplished.
Survey weights are most commonly used to produce estimates of aggregates, means, and quantiles for a population. The use of weights for more complex survey analysis is less clearcut. There is currently no one unified view by statisticians on if, and how, weights should be used when undertaking a modelling exercise.
The decision whether to use weights or not could have an effect on point estimates produced from the model, not only in terms of magnitude but possibly also in terms of sign and significance. Standard errors of these estimates will also differ in magnitude depending on whether weights are used. In general, use of survey weights in the model should lead to accurate point estimates. However, the standard errors of these estimates will not tend to be as accurate.
One approach that may help in partially addressing the issue of less accurate standard errors is to first normalize the weights. Normalizing of weights involves multiplying the weight of each person by a constant factor so that they add up to a desired value. For the purposes of survey analysis, it is common to normalize the weights to add up to the sample total. It is suggested that normalized weights be calculated at state level where possible due to the different sampling fractions that are used in each state.
State-level normalized weights can be calculated by multiplying the person weight of an individual within state h by the proportion of people in the state that were in the sample:
Normalization using SAS, SPSS and Stata
Support for normalized weights in modelling varies for each of the different software packages available in the ABS RADL.
In SAS: When using the LOGISTIC or PHREG procedures, users may specify an option /NORMALIZE when using the weight statement. This will cause the weights specified by the weight statement to add to the total sample size. Note that this option does not necessarily produce normalized weights at state-level.
In SPSS: Minimal support for normalized weights.
In Stata: When producing some models, Stata will rescale the specified weights to add to the total sample size.
Useful references on the use of survey weights in modelling
The following references may assist your decision on whether or not to use survey weights for your particular purpose. They have been selected as an authoritative cross-section of discussion on the issue of using weights in models.
Chambers, R.L. and C.J. Skinner (eds.) (2003), Analysis of Survey Data, Chichester: Wiley.
DuMouchel, W.H. and G.J. Duncan, (1983) "Using Sample Survey Weights in Multiple Regression Analyses of Stratified Samples", Journal of the American Statistical Association, Vol. 78, No. 383. (September), pp. 535-543.
Magee, L., Robb, A.L., and J. B. Burbidge, (1998), "On the use of sampling weights when estimating regression models with survey data", Journal of Econometrics, Volume 84, Issue 2 (June), Pages 251-27.1
Pfeffermann, D. (1993), "The Role of Sampling Weights When Modeling Survey Data", International Statistical Review, Vol. 61, No. 2. (August), pp. 317-337.
Pfeffermann, D (1996), "The Use of Sampling Weights for Survey Data Analysis", Statistical Methods in Medical Research, 5, pp. 239-261.
Skinner, C.J., Holt, D., and T.M.F Smith (1989), Analysis of Complex Surveys, Chichester: Wiley.
Winship, C., and L. Radbill (1994), "Sampling Weights and Regression Analysis." Sociological Methods and Research 23(2):230-257.
This page first published 27 September 2007, last updated 4 July 2008