Weighting in MLwiN (PDF, 459kB)
An explanation of what weights are and details on how to perform weighted analyses in MLwiN, along with the results of a simulation study that tested the performance of MLwiN's weighting facility.
Downloadable MLwiN demonstration files referred to in Appendix A, page 28, in the above PDF manual:
- weights-demonstration-start.wsz (MLwiN worksheet, 0.1 mb)
- weights-demonstration-completed.wsz (MLwiN worksheet, 0.1 mb)
- weights-demonstration.txtweights-demonstration.txtweights-demonstration.txt (Text file, 0.1 mb)
We received some feedback that suggested MLwiN's weighting facility might not be operating correctly. We therefore carried out a simulation study to check MLwiN's performance in weighted analyses. This investigation showed that MLwiN's weighting facility was performing correctly for the models we considered. (We did not investigate discrete response models, for which we do not recommend performing weighted analyses in MLwiN, and it is possible that the weighting facility may not be operating as expected for these models). This document also contains an explanation of what weights are and details on how to perform weighted analyses in MLwiN.
Example question: I seem to remember that the facility to specify sampling
weights (via the Weights option from the Model menu or the WEIGhts
command) was introduced as an experimental feature. Is this still the
case, or has thorough testing now been completed?
We have now finished our testing, which showed that weighted analyses were giving correct results, and so the weights facility is no longer regarded as experimental. (Results are available here: Weighting in MLwiN (PDF, 457kB). Please note that MLwiN does not offer the facility to use weights when using MCMC estimation: any weights specified will by ignored by MCMC. For this reason we do not recommend the use of weights in discrete response models (since we recommend MCMC estimation for these models), and we have not checked that the the weights facility works correctly for discrete response models run using IGLS.
More about differential weightings
Example question: I am hoping to use MLwiN to estimate a multiple membership model. However, I'm a little confused. My factor that is the multiple membership factor is "rater". So "rater" is analogous to secondary school in the example in the Mlwin manual. I'm trying to use the "WTCOl" command but it tells me that "categories in the ID column must run from 1..N with no gaps. Use MLREcode to create consecutive codes". However, this does not make sense for multiple membership models as there are several ID columns which would all need recoding together.
We strongly suggest that you estimate cross-classified and multiple membership models using MCMC estimation within MLwiN rather than IGLS. For full details see the MCMC manual. The manual details how the data needs to be set up and how to choose the weights etc