Weighting in MLwiN

Home Forums Default Forum Weighting in MLwiN

Viewing 4 posts - 1 through 4 (of 4 total)
  • Author
  • #5441

    I have a quick query. In my thesis I am touching on multi-level analysis in one particular section and I am using a dataset there that I have used elsewhere in the document. When I previously used the data, I applied the appropriate weights and ran the analysis in SPSS. Now I want to run a simple analysis (one on a continuous response variable, and one on a dichotomous response variable) in MLwiN too yet the FAQ’s on the website say that the weighting function is still experimental but I have not found much practical advice on using it. Does anyone have experience with using the weighting function in MLwiN? If so, what would your advice be? I’d really like to be able to weight the data consistently across the document.
    Many thanks!


    Hello Ms. Sarah
    Let me make it clear on the onset that I am not acquainted with MLwiN but, of course, with SPSS. It is true that weighting function is still experimental and researchers generally give subjective weights on their understanding of importance of variables. Those who insist on objective weights they mostly use “principal components analysis” and use component loadings as objective weights to calculate component scores in the case of a single dominant principal component. In cases where more than one principal components are significant as dimensions of a single “theoretical construct”, investigators in many cases resort to the practice of using eigen values or percentage proportions of variance explained by components as weights and after multiplying components scores of each principal component with these “supposed objective weights” add them together to get a single variable representing the theoretical construct.There are, of course, some simple methods of weighting which may be derived on the basis of concept of “intrarater reliamility” as explained by Schroeder, H. W. (1894) in his paper entitled, “Environmental Perception Rating Scales: A Case for Simple Methods of Analysis”, Environment & Behavior 16: 573-598. The True-Score Model described in the paper may be adapted for your use. In the case of dichotomous response variable, non-parametric phi-test value may be used as correlation in the evaluation/determination of “intrarater reliability” as explained in the paper cited above. Besides there is a depository of scaling methods in books and papers on “psychometric analyses” which can successfully and with great conviction be used in both of your sets of data.
    with regards,
    Mohammad Firoz Khan,
    Department of Geography,
    Jamia Millia Islamia,
    New Delhi,

    Jeremy Miles

    There are two kinds of weights, sampling weights and frequency weights. Usually when people talk about weights, they mean sampling weights.

    Be very, very careful about using (sampling) weights in multilevel models – they are weird, and I don’t think they’ve been completely worked out. That’s what they mean by experimental. Proc mixed and proc glimmix in SAS will take survey weights, and will give the answers you expect in a single level model, but in multilevel models there isn’t consensus.

    You could look at: http://www.statmodel.com/download/asparouhovgmms.pdf but frankly that sort of thing just confuses me.


    Hi Sarah,

    You may find it useful to post your question to the MLwiN User Forum which has just been set up to support MLwiN users.

    Regards, Kaisa

Viewing 4 posts - 1 through 4 (of 4 total)
  • The forum ‘Default Forum’ is closed to new topics and replies.