Using probability weights given by dataset

[Luke McGee]

I am currently trying to carry out the Mundlak approach to NELS:88-92 dataset. I was wondering if someone knew how best to use the probability weights the dataset has given in order to make the results less biased and more representative of the total population, as well as deal with the issue of non-response. If I turn the data into 'survey data' I no longer know how to use the Mundlak approach/xthybrid appraoch and so any help with that would also be much appreciated

[Clyde Schechter]

This is a multi-disciplinary international forum. Acronyms and abbreviations should only be used if they would be readily understood by anybody with a college education. I, and I suspect many others here, have no idea what NELS:88-92 refers to.

That said, you are correct that you cannot use weights with -xthybrid-. But -xthybrid- just implements random-effects regression of a model with within and between representations of the predictcors. And you can estimate that same random effects model using -meglm- with an id link and gaussian family. The estimation procedure is different, as is the syntax, but the model is identical.

Not knowing what NELS:88-92 is, I can't advise you about which variables in the data are the appropriate ones to use as weights for what analyses. Usually the survey data comes with documentation that explains that. Or perhaps somebody else on the Forum works with this data and can answer that question quickly for you. Once you have figured out what the appropriate weight variable is, your analysis will look something like this:

appropriate svyset command here
foreach v of varlist predictor* {
    by(grouping_variable), sort: egen between_`v' = mean(`v')
    gen within_`v' = `v' - between_`v'
}

svy: meglm outcome_var between_* within_* || grouping_variable:, link(id) family(gaussian)

Bear in mind that an appropriate svyset command will require setting weights at both the observation and group level, and for correct standard errors must also account for stratification and primary or other level sampling units. Again, the documentation that comes with your survey should provide you with the necessary information to write the -svyset- command.

You will not get the kind of pre-digested easily interpretable output that comes from -xthybrid-, so you will have to think that all through yourself, but it is all there.