mi estimate

[schneemann]

Dear all,

I am currently trying to analyse data with plausible values as dependent variables (PISA) with runmlwin and Stata's mi estimate integrating your suggestions in this thread: http://www.cmm.bristol.ac.uk/forum/view ... ?f=3&t=918

As the mi estimate command requires original variables before imputation and PISA as well as TIMMS only provide already imputed plausible values, I created a variable that only has missing values:

Code: Select all

egen pv = .

mi import wide, imputed(pv = pv1math pv2math pv3math pv4math pv5math) 

So that my estimation command becomes e.g.:

Code: Select all

mi estimate, cmdok noisily: runmlwin pv cons ///
escs immig ictattpos intmat instmot, ///
level3(country: cons) level2(school: cons, weightvar(w_fschwt)) level1(student: cons, weightvar(w_fstuwt)) nopause

I would like to ask you three things regarding this approach:

1. Do you see any major issues with the first step of creating pv? I am afraid, I might overlook a problem here, but from my understanding this variable does not exist as the values are missing by design?

2. I had some problems with identifying how the requested weights will be included in MLwiN by weightvar (raw or PWIGLS Step A): If I drop the nopause statement and look in MLwiN, it seems they are used as raw weights, but if I simply include a small mistake in the specification, the error messages imply standardised weights, e.g.:

Standardised weights have been requested for level 2, this requires non-equal standardised weights for all levels below 2. Please define weights for these levels


3. Do you think one can trust the resulting p-values using the cmdok command, as runmlwin is not specifically supported by mi estimate? I also understand that the combination of p-values and fit indices such as deviance is quite problematic with multiple imputation, but I am somewhat unsure whether to build the model simply using the plausible values separately so that I can use lrtest?


I am using the 64-bit version of MLwiN and I would really appreciate any ideas regarding this.

Best regards and thank you for your consideration!
David

[GeorgeLeckie]

Dear David,

In terms of (1) this is a more a question about how Stata's -mi- suite of commands work rather than -runmlwin-, but as far as I am aware yes, this looks fine

In terms of (2) I have taken a look and I believe -runmlwin- is working in so far as being able to specify the different weighting options in MLwiN. The way for you to double check is to fit the model twice, once manually in MLwiN via the point-and-click GUI and one via runmlwin specifying the equivalent weighting options. You should obtain the same results. However, I should point out that I am not an expert on MLwiN's treatment of weights, so any further questions about weights would be best placed on the MLwiN forum.

In terms of (3) I believe that the -mi estimate- prefix simply applies Rubin's rules to the provided model results. If -mixed- is supported by -mi estimate- then I would have thought that it would be reasonable to use it with -runmlwin- since -runmlwin- and -mixed- results should coincide for three-level random-intercept models such as yours.

Best wishes

George