Multilevel with plausible value

[Luis Velastegui]

Hi everyone. I need to help. I am trying run a multilevel model with five plausible value using the database called Terce-Unesco. But, I have read that when someone has five plausible value is necessesary run five twice the multilevel for each plausible value. Pisa recommend it. I have read that is wrong get the mean of five plausible value or chose one plausible value. My routine is this:

/*MODEL MULTINIVEL NULL with VP Y weight*/
pv, pv(PV1LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv || IDSCHOOL :@w
pv, pv(PV2LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv || IDSCHOOL :@w
pv, pv(PV3LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv || IDSCHOOL :@w
pv, pv(PV4LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv || IDSCHOOL :@w
pv, pv(PV5LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv || IDSCHOOL :@w
/*1. MODEL MULTINIVEL with VP and weight*/
pv, pv(PV1LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv repitel3 || IDSCHOOL :@w
pv, pv(PV2LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv repitel3 || IDSCHOOL :@w
pv, pv(PV3LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv repitel3 || IDSCHOOL :@w
pv, pv(PV4LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv repitel3 || IDSCHOOL :@w
pv, pv(PV5LEN3) weight(wgl) brr rw(w_fstr*) fays(0.5) :xtmixed @pv repitel3 || IDSCHOOL :@w

The literature explains that is neccesary get the mean of each coefficients. I wish know if I am doing the correct to estimate a multilevel model with five plausible value. If you know courses where I can practice multilevel model using plausible value, I will appreciate it.

Regards,
[email protected]

[Stephen Jenkins]

Plausible values are multiply-imputed values, aren't they? In which case, why not read your 5 datasets including the plausible values into Stata and then mi set them, and then follow that with mi estimate: mixed ...

PS Welcome to Statalist, Luis! Please read through the Forum FAQ and note the points about how to post code to the list using "CODE" delimiters. Also, you should state the provenance of user-written programs like pv

[Luis Velastegui]

Thanks Stephen. The five plausible value are grades in an evaluation. I am run the code in Stata 13 and the "pv" command I got from ssc. In the database I have for each student five plausible value (grades in an evaluation).

[wbuchanan]

Stephen was giving you sound advice, but you may not want to take the multilevel approach in your example above. The PISA, TIMSS, and other similar assessments usually include a combination of jackknife replicate weights and/or weights for inverse probability sampling. Fitting the model multiple times won't do too much, since you still need to estimate the variance parameters as a function of with and between imputation estimates. Another reason you may want to not consider a mixed effect model is that the majority of the assessments are designed and implemented in a way that only provides support for population level inferences (e.g., comparing the countries as a whole and not necessarily the sampling strata or clusters therein). In short, svyset your data (takes care of the multilevel stuff if you specify the weighting scheme correctly) and use the MI tools in Stata to take care of combining and the estimates and deriving the variance parameters for you.