[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Multilevel models with plausible values as dependent variable

From   Austin Nichols <>
Subject   Re: st: Multilevel models with plausible values as dependent variable
Date   Fri, 29 May 2009 12:17:09 -0400

Nailing Xia <> :
Instead of using -pv-, first -reshape- the data so that each plausible
value is a separate observation (on a person? are these math test
scores?) and specify that as another level (multiple observations on a
unit at a point in time) in -gllamm- etc. You may also want to weight
by the variance of plausible outcomes, taking account of the greater
uncertainty in some units' plausible values--see e.g.
though this would be more of an issue if you were to take the mean of
the plausible values as your outcome variable rather than using each
as a separate observation (AFAIK you can only have one weight for each
observation in Stata, not separate precision and sampling weights as
in HLM, so I would multiply them all together and treat as pweights if
you take this approach).

You may want to check you get the same answer using my -reshape-
approach in the unweighted case...  I have never used the user-written
-pv- but some experimentation may be required to get matching results.
-pv- seems to take various different approaches to combining
estimates, and I have not looked up its references--maybe you can
provide us some more context on your data.

You did not answer: if the only random effect is the intercept, why
not use fixed effects?  If you are not trying to estimate variability
in heterogeneous coefficients, go with the method that requires fewer
assumptions.  Or do you not have multiple points in time for each

You can't treat your sampling weights as frequency weights to -expand-
for other reasons (inappropriate variance estimates).

On Fri, May 29, 2009 at 11:41 AM, Nailing Xia <> wrote:
> Thanks for the reply. Here are more details on my problem.
> I am estimating multilevel models with random effect intercepts. The
> dependent variable is measured using plausible values, therefore I
> tried to use the user-written package -pv- for estimation as follows:
> pv escs age female grade immg hmlang public city rural clsize scmatedu
> [pw=w_fstuwt], pv(pv*math) cmd("xtmixed") cmdops("|| schid:") brr
> rw(w_fstr*) fays(0.5)
> The error message is "weights not allowed", so I guess -xtmixed- does
> not work with -pv-. I have tried -gllamm- with -pv- as well, and it
> gives the same error message.
> I believe that making the dataset unweighted or using frequency
> weights will not solve my problem either. My main issue is that the
> dependent variable uses plausible values, and hence I have to use -pv-
> for estimation. -pv- requires weights, so I need either find a command
> for multilevel random effects models that allows weights, or find an
> alternative to -pv- that handles plausible values.
> Any thoughts would be very much appreciated!
> Nailing Xia
> On Fri, May 29, 2009 at 10:47 AM, Austin Nichols
> <> wrote:
>> Nailing Xia <> :
>> -gllamm- can handle this case; -findit gllamm- and read a few hundred
>> pages of references...
>> Do you need a multilevel model or can you specify fixed effects
>> (faster and easier, and often more theoretically defensible)?
>> On Fri, May 29, 2009 at 9:50 AM, Nailing Xia <> wrote:
>>> I am estimating a multilevel model using command -xtmixed-. The
>>> dependent variable is measured with plausible values and Stata has the
>>> command -pv- that handles such situation. However, -pv- does not seem
>>> to work with -xtmixed-, it gives me error message "weights not
>>> allowed". Does anyone know why I am getting this error message, or
>>> other Stata command that handles plausible values with multilevel
>>> models? I appreciate your help!
>>> Nailing Xia
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index