st: RE: Re: xtmixed with log-transfered dependent variable: back to non-log on margins and marginsplot

["Sun, Wensheng" <wsun@bcm.edu>]

Hi, Joseph,

Thank you for the suggetion of log link and gllamm. I am new to gllamm. The following is the error message when I was trying gllamm. Please let me know how should I fix that. Thank you very much!

Wensheng

. webuse childweight
(Weight data on Asian children)
. generate byte k = 1
. eq cons: k
. eq age: age
. gllamm weight age c.age##i.girl  || id: age, cov(uns) res(ind) mle variance , i(id) nrf(2) eqs(cons age)  || id: age, cov(uns) res(ind) mle variance family(gaussian) link(log)
>  adapt
factor variables and time-series operators not allowed
r(101);
. gllamm weight age c.age##i.girl  || id: age, cov(uns) res(ind) mle variance , i(id) nrf(2) eqs(cons age)   family(gaussian) link(log) adapt
factor variables and time-series operators not allowed
r(101);


________________________________________
From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] On Behalf Of Joseph Coveney [stajc2@gmail.com]
Sent: Tuesday, March 19, 2013 9:49 PM
To: statalist@hsphsun2.harvard.edu
Subject: st: Re: xtmixed with log-transfered dependent variable: back to non-log on margins and marginsplot

Wensheng Sun wrote:

Hi, I have a question regarding multilevel model when I use log-transformation
on the dependent variable.

. webuse childweight
. gen ln_weight=ln(weight)
. xtmixed ln_weight  c.age##i.girl  || id: age, cov(uns) res(ind) mle variance
. margins girl, at(age=(0 (0.5)2.5))
. marginsplot

Is there a way if I change the above code a bit, I can let margins prediction
and marginsplot show me back log estimation and back log values on the
marginsplot?

--------------------------------------------------------------------------------

I'm not sure how to get back-transformation right with -margins-, but you could
look at the user-written command -predlog- (-findit predlog-) and the
accompanying _Stata Technical Bulletin_ article for inspiration about the
back-transforming the fixed effects (random effects set to zero).

Also, you could fit the model without log-transformation in the first place by
using a generalized linear mixed model with a log link.  Something like:

generate byte k = 1
eq cons: k
eq age: age
gllamm weight age _I*, i(id) nrf(2) eqs(cons age) ///
    family(gaussian) link(log) adapt

The upside is that the response variable is in its untransformed metric and so
there's no need for back-transformation.  The downside is that you'll have to do
manually (using -lincom-) what -margins- and -marginsplot- does for you.

Joseph Coveney

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/