Re: st: Fw: appling a gamma distribution random effects model. SAS to Stata

["Brent McSharry" <brent@focused-light.net>]

Thanks for the suggestion. the entire SAS code is:

proc nlmixed data=LOS_2009;
parms b0=2.6 b1=0.005 b2=-0.7 b3=-0.3 b4=0.3 b5=0.6 b6=1.17 b7=-0.2 b8=-0.2 
b9=-0.1 b10=-.01 b11=0.01 b12=-0.18 b13=0.07 b14=0.2 b16=-0.02 b17=0 b19=0 
b20=0.25 b21=0.1 s2u=1 a=2;
eta=b0 + b1*rs_hr124 + b2*_Ipdxcat_1 + b3*_Ipdxcat_2 + b4*_Ipdxcat_4 + 
b5*_Ipdxcat_5 + b6*pupils + b7*elective + b8*_Irecovbyp_1 + b9*_Irecovbyp_2 
+ b10*agey + b11*_Iiadm_sc_1 + b12*_Iiadm_sc_2 + b13*_Iiadm_sc_3 + 
b14*_Iiadm_sc_4 + b16*sbpa + b17*sbpasq + b19*_Iprev_ad_1 + b20*_Iprev_ad_2 
+ b21*retriev + u;
mu=exp(eta);
b=mu/a;
model los_days ~ gamma(a,b);
random u ~ normal(0,s2u) subject=PCODE out=random;
predict mu out=predout;
run;

PCODE is a variable representing the post (zip) code for each patient, which 
I do not believe is being used as any kind of independant variable, but 
seems to be somehow assisting in the generation of random normally 
distributed numbers.

> It has more of a flavour of fitting a model than of generating random 
> deviates

I am relieved to hear you say that (as a non-mathematician used to applying 
logistic regression models to generate proabilities of dying  - ie a 
relatively simple & logical task), the addition of random numbers to a model 
already calibrated to a large external data set, in order to predict an 
expected outcome (in this case length of stay in days) seems 
counter-intuitive. However random effects modeling is completely outside my 
field of expertise!

Thank you Maarten - I will look up the STATA random number functions & any 
SAS functions with a few more useful keywords.

----- Original Message ----- 
From: "Nick Cox" <n.j.cox@durham.ac.uk>
To: <statalist@hsphsun2.harvard.edu>
Sent: Thursday, December 08, 2011 12:30 AM
Subject: RE: st: Fw: appling a gamma distribution random effects model. SAS 
to Stata


> I suspect that Brent would be better off showing the whole of the SAS code 
> so that someone bilingual (emphatically not me) can translate. It has more 
> of a flavour of fitting a model than of generating random deviates to me, 
> but no amount of speculation can replace expertise.
>
> Nick
> n.j.cox@durham.ac.uk
>
> Maarten Buis
>
> On Wed, Dec 7, 2011 at 10:31 AM, Brent McSharry wrote:
>
> > I have some SAS code to apply a model to patient data, and I would like 
> > to
> > apply this to our local data (the model is externally validated to 
> > predict
> > length of ICU stay in days).
> > Unfortunately I have no idea about SAS syntax, and exclusively use stata.
> >
> > the first part of the SAS code obviously generates a linear coefficient &
> > applies an exp link
> > ie in Stata:
> >
> > scalar s2u = 1
> > scalar a = 2
> > generate linCoef = 0.2 + varA * 0.4 + varB*-0.02 ...
> > generate mu= exp(linCoef)
> > gen b = mu/a
> >
> > and then i am affraid I am lost (particularly as the = signs stop and the
> > tildas ~ begin)
> >
> > model los_days ~ gamma(a,b);
> > random u ~ normal(0,s2u) subject=varY out=random;
> > predict mu out=predout;
> > run;
>
> I have no idea about the syntax in Some Alternative Software either,
> but the tilde is often used for "distributed as", which makes sense in
> your case. I guess you want to create a random variables from that
> validated model. Now you need to figure out where exactly the random
> term enters the model. I guess it is added to the linear predictor,
> but you need to check that in the documentation of that model. You
> also need to check the exact parameterization of the gamma
> distribution used in that model and possibly translate that to the
> parameterization used in -rgamma()-. After that it will be a couple of
> -generate- commands calling the -normal()- and -rgamma()- functions,
> see -help random_number_functions-.

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