topic Re: intial values for parms in proc nlmixed in Statistical Procedures

intial values for parms in proc nlmixed

Abbas_Arkawazi — Thu, 24 Jun 2021 07:54:00 GMT

hello everybody can you help me in the problem in how we can chose the intial values for parms in the proc nlmixed? is there specified method can I chose from it the intial values for the parameters

please help me, my distribution and the program is:

proc nlmixed data=WORK.y;
parms b0=1 b1=2 b2=2=1 a=1 b=1 s=2;
m =b0+(b1*age)+(b2*creatin)+(b3*gender2);
y=urea;
if status_u=1 then f=((a*b)*((1-exp(-exp((y-m)/s)))**(a-1))*((1+((exp(exp((y-m)/s))-1)**(a)))**(-(b+1)))) /(s*exp((-a*exp((y-m)/s))-((y-m)/s))) ;
else f= (1+(((1-exp(-exp((y-m)/s)))/(exp(-exp((y-m)/s))))**a))**-b ;
ll=log(f);
model y ~ general(ll);
run;

Re: intial values for parms in proc nlmixed

Rick_SAS — Thu, 24 Jun 2021 10:24:34 GMT

It looks like you have a typo on the PARMS statement?
parms b0=1 b1=2 b2=2=1 a=1 b=1 s=2;

I know of three ways to choose initial parameters:

The brute force method: Use a grid search to find initial parameter values
Fit a reduced model. Sometimes your model is a generalization of a simpler model that has fewer parameters. You can estimate the simpler model and use those estimates as initial guesses for the full model.
Use the method of moments for parameters that are connected to means, standard deviations, etc.

Re: intial values for parms in proc nlmixed

Abbas_Arkawazi — Thu, 24 Jun 2021 11:28:25 GMT

Thank u very very much for your help, really it is a very important methods to chose initial values.