Execution error for observation 97

[m_hojai82]

Hi there

I wanted to apply proc nlmixed on real data but I got this error:  Execution error for observation 97.

Is there any body here to help me how to solve it?

Best wish 

proc nlmixed data=data qpoints=10 gconv=1e-2;
bounds 0 <=sigma2;
/* Initail Values ---------------- */
parms a0=5.3 a1=1.3
b0=14.3 b1=0.01 b2=0.2 b3=2.5
g1=1.1 g2=-0.1 g3=0.5
sigma2=1
l1=1 l2=2 l3=3 k=0.01;

/* Log Likelihood ---------------- */
if t=0 then do;
lam = exp(g1*GENDER + g2*AGE +g3*CITY + l2*a + l3*b);
loglik=(censored=0)*(log(lam)-lam * time) + (censored=1)*(-lam * time);
end;
else do;
pi = 1/(1 + exp(-(a0 + a1*ADTYPE + a)));
eta = abs(k)**(-2);
mu = b0 + b1*ADTYPE +b2*AGE +b3*GENDER - log(pi)-sigma2*log(k**2)/k-lgamma(eta*(k+k*sigma2))+lgamma(eta) + l1*a + b;
u = sign(k)*(log(Y) - mu)/sigma2;
if Y=0 then loglik=log(1-pi);
else if Y>0 then loglik=log(pi)- lgamma(eta) + eta*log(eta) - log(sigma2) - log(Y) + u*sqrt(eta) - eta*exp(abs(k)*u);
else if Y = . then loglik=0;
end;

model ID ~ general(loglik);
random a b ~ normal([0,0], [1, 0, 1]) subject=ID;
run;

[PaigeMiller]

Typically, this means that the procedure can't perform the math. Example: you can't use the LOG() function on a negative number or a zero value.

[m_hojai82]

Thank you for reply? So where is the problem? In formula or initial parameters or both?
Thank you again.

[PaigeMiller]

You have the data. 

You need to trace through the code using data from observation 97 to see where a problem could happen. I can't do that for you.

[m_hojai82]

Actually I applied this code for log-normal distribution before and I wanted to use generalized gamma distribution instead of LN. As you know LN is a special case of GG when 𝑘 → 0. I applied different amount of K and I got the same error but in different observation. for K near 0 I got this error for observation 97 (Y=0) but for K=500 I got this error for observation 1. I applied LN before on my data so it supposed to get the same result when 𝑘 → 0. anyway thank you for your response. As you said I need to track.

[FreelanceReinh]

Hello @m_hojai82 and welcome to the SAS Support Communities!

---

@m_hojai82 wrote:
So where is the problem? In formula or initial parameters or both?

---

[and earlier:]

...

u = sign(k)*(log(Y) - mu)/sigma2;
if Y=0 then ...

---

---

I guess the term log(Y) in the definition of u is a good candidate since you consider the case Y=0 immediately afterwards.