Convergence Of The Proc NLMIXED procedure

[fatso33]

Dear Friends

I am running a Proc NLmixed procedure but I am getting an output with missing values but the log reports that convergence has been met. I have used parameter estimates from the random effects in R and also used the appropriate optimization techinque. Despite all that I am still receiving an output with missing entries. May you please help me.....I got a thesis deadline soon. I will post the SAS code but  the results are in the document attached to this post...

proc nlmixed data=ELISAcalibration4 TECH=CONGRA COV noad qpoints=10;
parms top=3.465 bottom=0.258 c50=0.1311 slope=1.515 theta=0.5114 var=0.000694 s2b1=0.0036 s2b2=0.0009
s2b3=0.0009 s2b4=0.00025;

ka= top + b1;
ke= bottom + b2;
ki= c50 + b3;
w= slope + b4;
num= ke-ka;
den= 1 + exp(w*(logconc - ki));
expect= ka + (num/den);
model DO~normal(expect,(expect**theta)*var);
random b1 b2 b3 b4 ~ normal([0,0,0,0],[s2b1,0,s2b2,0,0,s2b3,0,0,0,s2b4])
subject= ID;
run;

[Rick_SAS]

So there are no WARNINGs or NOTEs in the log?

I think this model might be overparameterized or you might also have degenerate data. Here are some questions:

1. How many observations in the data?

2. Do you have any missing values?

3. Is the b1 variable constant? The estimate for the stderr of s2b1 is missing, which is not good.

4. Same question for b4. The s2b4 estimate is missing.  

[fatso33]

Hello Rick

Thank you for the response. I was tasked to fit a Nonlinear random effects shown by that model where b1 b2 b3 and b4 are the random effects parameters. Its the model for the 4-parameter logistic curve used in Elisa Calibration. I will answer your questions exactly in their order

1. there are 99 observations and 6 variables

2. No missing values

3. b1 is the random effects parameter for the upper asymptote(top) and s2b1 is the variance estimate for b1

4. b4 is the random effects parameter for the slope and s2b4 is the variance estimate for b4

I hope i have shed enough light....

[Rick_SAS]

I am not experienced with this particular model. Do you have a reference to the literature that states the form of the mixed model? The reason I ask is that you're specifying10 parameters for this "4-parameter logistic curve." 

[fatso33]

Hey Rick

Yes the literature that supports the random effects model makes use of the DRC R package but not SAS.....I have not yet come across any literature for this particular model in SAS

[Rick_SAS]

---

@fatso33 wrote:\

Yes the literature that supports the random effects model makes use of the DRC R package but not SAS.....I have not yet come across any literature for this particular model in SAS

---

Can you provide a link? I'd like to see the statistical formulation. The papers that I found do not use random effects:

Ritz C, Baty F, Streibig JC, Gerhard D (2015): https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0146021

In fact, the formulation in that paper can be solved by using PROC NLIN to perform nonlinear least squares on a 4-parameter model. 

[fatso33]

yes the literature supports a fixed effects model without random effects. but I am supposed to adapt it to a random effects model.

[Rick_SAS]

I encourage you to talk to your thesis advisor about the form of the mixed model. Good luck.

[fatso33]

thanks Rick...

[fatso33]

rick

you are definitely right about the Proc NLIN procedure. I have that paper too. But I have to adapt it to Random effects modelto cater for the 9 plates used...