My null model is

proc nlmixed data = temp0;

parms b0=-1 s2u=2;

eta = b0 + u;

p = exp(eta)/(1+exp(eta));

model ovpay ~ binomial(1,p);

random u ~ normal(0,s2u) subject=new_laid;

predict eta out=eta_0;

run;

And I'm adding in my potential variables one at a time to judge which variables are worth keeping in the model, by what they do to the s2u estimate. E.g. this "model 1":

proc nlmixed data = temp0;

parms b0=-1 s2u=2 b1=0;

eta = b0 + b1*density + u;

p = exp(eta)/(1+exp(eta));

model ovpay ~ binomial(1,p);

random u ~ normal(0,s2u) subject=new_laid;

predict eta out=eta_0_1;

run;

For my null model, the s2u is 0.1950. But for "model 1", the s2u estimate is 111 x 10^-14. It seems like a ridiculous decrease, and what does it show?! That the variable isn't appropriate for modelling? That the variable is amazing and should be used because it explains all random variation? Thing is, it seems to happen for around half of the variables that I am considering, so at present I can't really come up with a model. Also, if two variables when singularly added to the null model give appropriate s2u values (e.g. 0.07), often when combined in an attempt at a 2 variable model, they also produce this strangely small value. The odd thing is that it is the same 111 x 10^-14 every time.

So is there something wrong with my model? I'm not really sure what to do about parameter initial values (how do I estimate them/ tell if they are appropriate?) so it could be that, though I've fiddled with them/ left the parm statement out and the "error" still occurs.

Any help at all would be greatly appreciated!!

Thanks