Weird real parameter values

[j.harv3y]

Hi,

I think I've done something very wrong in my MSORD- some of the real parameters are tiny (4.884644e-15)!
I would really appreciate any feedback as I have no idea whats wrong and why I might get such strange values!

Below is the code for the design data formatting alongside the model I ran (with the lowest AIC) and a section of the output (I used mlogit for pent);

rd <- process.data(data = capt.hist,
model = "ORDMS",
time.intervals = time.intervals, strata.labels=c("1","U"))

rd.ddl<-make.design.data(rd,
parameters=list(Psi=list(subtract.stratum=c("1","1"))))

up=as.numeric(row.names(rd.ddl$p[rd.ddl$p$stratum=="U",]))

Psiuu1=as.numeric(row.names(rd.ddl$Psi[rd.ddl$Psi$stratum=="U"&
rd.ddl$Psi$time==1,]))


Output summary for ORDMS model
Name : S(~cohort)Psi(~time)pent(~time)Phi(~time)p(~time)

Npar : 46 (unadjusted=30)
-2lnL: 4289.245
AICc : 4386.32 (unadjusted=4351.3875)

Real Parameter S
Stratum:1
1 2 3 4 5 6
1 0.8092046 0.8092046 0.8092046 0.8092046 8.092046e-01 8.092046e-01
2 0.7606248 0.7606248 0.7606248 7.606248e-01 7.606248e-01
3 0.8193121 0.8193121 8.193121e-01 8.193121e-01
4 0.6492745 6.492745e-01 6.492745e-01
5 4.884644e-15 4.884644e-15
6 1.896823e-14

Real Parameter Psi
Stratum:1 To:U
1 2 3 4 5 6
1 8.570178e-12 8.952168e-12 8.739797e-12 2.975294e-12 9.457114e-12 8.215046e-12
2 8.952168e-12 8.739797e-12 2.975294e-12 9.457114e-12 8.215046e-12
3 8.739797e-12 2.975294e-12 9.457114e-12 8.215046e-12
4 2.975294e-12 9.457114e-12 8.215046e-12
5 9.457114e-12 8.215046e-12
6 8.215046e-12

Thank you so much in advance and really looking forward to finding out where I've gone wrong!

Jess

[jlaake]

Have you tried simpler models? Very small values represent those at boundary of a very large negative value for beta with a login link.

[j.harv3y]

Ah thanks so much for explaining things! That makes sense!
I'm not sure how I could do a simpler model? Am I right in thinking that for a msord I have to run parameters for psi,phi, pent,p and s?

Incredibly greatful for your advice!

[jlaake]

Yes you need a model for each of those parameters but they don't have to be time dependent.

[j.harv3y]

Ah Amazing thanks so much! I was under the impression that a .dot model for a parameter wasn't the best to include in a final model due to being not biologically relevant for some parameters/ is also not great because if you're using it as a structural constraint to allow other parameters to be estimable it means your data is insufficient.
Thanks again,
Jess

[cooch]

Ah Amazing thanks so much! I was under the impression that a .dot model for a parameter wasn't the best to include in a final model due to being not biologically relevant for some parameters/ is also not great because if you're using it as a structural constraint to allow other parameters to be estimable it means your data is insufficient.
Thanks again,
Jess

Yes, which is more or less the same point I made when I answered a version of this question you posted in the 'MARK' subforum:

viewtopic.php?f=1&t=3780