Discrepancy between PIM & DM results

[louise.fairless]

Dear all,

I have exhausted the forum and Mark handbook for answers, but cannot resolve the fact that there is a discrepancy between my PIM and DM results.

I have seen that I am not the first to come across this problem: see Discrepancies between age structure model using PIM and DM
by cmb » Wed Jun 15, 2005 2:42 pm for a similar problem. This forum post did not seem to be properly concluded (maybe it was done so offline?) but it still leaves me with questions!!

I have 6 groups in my .inp file with 3 colonies (1,2,3) and 2 sexes (MALE and FEMALE). Groups are listed in the following order,
colony 1 female, colony 1 male, colony 2 female, colony 2 male, colony 3 female, colony 3 male.

I have fitted the general model phi(c*s*t)p(c*s*t) to my data via PIMs and also with the design matrix as below.

I have not included my full design matrix as often requested on this forum as it is far too large! But here is the break down of it:

1 intercept columnn
2 columns for colony (to code the three colonies, with the first group as the reference groups as recommended).
1 column for sex (coding for male and female)
32 columns for time (34 capture events, therefore 34-2 columns in DM).

2 columns for the colony*sex interaction.
32 columns for sex* time interaction.
64(32x2) columns for colony* time interaction
Finally I have 64(32x2) columns for the three way colony*sex*time interaction. Therefore 198 columns each for survival and recapture parameters (396 in total) with 390 estimable parameters.

I want to use the design matrix so that I can go on to run additive models. But, I first checked the results of the DM verses PM for the global model phi(c*s*t) p(c*s*t) as a quality control measure. My results from the PIM and DM are shown below:

.................................................Delta
............................................AICc AICc Weight Likelihood #Par Deviance
{Phi(c*s*t) p(c*s*t) PIM} Sin link 3675.483 623.37 0.00000 0.0000 390.00 1561.019
{phi(c*s*t) p(c*s*t) DM} Logit link 3675.532 623.42 0.00000 0.0000 390.00 1561.068

As you can see, the results are very similar, but not identical!

Am I safe to proceed with further analyses and construct additive models with such a discrepancy in place?

Thank you for taking the time to read this message and I look forward to your thoughts. I understand this message is similar to a previous post mentioned above, but as there was no final conclusion I am therefore stuck in the same situation myself.

Best wishes and thank you in anticipation of your reply,

Louise

[dhewitt]

In the other thread the original poster indicated that some estimates were near boundaries (resighting ~ 1.0). I suspect you have something similar happening. In these cases, which I often run into with analyses of long-lived fish that have survival near 1.0 in some years, that amount of difference between the link functions is not uncommon. It isn't "supposed to happen", but it does. I suspect this will come down to data limitations, estimates near boundaries, or both. I wouldn't be too worried about it as you move forward... you'll need the LOGIT link for additive models.

[louise.fairless]

Thank you very much for your prompt reply. Indeed I do have estimates near the 0,1 boundary which may be the problem, but I am now confident to proceed.

Best wishes,

Louise