Model averaging derived parameters in robust design

[bear16]

I used the Robust design with a Huggins estimator and hence want to do a model average of derived parameters (N by group and year). However, when I try it, MARK aborts and closes. Is it not possible to do model averaging under these conditions? Or is there something else I might try to avoid aborting?

[abreton]

I dug into some old Mark Workshop files (from Intermediate Program Mark Workshops -- Fort Collins, CO) and found a robust design example that deployed the Huggins abundance estimator. In my files this was called simply,

ROBUST DESIGN HUGGINS.DBF
ROBUST DESIGN HUGGINS.FPT

I ran the model averaging procedure and the request was completed successfully (no errors). This file contained only one group...I'm still searching for an example file that contains >1 group. Have you tried just model averaging one of your abundance estimates at a time? I wonder if one parameter is causing a problem...

I'll let you know if I find a multi-group example ... whether or not I run into an error.

andre

[bacollier]

I'll let you know if I find a multi-group example ... whether or not I run into an error.

andre

I am not the smartest guy on model averaging on this board so hopefully someone will quash me if I am wrong, but I wonder if one would want to model average across a model set with multiple groups being selected as the best model? Wouldn't model averaging across groups pull you back towards a mean that is irrespective of the group structure?

Bret

[abreton]

Recall that model averaging deploys the model weights to estimate a weighted average of a parameter(s)-- so models with more weight will contribute more to the model averaged estimates than relatively inferior models (lower weights). Take some extreme cases, if your top model had a weight of 0.99 then your model average estimates would be essentially identical to those from the top model. In contrast, if you had five models in your set, and the first two each had a weight of 0.50, then, in this hypothetical case, the model average estimate would be the average of the estimates from the two top models. To answer your question,

Wouldn't model averaging across groups pull you back towards a mean that is irrespective of the group structure?

Depends on the model weights, if the top model is group dependent and the next model is not, and each have fairly high weights, then yes, model averaging will pull you towards the mean of these models (assuming the two top models have the majority of the weight in the set). But it sounds like in your case 'group' has strong support, if that's the case then the weights will reflect this and the "pull" will be marginal.

Does your model set include models with/without mixtures and/or models specified with classical closed capture vs. Huggins abundance estimators? I hadn't considered this before but mixing these and related options might be causing the problem as well. Are all your models huggins with no mixtures? (simplest case)

andre

[cooch]

I used the Robust design with a Huggins estimator and hence want to do a model average of derived parameters (N by group and year). However, when I try it, MARK aborts and closes. Is it not possible to do model averaging under these conditions? Or is there something else I might try to avoid aborting?

No reason in principle. You can prove this to yourself. Simulate 2 groups, RD, Huggins, 9 occasions, 3 primary (3 secondary per primary). Then model average. You'll see you can model average real or derived parameters just fine.

If you can't get it to work, issue is at your end. One option might be to try the non-interactive mode for model-averaging. Occasionally the GUI-based (standard) averaging approach is a bit twitchy. Normally this is an issue for huge (many parameter) problems, but it might also just be inherently more stable than the standard model averaging interface.

[darryl]

I've recently been doing this with the full Huggins estimator (ie with heterogeneity, recapture probabilities, etc) with 10 groups via RMark and haven't had a problem. Although I also had missing data for some groups and discovered that the Huggins estimator wasn't being calculated properly in that case which Gary has since corrected. If you have missing data you may also want to download the update file form Gary's website.

[cooch]

I've recently been doing this with the full Huggins estimator (ie with heterogeneity, recapture probabilities, etc) with 10 groups via RMark and haven't had a problem. Although I also had missing data for some groups and discovered that the Huggins estimator wasn't being calculated properly in that case which Gary has since corrected. If you have missing data you may also want to download the update file form Gary's website.

Or here:

http://www.phidot.org/software/mark/dow ... /setup.exe

As a general strategy, when something isn't working, one quick and dirty thing to try is to update MARK.

[bear16]

Thank you so much for your responses! It is very helpful to know that I should be able to calculate the averages; now I'll work on troubleshooting as you suggested.

I tried re-installing MARK and it still aborts when I try to model average the derived parameters. It works just fine for the real parameters.

4 of the top models include group (75% weight) and those same models include session.

I have 6 possible derived parameters, i.e. 2 groups, 3 primary sessions. I have tried running one at a time and it still crashes.

It is the basic Huggins model with no mixtures.

[abreton]

Bear16 --

If you're comfortable sending me the data files (DBF, FPT) then I'll see if I can trouble shoot on my end...consider sending me the inp file as well. You'll can find my email address at,
http://www.coopunits.org/Colorado/People/Andre_Breton/index.html

andre

[bear16]

Thank you all! Your answers allowed me to solve the issue. One model was causing the problem.