variable selction in MARK

[EricM]

I recently completed my 2 year woodcock nesting study. I didn't get the n that I wanted but that is wildlife research, or so I'm told.

I recorded 7 different variables from each nest location as well as paired random sites. Variables were recorded in a 11.3 diameter area around the nest and included: # invasive stems, # native stems, pH, soil moisture, total stems and percent invasive. The stusy is 2 pronged; nest success and use / availability. Is it necessary to run any univariate logistic regression tests on the variables to weed out any non-significants prior to building the models for MARK?

Thanks,

Eric Miller
Wildlife Biologist
PA Game Commission
homiller@state.pa.us

[bacollier]

Eric,
There have been (too) many papers discussing the differences between using a 'IC' approach for modeling your data and using a 'statistical significance testing' method and whether the 2 should be mixed: I will not attempt to recreate that discussion/argument here.

But, my thoughts on your question are as follows (focused on the nest success side of things).

Assuming you have a well thought out set of variables you collected, I think that the largest mistake most folks make when preparing to conduct analysis is 'not' looking at your data using some kind of quick summary. Personally, I like to use plots to look for obvious relationships between predictor variables, or between predictor variable and the response. I tend to agree with what Harrell(2001: Regression modeling strategies) say's: The only thing worse than looking at your data before analysis is not looking at your data before analysis. Anscombe (1973 I think) showed some compelling reasons for giving your data a look before analysis as well.

So, I would say, yes, you can/should look at some summaries of your data, see if anything is obviously screwy, and go from there. Some may disagree, some may agree, but I would rather develop a set of models which makes sense by taking some time looking at my data and thinking about the biology rather than flood the literature with a table(s) containing a bunch of models which are not 'selected' based on various statistical criterion (stepwise, all subsets, various IC's, etc.).

Bret

[sbonner]

Eric,
Assuming you have a well thought out set of variables you collected...

Eric,

I noticed that some of the variables you have suggested are redundant and this will result in identifiability problems when you try to fit models.

Unless I'm missing something: total stems = # invasive stem + # native stems and percent invasive= 100 * #invasive stem / total stem. The first equality means that any (linear)model including all three of these variables will have unidentifiable parameters. You can only use 2 of them. Which you choose does not matter, because you can always transform from one to the other. The second equality does not theoretically create an unidentifiable model, because percent invasive is not a linear function of #invasive stem and total stem, but I'd be surprised by a model that included both. I would suggest models with total stems (an overall measure of ground cover?) and then either #invasive stems or % invasive stems (as measures of cover type?).

Just some thoughts.

Cheers,

Simon

[EricM]

Thanks gents! Good points from both. Much appreciated.

I ended up using pH, Soil Moisture, Percent Invasive and then dummy coded nests found in Invasive Habitat and Native Habitat as was performed in the mallard nest example in the MARK guide. Taking your comments into consideration, I left out total stems, and native & invasie stem counts.

Now after I run the models and get AIC weights, etc. what test do I need to run to verify the results? Bootstrap GOF? This is the part that I am unfamiliar with.

Thanks!

[bacollier]

Eric,
See 17.8 of the MARKBOOK. After that, see Sturdivant et al. 2007: Studies in Avian biology 34:45-54.

Can you better define (at least for me) what you mean by "verify the results"? Did you have any models that were the 'best' (from a IC perspective?-weight>.80 or so), or did you have a bunch of models that seemed to fall out about the same in rank? Have you model averaged your real parameters to account for model selection uncertainty?

Bret

[EricM]

Thanks for the reply Bret. Sure, for some reason I had thought that there was one test I needed to run on the best fitting models, like the LR test.

To answer the other question, I had 1 model that came in at .24. The next closest was .17 which had a negative effect based on the beta estimate. The rest tapered off from there., I'll do a model average and see what I get.

Thanks!

[EricM]

I ran model averaging for real estimates on all 50 occasions and all came out with the same weight, B and SE plus weighted average and SE. 95% CI for the average was .842 - .996. Variation was 15.87%.

Am I good to go here?

Also, what data do I need to use in order to create a graph in Excel? I can't export them as I get an error.

Thanks!

[bacollier]

Eric,
Sorry for slow reply. Need some more information from you; by same weight, do you mean same estimated daily survival? If so, I this is because MARK is estimating survival at the mean of the covariate value. If you are interested in predicting, then you can use the beta parameter estimates output by MARK and then just build a plot of your predictions for each level of your covariate data using excel, open office, or any stat program.

Do that help? If you are still stuck, post some of your output (beta estimates, model selection table, etc) and we can go from there (or you can contact me offlist: bret@tamu.edu).

Also, normally we could take this offlist so as not to flood folks inboxes, but one point of this board is to archive issues so that others can see it. If we go offlist, we will post a summary later so others can access it.

Bret

[EricM]

Email just sent Bret. Thanks for the offer!