PIMs

questions concerning analysis/theory using program MARK

PIMs

Postby eronje » Thu Oct 18, 2018 12:42 pm

Hi,

I want to make sure I understand my PIMs. I've read through the chapters and much of this forum, but just don't "get it" yet. I have a simple data set: 1 group, 2 primaries with 3 secondaries each. My interest is abundance and only abundance, one estimate for each primary. I have been advised to use the Robust Design method even though we only have 2 primaries (I also have read on this forum how a closed capture analysis on each primary would work as well and be easier, but I digress). Will be doing full-likelihood p and c, and Huggins p and c.

Please see the image of my PIMS and my data file. I have assumed I need to fix S=1 and gamma'' = 0. If I want to run an Mnot model, it looks like I would need to set p=c for each primary, and for Mtb, I could run the model as is?

first few rows of data:
/* 1000 */ 011111 1;
/* 2000 */ 011111 1;
/* 2001 */ 011111 1;
/* 2002 */ 011011 1;
/* 2003 */ 011011 1;
/* 2004 */ 011111 1;

Image
eronje
 
Posts: 16
Joined: Sun May 27, 2018 12:23 pm

Re: PIMs

Postby eronje » Fri Oct 19, 2018 11:44 am

I think I mostly answered my own question...I ended up not fixing any parameters as it either did not change the results or crashed MARK.

The Mnot and Mt models I made are generating reasonable results now but I am a little stumped on the Mb. Should I be setting the last p of each primary to = the last c of each primary?
eronje
 
Posts: 16
Joined: Sun May 27, 2018 12:23 pm

Re: PIMs

Postby cooch » Fri Oct 19, 2018 11:49 am

That would be chapter 14.
cooch
 
Posts: 1628
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University

Re: PIMs

Postby eronje » Mon Oct 22, 2018 4:40 pm

Thank you. I've been doing my best to get my head around ch. 14 and others. Thanks for your books by the way. I've realized my mistake with the above model parameters and have since changed it so that the last 2 p's of each primary are equal.

Follow up question:

If I run the Mtb model multiple times I will get the same parameter estimates for everything but S and gamma'' (which I understand to be unestimable in this 2 primary RD scenario), and for those the SE and CIs change.

I assumed the SE for those 2 parameters will be 0.00 and there will not be an interval. Sometimes, but not always, that happens. I get the following different results. I can't find a difference in the PIM. Why does this occur?

Real Function Parameters of {Mtb}
Parm Estimate Standard Error Lower Upper
-------------------------- -------------- -------------- -------------- --------------
1:S 0.8122509 0.0000000 0.8122509 0.8122509
2:G"0.0950546 0.0000000 0.0950546 0.0950546

Real Function Parameters of {Mtb2}
Parm Estimate Standard Error Lower Upper
-------------------------- -------------- -------------- -------------- --------------
1:S 0.8122518 132.00608 0.2406575E-307 1.0000000
2:G"0.0950557 147.07030 0.5843063E-309 1.0000000

Real Function Parameters of {Mtb3}
Parm Estimate Standard Error Lower Upper
-------------------------- -------------- -------------- -------------- --------------
1:S 0.8122488 231.83443 0.2406527E-307 1.0000000
2:G" 0.0950522 258.29282 0.5842827E-309 1.0000000
eronje
 
Posts: 16
Joined: Sun May 27, 2018 12:23 pm

Re: PIMs

Postby cooch » Mon Oct 22, 2018 9:14 pm

With only 2 primaries, and 3 secondaries per primary, (i) there is no reason to run a constrained RD - do yourself a favor, and run each primary as a separate analysis; (ii) with only 3 secondaries, there are precious few models you can run anyway. Do yourself another favor, and use the pre-defined models (discussed in Chapter 14): M(0), M(t), M(b) (won't much matter if your use full likelihood or Huggins), model average the estimates, and be done with it. With only 3 occasions, those are arguably the only robust closed-population abundance models anyway.
cooch
 
Posts: 1628
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University

Re: PIMs

Postby eronje » Thu Jan 31, 2019 7:03 pm

So I took your advice and am using the Full Like and Huggins closed capture models to estimate abundance for each primary separately. When I run the pre-defined models, the Mb model returns an N estimate in the hundreds of thousands, while the Mt and M0 are more reasonable (approx. 65). The Mtb model misbehaves as well, though not as poorly.
What am I missing?
eronje
 
Posts: 16
Joined: Sun May 27, 2018 12:23 pm

Re: PIMs

Postby cooch » Thu Jan 31, 2019 8:06 pm

eronje wrote:So I took your advice and am using the Full Like and Huggins closed capture models to estimate abundance for each primary separately. When I run the pre-defined models, the Mb model returns an N estimate in the hundreds of thousands, while the Mt and M0 are more reasonable (approx. 65). The Mtb model misbehaves as well, though not as poorly.
What am I missing?


Reality of the constraints of your data is what you're 'missing'.

You have only 3 sampling occasions per primary interval. As such, there are only 3 models you could fit: M(0), M(t) and M(b). Anything else will blow up, because you don't have enough 'degrees of freedom' to do anything more. M(0) and M(t) almost always do reasonably well. M(b) can be twitchy, because of the shape of the likelihood profile (see last couple of pages of Chapter 14). The other problem (especially for M(b)) might be, low p, exacerbated with siginficiant individual heterogeneity. Basically, the worst possible scenario for estimation of abundance is: (1) <5 encounter sessions, (2) low encounter probabilities (p<=0.2), and (3) moderate to high levels of individual heterogoeneity. If you're stuck with (1) -> (3), you're stuck in a bad place. In fact, Gary and I end up recommending no fewer than 5 sampling occasions to even do reasonably well (White & Cooch, JWM, 2016 for the messy details).

About all you do at this point is model average your estimates from M(0) and M(t), and use those.

It is worth remembering: 'The methods are statistical, not magical' (no truth to the rumour I get a stipend from Darryl MacKenzie every time I borrow this quote from him).
cooch
 
Posts: 1628
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University

Re: PIMs

Postby cooch » Thu Jan 31, 2019 8:10 pm

cooch wrote:Do yourself another favor, and use the pre-defined models (discussed in Chapter 14): M(0), M(t), M(b) (won't much matter if your use full likelihood or Huggins), model average the estimates, and be done with it. With only 3 occasions, those are arguably the only robust closed-population abundance models anyway.


Note that in my original suggestion (above) I mentioned only these 3 models: M(0), M(t), and M(b). There was no basis for you running anything else given the limits of your data (I'm guessing you simply pulled up 'pre-defined models', clicked a button, and then watched most of them fail).
cooch
 
Posts: 1628
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University

Re: PIMs

Postby eronje » Thu Jan 31, 2019 9:23 pm

Thank you! Would you recommend averaging the Mt and M0 even if they are not that close (within 2 AIC points?)
eronje
 
Posts: 16
Joined: Sun May 27, 2018 12:23 pm

Re: PIMs

Postby cooch » Fri Feb 01, 2019 9:10 am

eronje wrote:Thank you! Would you recommend averaging the Mt and M0 even if they are not that close (within 2 AIC points?)


Yes. And, the reason the models are 'relatively close to each other in terms of AIC shouldn't suprise you -- with only 3 sampling occasions, there is precious little difference (1 df) between M(0) and M(t).
cooch
 
Posts: 1628
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University


Return to analysis help

Who is online

Users browsing this forum: No registered users and 10 guests