Fortran error in RMark ORDMS estimation

[gstauffer]

I see that this topic was posted last year, but I did not see an answer, so am unsure how/if the issue got resolved. http://www.phidot.org/forum/viewtopic.php?f=2&t=1730&p=5098

I am trying to run an Open Robust Design Multi-state model with 4 observable and 3 unobservable states (many transitions fixed to zero) for 27 years (182 total secondary occasions). The eh has 5500 records. I keep getting the following error:

"
Visual Fortran run-time error

forrtl: severe (157): Program Exception – access violation
Image PC Routine Line Source
mark.exe 005635B8 ESTMAT 2579 estmat.for
"

and then in RMark:

"
Warning: output from MARK was not complete
Error in if (x4 > x2) { : argument of length zero

*********Following model failed to run: S.1 – p.1 ***********
"

During the model run, I could see the mark.exe window counting iterations, then begin assembling the v-c matrix. I assume it failed sometime after or during this step (I last saw that the v-c matrix was 31% complete, when I headed out the door). The mark.out file contains the beta estimates and then just ends. The files mark.res and mark.vcv also are created, but the mark.vcv file clearly is incomplete (it is much too small). The mark.rda file is not created, although I specified "external = TRUE".

I also see that others got a similar error about array bounds being exceeded, and they solved it by setting pim.type="time" for Psi.
http://www.phidot.org/forum/viewtopic.php?f=21&t=1833&p=5451
I'm not sure I can do that because I need age structure in my Psi pims.
In any case, I'm not even sure that the size of the analysis is the problem.

Does anyone know how I can prevent this error? Would it help to specify a logit link for Psi? When I tried this with previous ORDMS analyses, I ran into convergence problems.

I am using R 2.14.1, RMark 2.1.1 (downloaded just over a week ago from Jeff's github site), and MARK 6.1 (build 7601) with Windows 7 and 4 GB of RAM.

Not sure where to go from here....
Glenn

[jlaake]

Glenn-

The only person that can address whether this is a size problem in MARK or something else is Gary. I suggest you send him the .inp file and the .out file and ask for his help. If MARK does not complete the analysis, clearly RMark will fail and cannot create the .rda file. I believe the error in MARK will likely be when it is creating the var-cov matrix of the real parameters - not the betas. If that is the case I suggest that you ask Gary to make computation of the real v-c matrix an option that can be turned off. You only other option is to reduce the number of real parameters which you can do by using the time pim or the logit. It does not sound like the time pim is an option for you. I have been requested to make an age pim for this case and will do so when I get time. For the time being you may want to try link=logit and use option="SIMANNEAL" to help with optimization issues. That's about all that I can suggest that will work with RMark at present. The underlying issue is that I cannot simplify the PIM structure due to the way mlogit is handled in MARK. SImplification works fine with logit.

--jeff

[bacollier]

Glenn,
I had hit that error on a fairly regular basis on a system about like yours for a big multi-strata analysis and I had to jump up to a 64bit Windows laptop with a dual quad core w/16g ram (and a slightly smaller Linux box) and it ran fine on both. In my case it was a memory issue.

Bret

[jlaake]

Bret-

Didn't you have to get a 64 bit version of mark.exe? Is that on the website somewhere for Glenn to get or did you get it from Evan?

--jeff

[gwhite]

Send me the input file that the MARK interface generates. I'll see if I can blow it up on my machine.

Gary

[bacollier]

Bret-

Didn't you have to get a 64 bit version of mark.exe? Is that on the website somewhere for Glenn to get or did you get it from Evan?

--jeff

Hmmm, I just checked and I am running the 32 bit version on a 64 bit machine (must have forgot to update to 64 bit MARK when I cloned my old 32 bit machine onto my new 64 bit one), yet oddly my models worked on the 64 bit system but not my previous 32 bit system...maybe I am not capitalizing on the memory I thought I was? I will get the 64 bit version and run some tests to see what is what (that is if my bootstrapping routine I am running ever ends)

Bret

[jlaake]

If you are using the 32 bit version, it could have still been a memory issue that was released with the added memory of that machine. You just didn't need as much as the 64 bit version would have provided.

--jeff

[gstauffer]

Thanks all for comments.
Bret, I've run big ORDMS models before on this machine, but I have now expanded the number of states. Perhaps that was the straw that broke the camel's back! A little while ago I tried to run some ORDMS models on a Linux cluster, but ran into issues where Linux couldn't seem to read some of the output files, so the R output objects weren't created. Never did get that resolved, but it might be worth another try if I can get access to a cluster again.

Jeff, In my call to mark, I didn't change the default realvcv=FALSE. I guess that affects only whether RMark stores the result, not whether MARK computes it?

I forgot to mention in my last post that I did provide what I thought were reasonable starting values, based on a similar analysis I did previously, and some intuition about what values some betas might take relative to others.

Anyway, I am presently trying the model again, with logit links on Psi and with SIMANNEAL instead of initials. I'm also trying a simpler model without time variation in Psi. I expect it will take "awhile" to run!

Glenn

[jlaake]

Jeff, In my call to mark, I didn't change the default realvcv=FALSE. I guess that affects only whether RMark stores the result, not whether MARK computes it?

Correct.

Increasing the number of states increases the number of real parameters due to the increase in the number of possible transitions for Psi. Obviously that pushed it over the edge.

I thought we resolved the binary read issue on Linux machine. I don't think Bret has a problem on his Linux machine.

regards --jeff

[gstauffer]

Jeff,
I did successfully run MS models on Linux, but I can't recall that it ever worked with ORDMS. Perhaps I am remembering incorrectly. I'll try again if I can get access to a fast machine.

Glenn