MARK under Linux

[egc]

As per the MARK download page - there are a number of ways to run MARK under Linux. These methods can be broken down into one of 2 basic approaches: (1) run MARK (in full GUI-based form) under emulation (using either wine, or VMWare - or equivalents on a Mac), or (2) as a standalone application, using RMark as the interface.

Gary and I have recently figured out the kinks of compiling mark.exe under Linux (in either 32- or 64-bit flavors). And, Jeff Laake and I have figured out how to get RMark to compile and play nice with mark.exe (while you can run individual and groups of models through mark.exe using the command line - fairly easily if you know some basic scripting languages - I suspect most folks taking the non-emulation approach to running MARK under Linux will use RMark as the front-end).

Full details on mark.exe, Linux, and RMark, can be found at

http://www.phidot.org/software/mark/rmark/linux/

Why go to this trouble? Well, other than the fact that Windows is increasingly a black hole of an OS from which it gets harder and harder to climb out of (entire world, pers. comm.), two primary reasons are:

- Linux is a very stable operating system, and RMark and mark.exe under Linux is less "crash-prone" than running RMark and MARK under Windows.

- more importantly, some of the inherent efficiencies in Linux make running RMark under Linux faster than under Windows (all other things being equal). Faster is always better...at least on a computer.

- using RMark, mark.exe under Linux scores major geek points, which means nothing to normal people (who are perhaps not overly well represented as a constituency on this forum; EG Cooch pers. obs.).

[dhewitt]

- using RMark, mark.exe under Linux scores major geek points, which means nothing to normal people (who are perhaps not overly well represented as a constituency on this forum; EG Cooch pers. obs.).

Jeff, you gonna take that?

Seriously, Jeff just helped me (a lot) in getting up and running with RMark and I highly recommend it. If you aren't afraid of a little computer code, the pros outweigh the cons for sure.

[jlaake]

While I share Evan's enthusiasm for the Linux machines for speed, I feel the need to temper some of his remarks before everyone races off to get a Linux machine to run RMark/MARK.

1) I've personally not had any problems with RMark/MARK crashing on Windows machine; however, it can be slow for big jobs. I have had the Rgui crash on Windows before but unrelated to RMark. The R version on Linux is not a gui so it is more stable for that reason. From my experience most problems with "crashing" with MARK has been due to its gui interface which is not used with RMark.

2) There is a major limitation of RMark/MARK on Linux. RMark does not duplicate every feature in MARK because some features are only available through the gui interface of MARK which is not contained in the Linux version because it is only mark.exe (the numerical portion of the code). The Linux version is awesome for large analysis or simulations because it executes so much faster because of the machine,operating system, and from the code optimization that Evan/Gary have done.

So if you have access to a Linux machine and have large analyses to do, you might want to consider RMark/MARK as an option. If you find that you still need some of the features in the MARK gui interface, you can port the output files to a Windows machine and go from there.

[cooch]

1) I've personally not had any problems with RMark/MARK crashing on Windows machine; however, it can be slow for big jobs. I have had the Rgui crash on Windows before but unrelated to RMark. The R version on Linux is not a gui so it is more stable for that reason. From my experience most problems with "crashing" with MARK has been due to its gui interface which is not used with RMark.

Actually, the lack of a GUI for R under Linux (as the default configuration - there are some existing packages which are superficially analogous to the Window GUI), the other primary reason for greater stability stems from the fact that (i) R is a memory pig (notoriously so), and (ii) Linux (and *nix in general) does *much* better with memory management than Windows.

2) There is a major limitation of RMark/MARK on Linux. RMark does not duplicate every feature in MARK because some features are only available through the gui interface of MARK which is not contained in the Linux version because it is only mark.exe (the numerical portion of the code). The Linux version is awesome for large analysis or simulations because it executes so much faster because of the machine,operating system, and from the code optimization that Evan/Gary have done.

The documentation that Jeff and Eric have put together goes into all the 'differences' in great detail. Perhaps the biggest structural difference (in my experience) is that RMark uses the logit link only, whereas MARK defaults to the sin, and allows access to the sin, logit, and 4-5 more - each of which has different utility depending on the need (e.g., some proportional hazards approaches can benefit from log links etc). As many MARK users will know, for 'twitchy' data, where one or more parameters are estimated near the boundary, the logit link doesn't have particularly good performance.

So if you have access to a Linux machine and have large analyses to do, you might want to consider RMark/MARK as an option. If you find that you still need some of the features in the MARK gui interface, you can port the output files to a Windows machine and go from there.

Agreed. For the general user, there is really no reason to not user either standard MARK, or RMark and R on a Windows platform. But, for 'enterprise scale' sorts of analyses - I've found dumping things off on a Linux rocket has a slew of advantages.

I'm not personally adopting an advocacy position - Jeff, Gary and others involved in promulgating use of MARK are simply trying to increase the platform base. Once I can lay my hands on a OS/X-based Mac ($$$ donations accepted, as always , I'll do a port for Mac users also.

[jlaake]

I agreed with most of what Evan said, but wanted to clarify one point. When RMark first started out it only supported the logit link but as I expanded it to support other models like closed, robust, multistrata I expanded the links to include

logit, mlogit, log, loglog, cloglog and identity.

Most parameters have a natural link: logit for probability, log for N, mlogit for Psi etc. You can change the link function in the model parameter specification but I've not tested that out to any great deal. I show an example in the updated appendix C.17 (page 87). It will certainly work to change the link, pass it to MARK and get the results, but I've not checked these cases for some of the post-processing (eg covariate.predictions etc).

The lack of a sin link is due to the fact that all models in RMark are specified with a design matrix which excludes its use.

At some point I may include the use of an identity matrix and sin link but I've not worked out all the details.