vcdExtra
vcdExtra copied to clipboard
friendly
CMHtest throws an uncatchable "Error in solve.default(AVA)"
For large sparse table (many strata) CMHTest will occasionally throw a Error in solve.default(AVA) Not surprising when I inspect the sub-tables as there are lots of zeros. However, I am unable to catch it with a tryCatch wrapped around the CMHtest call. Are there any downsides to replacing the solve call with a call to MASS::ginv as I would normally do for a linear model? (aside from speed issues)
As you say, your result for a large, sparse table with many strata is unsurprising. CMHtest was not designed for the situation of sparse tables, where solve doesn't work. We would not just substitute MASS::ginv without extensive testing.
If this issue is serious for you, you could clone the package, do some testing of an alternative, and file a patch or pull-request when you think you have a workable replacement.
What kind of testing would you suggest? Just a simple drop-in seems to work in the few cases I tested.. UNC-CH was where I learned about GENCAT, PARCAT, CATMOD and the rest of the GSK FARM stuff. Dick Landis was a year or two ahead of me. IIRC generalized inverses were being used that context, since GSK is a linear approximation. Its been 40+ years, though.
This issue is still open, but I'd like to close it. As I said,
If this issue is serious for you, you could clone the package, do some testing of an alternative, and file a patch or pull-request when you think you have a workable replacement.
Excuse me Mike, I thought you had closed it after our last conversation. I cloned it, replaced it with a g-inverse and the problem is done.
Thanks Bill
On Feb 9, 2019, at 12:22 PM, Michael Friendly [email protected] wrote:
This issue is still open, but I'd like to close it. As I said,
If this issue is serious for you, you could clone the package, do some testing of an alternative, and file a patch or pull-request when you think you have a workable replacement.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/friendly/vcdExtra/issues/3#issuecomment-462067209, or mute the thread https://github.com/notifications/unsubscribe-auth/AB_B_sJ1T7z1jkFWk0ooYb419T7lKQeaks5vLxHTgaJpZM4UjTOU.