I'll re-enter the fray. The data set is an example where coxph is incorrect; due to round off error it is treating a 5 column rank 3 matrix as if it were rank 4. This of course results in 0 digits of precision. Immediate fix, for the user, is to add "toler.chol= 1e-10" to their coxph call. It is very likely that they will never ever have to change this value. I will look into changing the default from its current value of .Machine$double.eps^.75. However, I first have to check that this does not break anything else. All "epsilon" constants are a delicate dance between mutiple data sets, and anyone with long experience in numerical anlysis will tell you that it is impossible to find constants that will work for every data set. This is true for linear models, logistic, Cox, ... you name it. In summary: I appreciate the example. I'll add to my list of "nasty" problems. I may be able to fix long term, and maybe not. Changing the constant may break something else. I've given a good short term fix. Terry T.
On 12/18/2013 05:00 AM, r-help-request at r-project.org wrote:
Your comprehension of the issue seem to be entirely wrong. Between r11513 and r11516, some tuning of internal parmeters were done, so the process of finding the rank of a singular matrix no longer converges (within the time/tolerance implicitly specified). There are warnings issued, but then there are misc warnings before and after (and one gets "desensitised" about them). Also the nature of the problem, which is to test for possibility of interactions - or lacking thereof - outcome ~ factor A + factor B + factor A x factor B or just extra terms in "outcome ~ factor A + factor B + ..." as an exploration of auxiliary effects, more often than not extra terms won't make any difference and the matrix involved just isn't the nicest to manipulate; it is in the nature of that kind of exploratory work. Professor Therneau replied that it is possible to get the older convergent behaviour by manual tuning of some of the convergence criteria parameters; I have responded that while that is possible, often one is simultaneously exploring many models with many possible auxiliary effects (and lacking thereof), manual tuning for each is neither feasible nor appropriate; and we sort of left it at that.