Dear list members,
I am using a counting process (Andersen and Gill 1982) to fit a number of
Cox proportional hazards models. I have a number of categorical and
continuous demographic (e.g., sex, age) and behavioural variables (e.g.,
vegetation type, distance from roads) describing each observation. For
those observations, non-mortalities (0s=10049 records) far outweigh
mortalities (1s=83 records). Some models that include variables with few
mortality observations will not converge (S-Plus diagnostic: Warning
messages: Warning in fitter(X, Y, strats, offset, init = init, iter.max =
: Loglik converged before variable 1 ; beta may be infinite.).
Have others witnessed similar instability with grossly unbalanced models?
Is there a "rule of thumb" for adequate sample size when applying a Cox
proportional hazards model? Beside the proportional hazards assumption
and empty cells what are other factors leading to poorly fit or unstable
models?
All suggestions are appreciated.
Chris Johnson
|