Hi,
I'm doing some coxph fits using the survival package. There are a large
number of potential predictors and so I was considering using stepAIC
for model selection. However, in the early stages I'm getting complaints
like the following:
Warning message:
In fitter(X, Y, strats, offset, init, control, weights = weights, :
Loglik converged before variable 1,2 ; beta may be infinite.
The source of the problem seems to be that I have a fairly small number
of data values ~500 and some of the predictor levels only have one or
two data points, which I imagine is breaking the fit.
I have two (related) questions:
1. Can I in general ignore this kind of warning and just pay attention to
predictors which produce significant p values?
2. In the specific case of stepAIC, can I ignore this warning and trust
that stepAIC will remove predictors that aren't useful contributors
to the model.
Thanks in advance for any help.
-Ekr
> Loglik converged before variable 1,2 ; beta may be infinite.> I have two (related) questions: > 1. Can I in general ignore this kind of warning and just pay attention to > predictors which produce significant p values? > 2. In the specific case of stepAIC, can I ignore this warning and trust > that stepAIC will remove predictors that aren't useful contributors > to the model.Usually you can ignore the message, it is mostly for your information. The key things to note are a. When one of the coefficients goes to infinity in a Cox model, the Wald test of significance beta/se(beta) breaks down, and is no longer reliable. The LR test however is still valid. Hence routines like stepAIC are ok. So are predicted values, residuals, etc etc. In fact it is pretty much only the Wald test that needs to be ignored: it is based on a Taylor series that simply doesn't work that far from zero. Oops -- confidence intervals based on the se are also useless. b. The actual value of beta that is reported depends on the convergence criteria for the routine. So this is one case where different Cox model functions can give results that look different. I work in medical research, and view these differences as unimportant: if I were to tell you that your relative risk of death was exp(11) = 59,774 fold greater than your compatriots, would the message be substantially changed for beta of 10 or 12? There is a statistical literature under the heading of "monotone likelihood ratio" that worries about these coefficients and tries to fix them. Much ado about nothing, IMHO. c. For a large beta and a very skewed covariate the message can sometimes be wrong. Beta is finite, just unstable. I might still prefer the LR to the Wald in this case. Spline fits based on the truncated power basis (which Frank Harrell uses) are one way to generate such spurious messages. Frank has argued with me that these messages may be shedding more confusion than illumination. He has a point. Terry Therneau
> -----Original Message----- > From: r-help-bounces at r-project.org [mailto:r-help-bounces at r- > project.org] On Behalf Of Terry Therneau > Sent: Thursday, September 18, 2008 6:49 AM > To: ekr at rtfm.com > Cc: r-help at r-project.org > Subject: Re: [R] Coxph and loglik converged before variable X[snip]> b. The actual value of beta that is reported depends on the > convergence > criteria for the routine. So this is one case where different Cox > model > functions can give results that look different. I work in medical > research, and > view these differences as unimportant: if I were to tell you that your > relative > risk of death was exp(11) = 59,774 fold greater than your compatriots, > would the > message be substantially changed for beta of 10 or 12? > There is a statistical literature under the heading of "monotone > likelihood > ratio" that worries about these coefficients and tries to fix them. > Much ado > about nothing, IMHO.Maybe it is just me, but the last line above just begs to have the following lines from "Much Ado About Nothing" added: DON PEDRO What a pretty thing man is when he goes in his doublet and hose and leaves off his wit! CLAUDIO He is then a giant to an ape; but then is an ape a doctor to such a man. [snip] -- Gregory (Greg) L. Snow Ph.D. Statistical Data Center Intermountain Healthcare greg.snow at imail.org 801.408.8111