[R] covariate selection in cox model (counting process)
mayeul.kauffmann at tiscali.fr
Mon Jul 26 14:15:42 CEST 2004
I am searching for a covariate selection procedure in a cox model
as a counting process.
I use intervals, my formula looks like coxph(Surv(start,stop,status)~
x1+x2+...+cluster(id),robust=T) where id is a country code (I study
occurence of civil wars from 1962 to 1997).
I'd like something not based on p-values, since they have several flaws
I turned to other criteria but all the articles I read seems to apply to
classical formulation of the cox model, not the counting process one (or
they apply to both but I am not aware of this)
I've tried AIC with
and BIC using
>step(cox.fit,k = log(n))
but there seems to be 2 theoretical problems to address:
(1) These values are based on partial loglikelihood ("loglik")
I wonder if this is correct with the cox model formulated as a *counting
process*, with many (consecutive) observations for a given
individual, and then some observation not being independent
Since "the likelihood ratio and score tests assume independence of
observations within a cluster, the Wald and robust score tests
do not" (R warning), and the likelihood ratio being based on loglik, can
use loglik in BIC with some dependent observations?
[I have 170 individuals (namely, countries) for 36 year, some single
countries having up to 140 very short observation intervals, other
the other extreme) only 1 long interval per year. That's because I
artificial covariates measuring the proximity since some
events: exp(-days.since.event/a.chosen.parameter). I splitted every
interval for which these covariates change rapidly (i.e.
when the events are recent) yielding up to 11 intervals a year]
(2) What penalized term to used?
It seems natural to include the number of covariates, k.
What about the number of observations?
I found several definitions:
AIC= -2 loglik(b) + 2.k
Schwartz Bayesian information criteria:
SBIC= -2 loglik(b) + k ln(n)
Frédérique Letué (author of PhD thesis "COX MODEL: ESTIMATION VIA MODEL
SELECTION AND BIVARIATE SHOCK MODEL",
AIC= - loglik(b) + 2.k/n
BIC= - loglik(b) + 2.ln(n).k/n, with other possible values for
"2" in this case (see her thesis p.100, but this section is in French)
All these do not tell *what to take for n*. There are 3 main
a) Taking the number of observations (including censored one) will give
huge n (around 6000 to 8000), which may seem meaningless
since some observations are only a few days long.
With n at the denominator (Letué's criteria), the penalized term would
low that it's like not having it:
(where loglik from summary(cox.fit) range from -155 to -175, dependig on
b) Volinsky & Raftery "propose a revision of the penalty term in BIC so
the penalty is defined in terms of the number of
uncensored events instead of the number of observations." (Volinsky &
Raftery , Bayesian Information Criterion for Censored
Survival Models, June 16, 1999,
This could be computed with
Letué's BIC penalized trerm with 50 events will then be
which will have more effects.
However, adding or removing a country which has data for the 36 years
event (then, it is censored) will not change this BIC.
Thus, it is not suitable to account for missing data that do not reduce
number of event.
I'd like the criteria to take this into account, because all covariates
not have the same missing data.
The question is: When I have the choice with adding a covariate, x10 or
which have different (not nested) set of missing
values, which one is best?
Estimating all subsets of the full model (full model = all covariates)
a dataset containing no missing data for the full model
would be a solution but would more than halve the dataset for many
of the covariates.
I should mention that step(cox.fit) gives a warning and stops:
"Error in step(cox.fit) : number of rows in use has changed: remove
which makes me ask whether the whole procedure is OK with model of
c) "For discrete time event history analysis, the same choice has been
while the total number of exposure time units has also
been used, for consistency with logistic regresion
and Kahn,1993;Raftery Lewisand Aghajanian,1994)"
(Raftery, Bayesian Model Selection in Social Research, 1994,
I am not sure what "exposure time units" mean. But since I could have
logit model with yearly observations [but with many
flaws...], I suggest I could use the number of years (sum of length of
intervals, in year)
This may still be too high.
Since I have datas from 1962 to 1997 (36 years), I have the folowing
of "complete cases"-equivalent:
This seems more resonable and would account for NAs in different models.
However, it might be to high, because some countries are not at risk
the all period: some did not existed because they gain
independence near the end of the period (E.G. ex-USRR countries in arly
1990's) or because they were experiencing an event (new
wars in countries already experiencing a war are not taken into
I may take the *proportion* of available data to time at risk to adjust
this: a country at risk during 1 year and for which
data are available for this entire year will increase n by 1, not by
If the data frame "dataset" contains all countries at risk (including
with NA), assuming id == id[complete.cases(dataset)]
(all countries have at least one complete observation) this will be
But this would be a rather empirical adjustment, maybe with no
And I don't think I can enter this as argument k to step()....
Thank you for having read this. Hope I was not too long.
And thank you a lot for any help, comment, etc.
(sorry for mistakes in English as I'm a non native English speaker)
Univ. Pierre Mendes France
Grenoble - France
More information about the R-help