# [R] coxme: model simplification using LR-test?

Simon Tragust Simon.Tragust at biologie.uni-regensburg.de
Sun Feb 19 16:45:04 CET 2012

```Hi
I'm encountering some problems with coxme

My data:
I'm looking at the survival of animals in an experiment with 3 treatments,
which came from 4 different populations, two of which were infected with a
parasite and two of which were not. I'm interested if infected animals
differe from uninfected ones across treatments.

Factor 1: treatment (3 levels)
Factor 2: infection state (infected/uninfected)
Random effect 1: (population nested within infection state)

modelling this with
m<-coxme(Surv(day,status)~condition*infection+(1|infection/population),data=all)

gives me the following

Cox mixed-effects model fit by maximum likelihood
Data: all
events, n = 476, 720
Iterations= 7 53
NULL Integrated    Fitted
Log-likelihood -2915.527  -2641.427 -2634.182

Chisq   df p    AIC    BIC
Integrated loglik 548.20 7.00 0 534.20 505.04
Penalized loglik 562.69 6.96 0 548.78 519.80

Model:  Surv(day, status) ~ condition * infestation + (1 | infestation/population1)
Fixed coefficients
coef  exp(coef)  se(coef)     z      p
conditionstarved                            3.3960657 29.8464431 0.3228277 10.52 0.0000
conditionwater                              3.3277968 27.8768547 0.3224368 10.32 0.0000
infestationinfestationyes                   1.5596539  4.7571747 0.7254405  2.15 0.0320
conditionstarved:infestationinfestationyes -1.1100987  0.3295264 0.3712690 -2.99 0.0028
conditionwater:infestationinfestationyes   -0.9150922  0.4004797 0.3709914 -2.47 0.0140

Random effects
Group                   Variable    Std Dev      Variance
infestation/population1 (Intercept) 0.6367618042 0.4054655953
infestation             (Intercept) 0.0199767654 0.0003990712

To assess if the interaction is needed I would normally do a model simplification

m1<-update(m,~.-condition:infection)

however, this gives me

error in formula.default(object, env = baseenv()) : invalid formula

I do not encounter this problem without a random effect in coxph. So my question is

(1)Is it not possible to do model simplification with coxme?
(2)Is there another way to assess an overall significant interaction with coxme?