I?d appreciate any advice on coxph with a frailty component:
When I try to run the model with ?site? representing the level of shared
frailties, the model fails. When I specify dfs, per T&G (2000), the model
runs. For each site, there is one constant covariate. (The positive stable
frailty macro (EM algorithm) for SAS (Shu and Klein), runs fine but gives
different results than the S+ frailty with dfs specified.) Is the problem
measuring shared frailties with a single constant covariate? How is this
intra-site comparison made? Is bootstrapping the answer?
> names(subjects)
[1] "site" "time" "status" "dummy1" "dummy2" "dummy3"
> model <-coxph(Surv(time,status)~dummy1+dummy2+dummy3+frailty(site),subjects)
Warning messages:
1: Inner loop failed to coverge for iterations 2 3 6 7 in: coxpenal.fit(X, Y,
strats, offset, init = init, control, weights = weights, method = method,
row.names(m), pcols,
pattr)
2: X matrix deemed to be singular; variable 3 in: coxph(Surv(time, status) ~
dummy1 + dummy2 + dummy3 + frailty(site), subjects)
> model2 <-coxph(Surv(time,status)~dummy1+dummy2+dummy3+frailty(site,
df=15),subjects)
> model2 coef se(coef) se2 Chisq DF p
dummy1 0.158 0.146 0.122 1.17 1 0.2800
dummy2 0.335 0.144 0.120 5.40 1 0.0200
dummy3 0.461 0.146 0.121 9.90 1 0.0016
frailty(site, df = 15) 142.96 15 0.0000
Iterations: 4 outer, 14 Newton-Raphson
Variance of random effect= 0.0369 I-likelihood = -3369.1
Degrees of freedom for terms= 0.7 0.7 0.7 15.0
Likelihood ratio test=263 on 17.08 df, p=0 n= 720
> model3 <-coxph(Surv(time,status)~dummy1+dummy2+dummy3+frailty(site,
df=3),subjects)
> model3 coef se(coef) se2 Chisq DF p
dummy1 0.153 0.125 0.121 1.49 1 2.2e-001
dummy2 0.334 0.124 0.119 7.28 1 7.0e-003
dummy3 0.392 0.124 0.120 9.90 1 1.7e-003
frailty(site, df = 3) 26.03 3 9.4e-006
Iterations: 3 outer, 8 Newton-Raphson
Variance of random effect= 0.00601 I-likelihood = -3417.1
Degrees of freedom for terms= 0.9 0.9 0.9 3.0
Likelihood ratio test=63.6 on 5.79 df, p=6.39e-012 n= 720
Thanks,
Tarek
|