Immanuel
2010-Nov-25 16:50 UTC
[R] aftreg vs survreg loglogistic aft model (different intercept term)
Hi, I'm estimating a loglogistic aft (accelerated failure time) model, just a simple plain vanilla one (without time dependent covariates), I'm comparing the results that I obtain between aftreg (eha package) and survreg(surv package). If I don't use any covariate the results are identical , if I add covariates all the coefficients are the same until a precision of 10^4 or 10^-5 except for the intercept term(in survreg)=log(scale) (in aftreg) that are similar but with a worse precision of 10^-1 or 10^-2. Any idea of why this happens? Probably the different covariates coefficients have a so big influence on the intercept term? Thank you Immanuel -- View this message in context: http://r.789695.n4.nabble.com/aftreg-vs-survreg-loglogistic-aft-model-different-intercept-term-tp3059250p3059250.html Sent from the R help mailing list archive at Nabble.com.
Immanuel
2010-Nov-26 16:17 UTC
[R] aftreg vs survreg loglogistic aft model (different intercept term)
I add an example , all the variables are mutually excluding dummy variables,
notice the different intercept: 5.627 vs 5.545:
survreg:
Value Std. Error z p
(Intercept) 5.627 0.00887 634.3 0.00e+00
Var1.recR2 -0.108 0.01026 -10.5 1.00e-25
Var1.recR3 -0.490 0.01099 -44.5 0.00e+00
Var1.recR4 -0.542 0.01303 -41.6 0.00e+00
Var1.recR5 -0.891 0.01095 -81.3 0.00e+00
Log(scale) -0.324 0.00350 -92.7 0.00e+00
Scale= 0.723
Log logistic distribution
Loglik(model)= -379503.5 Loglik(intercept only)= -383388.9
Chisq= 7770.76 on 4 degrees of freedom, p= 0
aftreg:
Covariate W.mean Coef Exp(Coef) se(Coef) Wald p
Var1.recR
1 0.253 0 1 (reference)
2 0.330 0.108 1.114 0.010 0.000
3 0.191 0.490 1.632 0.011 0.000
4 0.106 0.542 1.720 0.013 0.000
5 0.120 0.891 2.437 0.011 0.000
log(scale) 5.545 256.029 0.008 0.000
log(shape) 0.324 1.383 0.003 0.000
Max. log. likelihood -379504
--
View this message in context:
http://r.789695.n4.nabble.com/aftreg-vs-survreg-loglogistic-aft-model-different-intercept-term-tp3059250p3060545.html
Sent from the R help mailing list archive at Nabble.com.
Terry Therneau
2010-Nov-29 13:35 UTC
[R] aftreg vs survreg loglogistic aft model (different intercept term)
Survreg maximizes the log-likelihood to a relative tolerance of 1e-9
(?survreg.control). The printout shows -379503.5, to see the rest of
the digits you need something like:
fit <- survreg(....
print(fit$loglik, digits=9)
Aftreg printed even less digits; you would have to do the same with it
to see which routine got closer to maximizing the actual log-likelihood.
That is of survreg showed -37903.5392 and aftreg -37903.6123 then
survreg "wins".
Likley all this means is that the default iteration tolerance is smaller
for one routine than for the other. When you consider that
"significant" changes in a log-likihood are on the order of 3.94/2 =2
units, I do not get very excited by a .08 difference in convergence.
Terry Therneau
---------- begin included message --------------
I add an example , all the variables are mutually excluding dummy
variables,
notice the different intercept: 5.627 vs 5.545:
survreg:
Value Std. Error z p
(Intercept) 5.627 0.00887 634.3 0.00e+00
Var1.recR2 -0.108 0.01026 -10.5 1.00e-25
Var1.recR3 -0.490 0.01099 -44.5 0.00e+00
Var1.recR4 -0.542 0.01303 -41.6 0.00e+00
Var1.recR5 -0.891 0.01095 -81.3 0.00e+00
Log(scale) -0.324 0.00350 -92.7 0.00e+00
Scale= 0.723
Log logistic distribution
Loglik(model)= -379503.5 Loglik(intercept only)= -383388.9
Chisq= 7770.76 on 4 degrees of freedom, p= 0
aftreg:
Covariate W.mean Coef Exp(Coef) se(Coef) Wald p
Var1.recR
1 0.253 0 1 (reference)
2 0.330 0.108 1.114 0.010 0.000
3 0.191 0.490 1.632 0.011 0.000
4 0.106 0.542 1.720 0.013 0.000
5 0.120 0.891 2.437 0.011 0.000
log(scale) 5.545 256.029 0.008 0.000
log(shape) 0.324 1.383 0.003 0.000
Max. log. likelihood -379504
-------- end inclusion -----------