[R] Different results from survreg with version 2.6.1 and 2.10.1

[nathalcs at student.matnat.uio.no]

---------------------------- Original Message ----------------------------
Subject: Different results from survreg with version 2.6.1 and 2.10.1
From:    nathalcs at ulrik.uio.no
Date:    Fri, March 19, 2010 16:00
To:      r-help at r-project.org
--------------------------------------------------------------------------

 Dear all

I'm using survreg command in package survival.

sur2=survreg(Surv((levetid-364),status==2)~vekt.ind,control=list(maxiter=50),data=ord.mfr)

When using

R version 2.6.1 (2007-11-26)
i386-pc-mingw32

I get this result

 summary(sur2)

Call:
survreg(formula = Surv((levetid - 364), status == 2) ~ malder +
    faar + gest + vekt.ind, data = ord.mfr, control = list(maxiter = 50))
               Value Std. Error       z        p
(Intercept) 1.49e+01    1.73105  8.5993 8.02e-18
malder      9.22e-02    0.01955  4.7178 2.38e-06
faar        2.78e-02    0.01308  2.1272 3.34e-02
gest        3.79e-04    0.00528  0.0718 9.43e-01
vekt.ind    8.46e-01    0.19575  4.3234 1.54e-05
Log(scale)  6.83e-01    0.03651 18.7086 4.22e-78

Scale= 1.98

Weibull distribution
Loglik(model)= -11108.6   Loglik(intercept only)= -11136.6
        Chisq= 55.87 on 4 degrees of freedom, p= 2.1e-11
Number of Newton-Raphson Iterations: 13
n= 254572

while when using

R version 2.10.1 (2009-12-14)
i386-pc-mingw32

this is what I get

ummary(sur2)

Call:
survreg(formula = Surv((levetid - 364), status == 2) ~ malder +
    faar + gest + vekt.ind, data = ord.mfr, control = list(maxiter = 50))
                Value Std. Error         z p
(Intercept)  -7.95477   2.83e-34 -2.81e+34 0
malder        0.01635   2.61e-36  6.25e+33 0
faar          0.29213   2.60e-36  1.12e+35 0
gest         -0.00275   7.54e-37 -3.64e+33 0
vekt.ind      0.28575   2.96e-35  9.64e+33 0
Log(scale)  -95.89809   5.88e-36 -1.63e+37 0

Scale= 2.25e-42

Weibull distribution
Loglik(model)= -82219.2   Loglik(intercept only)= -11136.6
        Chisq= -142165.2 on 4 degrees of freedom, p= 1
Number of Newton-Raphson Iterations: 2
n= 254572

which is quite different results, with the result with version 2.6.1 being
 the most reasonable.

But if I for instance do a cox-regression on on the same data with the
same covariate with the two different versions I do get the same result,
which is how it should be

R version 2.6.1 (2007-11-26):
print(coxfit2)
Call:
coxph(formula = Surv(levetid, status == 2) ~ malder + faar +
    gest + vekt.ind, data = ord.mfr)


              coef exp(coef) se(coef)      z       p
malder   -0.046320     0.955  0.00972 -4.767 1.9e-06
faar     -0.040658     0.960  0.00639 -6.367 1.9e-10
gest     -0.000211     1.000  0.00267 -0.079 9.4e-01
vekt.ind -0.427916     0.652  0.09765 -4.382 1.2e-05

Likelihood ratio test=102  on 4 df, p=0  n= 254572


R version 2.10.1 (2009-12-14):
 print(coxfit2)
Call:
coxph(formula = Surv(levetid, status == 2) ~ malder + faar +
    gest + vekt.ind, data = ord.mfr)


              coef exp(coef) se(coef)      z       p
malder   -0.046320     0.955  0.00972 -4.767 1.9e-06
faar     -0.040658     0.960  0.00639 -6.367 1.9e-10
gest     -0.000211     1.000  0.00267 -0.079 9.4e-01
vekt.ind -0.427916     0.652  0.09765 -4.382 1.2e-05

Likelihood ratio test=102  on 4 df, p=0  n= 254572

My question is why do this happen with survreg?

Best regards
Nathalie Støer