[R] Survival analysis MLE gives NA or enormous standard errors

[Charles C. Berry]

On Fri, 23 Jul 2010, Christopher David Desjardins wrote:

> Hi,
> I am trying to fit the following model:
>
> sr.reg.s4.nore <- survreg(Surv(age_sym4,sym4), as.factor(lifedxm),
> data=bip.surv)

Next time include a reproducible example. i.e. something we can run.

Now, Google "Hauck Donner Effect" to understand why

 	anova(sr.reg.s4.nore)

is preferred.

Chuck


>
> Where age_sym4 is the age that a subject develops clinical thought
> problems; sym4 is whether they develop clinical thoughts problems (0 or
> 1); and lifedxm is mother's diagnosis: BIPOLAR, MAJOR DEPRESSION, or
> CONTROL.
>
> I am interested in whether or not survival differs by this covariate.
>
> When I run my model, I am getting the following output:
>
>> summary(sr.reg.s4.nore)
>
> Call:
> survreg(formula = Surv(age_sym4, sym4) ~ as.factor(lifedxm),
>    data = bip.surv)
>                           Value Std. Error     z       p
> (Intercept)                4.037      0.455  8.86643
> 0.000000000000000000755
> as.factor(lifedxm)CONTROL 14.844   4707.383  0.00315
> 0.997484052845082791450
> as.factor(lifedxm)MAJOR    0.706      0.447  1.58037
> 0.114022774867277756905
> Log(scale)                -0.290      0.267 -1.08493
> 0.277952437474223823521
>
> Scale= 0.748
>
> Weibull distribution
> Loglik(model)= -76.3   Loglik(intercept only)= -82.6
> 	Chisq= 12.73 on 2 degrees of freedom, p= 0.0017
> Number of Newton-Raphson Iterations: 21
> n=186 (6 observations deleted due to missingness)
>
>
> I am concerned about the p-value of 0.997 and the SE of 4707. I am
> curious if it has to do with the fact that the CONTROL group doesn't
> have a mixed response, meaning that all my subjects do not develop
> clinical levels of thought problems and subsequently 'survive'.
>
>> table(bip.surv$sym4,bip.surv$lifedxm)
>
>    BIPOLAR CONTROL MAJOR
>  0      41      60    78
>  1       7       0     6
>
> Is there some sort of way that I can overcome this? Is my model
> misspecified? Is this better suited to be run as a Bayesian model using
> priors to overcome the lack of a mixed response?
>
> Also, please cc me on an email as I am a digest subscriber.
> Thanks,
> Chris
>
>
> -- 
> Christopher David Desjardins
> PhD student, Quantitative Methods in Education
> MS student, Statistics
> University of Minnesota
> 192 Education Sciences Building
> http://cddesjardins.wordpress.com
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

Charles C. Berry                            (858) 534-2098
                                             Dept of Family/Preventive Medicine
E mailto:cberry at tajo.ucsd.edu	            UC San Diego
http://famprevmed.ucsd.edu/faculty/cberry/  La Jolla, San Diego 92093-0901