[R] Overall p-value from a factor in a coxph fit

[David Winsemius]

Kåre Edvardsen <ked at nilu.no> wrote in
news:1208767578.25737.127.camel at localhost.localdomain: 

> Prof. Paul,  Prof. Frank.
> 
> Thank you very much for helping me out. The Design package did the
> trick.
> 
> Here is how the anova table looks like without using the Design
> package: 
> 
>> anova(Fit1)
> Analysis of Deviance Table
> Cox model: response is Surv(Time, cancer)
> Terms added sequentially (first to last)
> 
>             Df Deviance Resid. Df Resid. Dev
> NULL                           16783     5341.8
> relativ      1      0.0        16782    14995.0
> hormone      3    939.4        16779    14055.6

> 
> As you see, no p-values reported
> 
> Here is how it looks with after implementing Design:
> 
>> anova(Fit1)
>          Wald Statistics       Response: Surv(Time, cancer) 
>  Factor     Chi-Square d.f.    P
>  relativ     6.08       1   0.0137
>  hormone     8.68       3   0.0339

(spacing added to hopefully look correct in monospaced font)

I was a puzzled looking at those two outputs. In the first the change 
in deviance with addition of "hormone" was 939.4 with 3 df. In the 
second the Wald chi-square, which I presume should be similar to the 
difference in deviance, is only 8.68. Were these outputs from 
completely different models or data? Or are there pitfalls in applying 
Design functions to Surv objects created with other packages?

-- 
David Winsemius



> Regards,
> Kare
> 
> 
> On Fri, 2008-04-18 at 11:03 -0500, Frank E Harrell Jr wrote:
> 
>> Paul Johnson wrote:
>> > On Fri, Apr 18, 2008 at 3:06 AM, Kåre Edvardsen <ked at nilu.no>
>> > wrote 
>:
>> >> Hi all.
>> >>
>> >>  If I run the simple regression when x is a categorical variable
>> >>  ( x 
>  <-
>> >>  factor(x) ):
>> >>
>> >>  > MyFit <-coxph( Surv(start, stop, event) ~ x )
>> >>
>> >>  How can I get the overall p-value on x other than for each
>> >>  dummy variable?
>> >>
>> >>  > anova(MyFit)
>> >>
>> >>  does NOT provide that information as previously suggested on
>> >>  the li 
> st.
>> >>
>> > 
>> > It "should" work...  Here's a self contained example showing that
>> > anova does give the desired significance test for an lm model.
>> > 
>> >> y <- rnorm(100)
>> >> x <- gl(5,20)
>> >> mod <- lm(y~x)
>> >> anova(mod)
>> > Analysis of Variance Table
>> > 
>> > Response: y
>> >           Df  Sum Sq Mean Sq F value Pr(>F)
>> > x          4   6.575   1.644  1.5125 0.2047
>> > Residuals 95 103.237   1.087
>> > 
>> > If you provide a similar self contained example leading up to a
>> > coxph 
> ,
>> > I would be glad to investigate your question.  You don't give
>> > enough information for me to tell which version of coxph you are
>> > running, an 
> d
>> > from what  package.
>> > 
>> > Suppose I "guess" that you are using the coxph from the package
>> > "survival". If so, it appears to me there is a bug in that
>> > package at the moment.  The methods anova.coxph and drop1.coxph
>> > did exist at one time, until very recently.  There is a thread in
>> > r-help (which I foun 
> d
>> > by typing "RSiteSearch("anova.coxph")" ) discussing recent
>> > troubles with anova.coxph.
>> > 
>> > http://finzi.psych.upenn.edu/R/Rhelp02a/archive/118481.html
>> > 
>> > As you see from the discussion in that thread, there used to be
>> > an anova method for coxph, and in the version of survival I have
>> > now, there is no such method.  The version I have is  2.34-1,
>> > Date: 
>> >  2008-03-31.
>> > 
>> > Here's what I see after I run "example(coxph)" in order to create
>> > som 
> e
>> > coxph objects, on which I can test the diagnostics:
>> > 
>> >> drop1(test2)
>> > Error in terms.default(terms1) : no terms component
>> >> anova(test2)
>> > Error in UseMethod("anova") : no applicable method for "anova"
>> > 
>> > In that survival package, I do find anova.survreg, but not
>> > anova.coxph. If you are using the survival package, I'd suggest
>> > you contact Thomas Lumley directly, since he maintains it.
>> > 
>> > I think if you had reported the exact error you saw, it would
>> > have been easier for me to diagnose the trouble.
>> > 
>> > HTH
>> > pj
>> > 
>> 
>> In the meantime you can do
>> 
>> library(Design)
>> f <- cph( . . . )
>> anova(f)      # multiple d.f. Wald statistics including tests of 
>> nonlinearity
>> 
>> cph uses coxph but anova.Design is separate from the survival
>> package. 
>> 
>> Frank
>> 
> 
>      [[alternative HTML version deleted]]
> 
> 
> --===============0807199787==
> ______________________________________________
> 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. 
> 
> Attachment decoded: untitled-3.txt
> --===============0807199787==--