[R] using pre-calculated coefficients and LP in coxph()?

[David Winsemius]

On Mar 13, 2011, at 2:43 PM, Ravi Varadhan wrote:

> Like David, I too thought that `offset' is the way to do this.  I  
> was actually in the midst of testing the differences between using  
> `offset' and `init' when David's email came.
>
> Here is what I could figure out so far:
>
> 1.  If you want to fix only a subset of regressors, but let others  
> be estimated, then you must use `offset'.  The `init' approach will  
> not work.
>
> 2. Even when all the regressors are fixed (I have to admit that I do  
> not see the point of this, like David said), there seems to be a  
> difference in using `init' and `offset'.

    I want to thank you, Ravi, for taking the next steps beyond my  
speculations. However, I did not mean to imply that I could see no  
point in using an offset with coxph(). I only meant to say that the OP  
had not yet provided a basis for doing so.

    If one were trying to test  a pre-determined classification rule  
against a new or augmented candidate rule, then entering an offset  
term could be very desirable. An example in my domain of interest  
might be to use a set of life-table estimates for the effect of sex  
and age , then  including other covariates, and even including a  
subject_age term to test whether there was a departure from the  
population expectations. I admit that I have not seen worked examples  
using coxph(), but Therneau has offered examples using Poisson models  
with glm() in his publications regarding "expected survival" both in  
Mayo Clinic Technical Reports and in his book with Grambsch, "Modeling  
Survival Data".

> First of all, we cannot interpret or use the standard errors, CIs,  
> abd p-values when iter.max=0.  Secondly, there is major disagreement  
> in the predictions between `offset' and `init' with no iterations.  
> You can run the following code to verify this:
>
> ans1 <- coxph(Surv(time, status) ~ age + ph.karno, data = lung, init  
> = c(0.05, -0.05), iter.max = 0)
> ans2 <- coxph(Surv(time, status) ~ offset(0.05*age) +  
> offset(-0.05*ph.karno), data = lung)
>
> lp1 <- predict(ans1, type="lp")
> lp2 <- predict(ans2, type="lp")
>
> all.equal(lp1, lp2)
>> all.equal(lp1, lp2)
> [1] "Mean relative difference: 1.463598"
>
> The results from `offset' are correct, i.e. lp2 can be readily  
> verified to be equal to 0.05 * (age - ph.karno).  I don't know how  
> lp1 is computed.
>
> Ravi.
> ____________________________________________________________________
>
> Ravi Varadhan, Ph.D.
> Assistant Professor,
> Division of Geriatric Medicine and Gerontology
> School of Medicine
> Johns Hopkins University
>
> Ph. (410) 502-2619
> email: rvaradhan at jhmi.edu
>
>
> ----- Original Message -----
> From: David Winsemius <dwinsemius at comcast.net>
> Date: Sunday, March 13, 2011 2:29 pm
> Subject: Re: [R] using pre-calculated coefficients and LP in coxph()?
> To: Dimitris Rizopoulos <d.rizopoulos at erasmusmc.nl>
> Cc: r-help at r-project.org, Angel Russo <angerusso1980 at gmail.com>
>
>
>> On Mar 13, 2011, at 1:32 PM, Dimitris Rizopoulos wrote:
>>
>>> probably you want to use the 'init' argument and 'iter.max'
>> control-argument of coxph(). For example, for the Lung dataset, we  
>> fix
>> the coefficients of age and ph.karno at 0.05 and -0.05, respectively:
>>>
>>> library(survival)
>>>
>>> coxph(Surv(time, status) ~ age + ph.karno, data = lung,
>>>  init = c(0.05, -0.05), iter.max = 0)
>>
>>>
>>>
>>> I hope it helps.
>>>
>>> Best,
>>> Dimitris
>>>
>>>
>>> On 3/13/2011 6:08 PM, Angel Russo wrote:
>>>> I need to force a coxph() function in R to use a pre-calculated set
>> of beta
>>>> coefficients of a gene signature consisting of xx genes and the  
>>>> gene
>>>> expression is also provided of those xx genes.
>>
>> I would have guessed (and that is all one can do without an example
>> and better description of what the setting and goal might be) that  
>> the
>> use of the offset capablity in coxph might be needed.
>>
>> -- 
>> David.
>>>>
>>>> If I try to use "coxph()" function in R using just the gene
>> expression data
>>>> alone, the beta coefficients and coxph$linear.predictors will
>> change and I
>>>> need to use the pre-calcuated linear predictor not re-computed
>> using coxph()
>>>> function. The reason is I need to compute a quantity that uses as
>> it's input
>>>> the coxph() output but I need this output to be pre-calculated
>>>> beta-coefficients and linear.predictor.
>>>>
>>>> Any one can show me how to do this in R?
>>>>
>>>> Thanks a lot.
>>
>> David Winsemius, MD
>> West Hartford, CT
>>
>> ______________________________________________
>> R-help at r-project.org mailing list
>>
>> PLEASE do read the posting guide
>> and provide commented, minimal, self-contained, reproducible code.

David Winsemius, MD
West Hartford, CT