[R] Re : Re : Re : Nomogram with multiple interactions (package rms)

[Frank E Harrell Jr]

On 05/23/2010 06:29 AM, Marc Carpentier wrote:
> Thanks for the answer.
> Unfortunately, I'm not yet skilled enough to do such a thing. I had a look on the code and I'll try to understand it, as a good exercise.
> I thought about sending fake fit objects to nomogram() derived from the original one :
> - orignal : f2<- cph(Surv(d.time,death) ~ sex*(rcs(cholesterol,4)+blood.pressure)
> - manually derived :
> * fMale : with coef rcs(cholesterol,4) and blood.pressure form f2, no sex effect
> * fFemale : with "agregated" coef sex:rcs(cholesterol,4) for cholesterol and sex:blood.pressure for BP and an obligatory sex effect.
> But I failed to fool your function. Had to try though...
>
> Marc

Marc,

Although this feature should really be implemented or fixed in 
nomogram(), you can always use ols to predict (with an R^2 of 1.0) the 
linear predictor from predict(cph fit) setting a variable to a constant 
in the newdata argument to predict, and not using that variable to 
predict the linear predictor.  Then you can make a nomogram from the ols 
model.

Frank


>
>
>
>
>
> ----- Message d'origine ----
> De : Frank E Harrell Jr<f.harrell at vanderbilt.edu>
> À : Marc Carpentier<marc.carpentier at ymail.com>
> Cc : r-help-request Mailing List<r-help at r-project.org>
> Envoyé le : Jeu 20 mai 2010, 15h 30min 27s
> Objet : Re: Re : Re : [R] Nomogram with multiple interactions (package rms)
>
> On 05/20/2010 01:42 AM, Marc Carpentier wrote:
>> Thank you for your responses, but I don't think you're right about the doc...
>> I carefully looked at it before posting and ran the examples, looked in Vanderbilt Biostat doc, and just looked again example(nomogram) :
>> 1st example : categorical*continous : two axes for each sex
>> f<- lrm(y ~ lsp(age,50)+sex*rcs(cholesterol,4)+blood.pressure)
>
> Hi Marc,
>
> My apologies; I misread my own example.  This will take some digging
> into the code.  If you have time to do this before I do, code change
> suggestions welcomed.
>
> Frank
>
>>
>>
>> 2nd : continous*continous : one "age" axe for each specified value of cholesterol
>> g<- lrm(y ~ sex + rcs(age,3)*rcs(cholesterol,3))
>>
>> 3rd and 4th : categorical*continous : two axes for each sex (4th with fun)
>> f<- psm(Surv(d.time,death) ~ sex*age, dist='lognormal')
>>
>> 5th : categorical*continous : two axes for each sex (with fun)
>> g<- lrm(Y ~ age+rcs(cholesterol,4)*sex)
>>
>> I'm desperately trying to represent a case of categorical*(continous+continous) :
>> f2<- cph(Surv(d.time,death) ~ sex*(rcs(cholesterol,4)+blood.pressure)
>> The best solution I can think of is to draw one nomogram for each sex :
>> Assuming 'male' is the ref level of sex :
>> 1st nomogram : one axe for rcs(cholesterol,4), one axe for blood.pressure
>> 2nd nomogram : one axe for sex:rcs(cholesterol,4), one axe for sex:blood.pressure, both shifted because of the sex own effect.
>> (I badly draw it in my previous mail)
>> I didn't see any example of this "adjustement" of nomogram to 'male' or 'female'...
>>
>> I hope I gave a clearer explanation and I'm not wrong about this unmentioned case.
>>
>> Marc
>>
>>
>>
>>
>> ----- Message d'origine ----
>> De : Frank E Harrell Jr<f.harrell at vanderbilt.edu>
>> À : Marc Carpentier<marc.carpentier at ymail.com>
>> Cc : r-help-request Mailing List<r-help at r-project.org>
>> Envoyé le : Jeu 20 mai 2010, 0h 55min 32s
>> Objet : Re: Re : [R] Nomogram with multiple interactions (package rms)
>>
>> On 05/19/2010 04:36 PM, Marc Carpentier wrote:
>>> I'm sorry. I don't understand the "omit" solution, and maybe I mislead you with my explanation.
>>>
>>> With the data from the "f" exemple of nomogram() :
>>> Let's declare :
>>> f2<- cph(Surv(d.time,death) ~ sex*(age+blood.pressure))
>>> I guess the best (and maybe the only) way to represent it with a nomogram is to plot two nomograms (I couldn't find better).
>>> Is there a way to have :
>>>
>>> Nomogram1 : "male" :
>>> - points 1-100 ---------------
>>> - age ("men") ---------------
>>> - blood.pressure ("men") ---------------
>>> - linear predictor ---------------
>>>
>>> And nomogram2 : "female" :
>>> - points 1-100 ---------------
>>> - age ("female") ---------------
>>> - blood.pressure ("female") ---------------
>>> - linear predictor ---------------
>>>
>>> As I said I tried and failed (nomogram() still wants me to define
>>> interact=list(...)) with :
>>> plot(nomorgam(f2, adj.to=list(sex="male")) #and "female" for the other one
>>>
>>> Marc
>>
>> I think the documentation tells you how to do this.  But you failed to
>> look at the output from example(nomogram).  In one of the examples two
>> continuous predictors have two axes each, with male and female in close
>> proximity.  Or maybe I'm just missing your point.
>>
>> Frank
>>
>>>
>>>
>>>
>>> ----- Message d'origine ----
>>> De : Frank E Harrell Jr<f.harrell at vanderbilt.edu>
>>> À : Marc Carpentier<marc.carpentier at ymail.com>; r-help-request Mailing List<r-help at r-project.org>
>>> Envoyé le : Mer 19 mai 2010, 22h 28min 51s
>>> Objet : Re: [R] Nomogram with multiple interactions (package rms)
>>>
>>> On 05/19/2010 03:17 PM, Marc Carpentier wrote:
>>>> Dear list, I'm facing the following problem : A cox model with my sex
>>>> variable interacting with several continuous variables :
>>>> cph(S~sex*(x1+x2+x3)) And I'd like to make a nomogram. I know it's a
>>>> bit tricky and one mights argue that nomogram is not a good a
>>>> choice... I could use the parameter
>>>> interact=list(sex=("male","female"),x1=c(a,b,c))... but with rcs or
>>>> pol transformations of x1, x2 and x3, the choice of the
>>>> categorization (a,b,c,...) is arbitrary and the nomogram not so
>>>> useful... Considering that sex is the problem, I thought I could draw
>>>> two nomograms, one for each sex... based on one model. These would be
>>>> great. Do you think it's possible ?
>>>
>>> Yes, you can specify constant predictors not to draw with the omit=
>>> argument.  But try first to draw everything.  Shouldn't you just get 2
>>> axes each for x1 x2 x3?
>>>
>>>>
>>>> Taking the exemple of the help of nomogram() (package "rms") : f<-
>>>> psm(Surv(d.time,death) ~ sex*age, dist=if(.R.)'lognormal' else
>>>> 'gaussian')
>>>
>>> Drop the if(.R.) which was just corrected in the documentation.  Use
>>> dist='lognormal'
>>>
>>> Frank
>>>
>>>>
>>>> Let's add the previously defined blood.pressure effect with an
>>>> interaction with sex too (with cph) : f2<- cph(Surv(d.time,death) ~
>>>> sex*(age+blood.pressure))
>>>>
>>>> I thought of the parameter adt.to : plot(nomorgam(f2,
>>>> adj.to=list(sex="male")) #and "female" for the other one
>>>>
>>>> But nomogram() still wants me to define interact=list(...) Thanks for
>>>> any advice you might have (with adj.to or any alternative...)
>>>>
>>>> Marc Carpentier
>>>>
>>>
>>>
>>
>>
>
>


-- 
Frank E Harrell Jr   Professor and Chairman        School of Medicine
                      Department of Biostatistics   Vanderbilt University