Skip to content

covariance matrix under null

6 messages · Devarajan, Karthik, Lanre Okusanya, Francisco J. Zagmutt +3 more

Original
Lanre Okusanya
# 
Kindly excuse a non-statistician newbie attempting to wrestle with R.

This might be a relatively easy question, but I am trying to perform nls 
regression and plot the fitted function through the data superimposed on 
the raw data. from reading the R-help, Rtips et al, I am only able to do 
that by extracting the parameter values manually and using it to create 
the plot.

Is there an easier way to do this,  (I have ~60 Plots), obtain an r^2,  
and also plot the x axis in the log domain (any attempts I have tried 
have screwed up).

NLS script

fit<- nls(y~-emax*x^h/(ec50^h+x^h),
       data= sample, start=list(emax=4,h=2,ec50=1))

summary(fit)

Thank you all for your help

Lanre Okusanya, Pharm.D.,BCPS
UB/Pfizer Pharmacometrics Fellow
University at Buffalo School of Pharmacy and Pharmaceutical Sciences
237 Cooke Hall
Buffalo, NY 14260
Email: ooo at buffalo.edu
Tel: (716)645-2828 x 275
Fax: (716)645-2886
Francisco J. Zagmutt
# 
Review the code in example(predict.nls)

For the R squared question check on this thread 
http://finzi.psych.upenn.edu/R/Rhelp02a/archive/51678.html

Cheers

Francisco
Brian Ripley
# 
You will need to tell us of _what_ you want the covariance matrix and what 
you mean by the `null hypothesis':  coxph does estimation, not testing.

If you want the covariance matrix of the parameter estimates, see vcov(), 
which has a coxph method.
On Thu, 25 Aug 2005, Devarajan, Karthik wrote:

            

      
      
      
    

        

      
      
    

        
PLEASE do (no HTML mail, useful subject, please)

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Thomas Lumley
    

    

      
      #
    

  


  

      
On Thu, 25 Aug 2005, Devarajan, Karthik wrote:

        

            

      
      
      
    

        
You can evaluate the second derivative of the partial loglikelihood at any 
specified beta with

   vcov(coxph(formula, data,iter=0, start, init=beta)

eg if you want to get score tests.

 	-thomas

  
  


  


      

    

    

      
      

        


  

        

      Ben Rich
    

    

      
      #
    

  


  

      
To get nice looking plots you can use trellis plots from the lattice
package.  First you need:

library(lattice)

Then you can define a custom panel function that will overlay the
fitted curve on top of the data points in a different color (you just
need to do this once; the fit you want plotted is specified as an
argument):

pred.overlay.panel <- function(x, y, fit, ...)
{
   panel.grid()
   panel.xyplot(x, y, ...)
   form <- as.list(sys.call(-2))[[2]]$call$formula
   resp <- deparse(form[[2]])
   covar <- deparse(form[[3]])
   xx <- seq(min(x), max(x), len=101)
   newdat <- data.frame(xx)
   colnames(newdat) <- covar
   panel.superpose(xx, predict(fit, newdata=newdat),
       subscripts=1:length(xx), groups=factor(rep(2, length(xx)),
       levels=1:2), type="l", ...)
}

Finally, you use the custom panel function in a call to xyplot:

xyplot(y ~ x, data=sample, panel=pred.overlay.panel, fit=fit,
scales=list(x=list(log=TRUE)))


Note how you specify that you want the x-axis to be in log-scale with
the scales parameter.

Hope this helps.

Ben

        
On 8/26/05, Lanre Okusanya <ooo at buffalo.edu> wrote: