Skip to content

Obtaining p-values for coefficients from LRM function (package Design) - plaintext

6 messages · David Winsemius, Joris Meys, Frank E Harrell Jr

Joris Meys
# 
Sent this mail in rich text format before. Excuse me for this.

------------------------
Dear all,

I'm using the lrm function from the package "Design", and I want to
extract the p-values from the results of that function. Given an lrm
object constructed as follows :

fit <- lrm(Y~(X1+X2+X3+X4+X5+X6+X7)^2, data=dataset)

I need the p-values for the coefficients printed by calling "fit".

fit$coef (gives a list of only the coefficients)
fit$pval, fit$p, fit$pvalue, fit$p.value,... : nothing works
str(fit) : no hints there
fit[1,4] : gives dimension errors

help files don't seem to give me a function that extracts them. Yet,
they are calculated and printed, based on the Wald statistics. So they
must be reachable.

Anybody knows how?

Thank you in advance
Kind regards
Joris
Frank E Harrell Jr
#
joris meys wrote:
The individual p-values are not very meaningful.  anova(fit) returns a 
matrix of meaningful tests with P-values.

Frank

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      David Winsemius
    

    

      
      #
    

  


  

      
On Dec 13, 2008, at 1:12 PM, joris meys wrote:

        

            

      
      
      
    

        
That link could create a montrous interpretation problem.

            

      
      
      
    

        
If you want to see how Harrell does it, you can work through the code  
that you get from:

print.lrm

The last element in the "stats" list is (1 - pchisq(z^2, 1), 4) )  
where z was defined as

z <- cof/sqrt(vv)

... and those were obtained further up as:

vv <- diag(x$var)
     cof <- x$coef

So you could try seeing if this is satisfying:

vv <- diag(fit$var) ;
cof <- fit$coef ;
z <- cof/sqrt(vv) ;
1 - pchisq(z^2, 1)

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Joris Meys
    

    

      
      #
    

  


  

      
Thanks for the answers.

@David : I am aware of that, but this is far from the last model actually.

@ Frank : I know the Anova procedure gives more relevant p-values, but
the attempt is to order the terms by interaction type from low
significance to high significance, based on their individual
difference from zero (if I'm making any sense here). I use this merely
as a quick guideline for model selection, the Anova I use later on for
model evaluation.

Therefore I would like to substract the p-values, as they're easier to
interprete in that respect than the anova values. Or am I missing
something?

Kind regards
Joris

        
On Sat, Dec 13, 2008 at 8:44 PM, David Winsemius <dwinsemius at comcast.net> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Joris Meys
    

    

      
      #
    

  


  

      
To clarify : I am aware of the interpretation problems. Thank you for
the tip! (it's getting late here...)

        
On Sat, Dec 13, 2008 at 9:56 PM, joris meys <jorismeys at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Frank E Harrell Jr
    

    

      
      #
    

  


  

      
joris meys wrote:

            

      
      
      
    

        
Be sure to use the hierarchy principle, which anova.Design respects.

Beware of doing model selection on the basis of P-values, R-square, 
partial R-square, AIC, BIC, regression coefficients, or Mallows' Cp.

Frank