Skip to content

Remove term from formula for predict.lm

8 messages · Werner W., Henrique Dallazuanna, Gabor Grothendieck +3 more

Werner W.
# 
Hi,

probably just a quick question: can I somehow change the formula used with predict? E.g., the regression was run on "y ~ u + v + w" but for the prediction the term v should be removed from the formula contained in the regression object and only "y ~ u + w" be used.

I could use model.matrix etc. to do the predictions but it would be very helpful to know a simpler way.

Thanks so much,
  Werner


__________________________
 verf?gt ?ber einen herausragenden Schutz gegen Massenmails. 
http://mail.yahoo.com
Henrique Dallazuanna
# 
Try this;

mod1 <- lm(y ~ u + v + w, data = d)
update(mod1, . ~ . -v)
On Tue, Jan 19, 2010 at 2:10 PM, Werner W. <pensterfuzzer at yahoo.de> wrote:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Gabor Grothendieck
    

    

      
      #
    

  


  

      
This recomputes the lm but if that is ok then:

mod <- lm(y1 ~ x1 + x2 + x3 + x4, anscombe)

mod$call$formula <- update(as.formula(mod$call$formula), ~ . - x1 - x2)
predict(eval(mod$call), list(x3 = 1, x4 = 1))

        
On Tue, Jan 19, 2010 at 11:10 AM, Werner W. <pensterfuzzer at yahoo.de> wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      werner w
    

    

      
      #
    

  


  

      
Thanks Gabor and Henrique!

Sorry for the imprecise question. I want predict() to use the coefficients
estimated by the original regression but to exclude terms from the
prediction formula. If I originally estimated y ~ x1 + x2 and got
coefficients b0, b1, b2, I would like to remove x2 and predict y = b0 +
b1*x1 using the the originally estimated coefficients b0 and b1.

@Gabor: I tried your suggestion but it seems although predict now accepts a
list with fewer variables, the new function is not used so that the
coefficient does not change.

            

      
      
      
    

        
1 
4.684909 
Warning message:
In predict.lm(eval(mod$call), list(x1 = 1, x2 = 1, x3 = 1, x4 = 1)) :
  prediction from a rank-deficient fit may be misleading

            

      
      
      
    

        
1 
4.684909

            

      
      
      
    

        
Many thanks,
  Werner

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Walmes Zeviani
    

    

      
      #
    

  


  

      
Werner,

You could set 0 to that regressors you don't want to consider for
prediction.

da <- expand.grid(A=1:20, B=rnorm(20, 4, 0.2), C=10^seq(1,2,l=20))
da$y <- rnorm(da$A, 0, 0.3)

m0 <- lm(y~A+B+C, data=da)

new <- da
new$C <- 0

predict(m0)[1:5]
predict(m0, newdata=new)[1:5]

At your disposal.
Walmes.

-----
..oooO
..................................................................................................
..(....)... 0ooo...                              Walmes Zeviani
...\..(.....(.....)...     Master in Statistics and Agricultural
Experimentation
....\_)..... )../....       walmeszeviani at hotmail.com, Lavras - MG, Brasil
............
(_/............................................................................................

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      David Winsemius
    

    

      
      #
    

  


  

      
On Jan 19, 2010, at 12:05 PM, werner w wrote:

        

            

      
      
      
    

        
So just set x2 = 0 in your newdata argument.

            

      
      
      
    

        

      
      
    

        
David Winsemius, MD
Heritage Laboratories
West Hartford, CT

  
  


  


      

    

    

      
      

        


  

        

      Gabor Grothendieck
    

    

      
      #
    

  


  

      
If you want to keep the same coefficients but ignore certain ones just
use 0 for the ones you don't want:

mod <- lm(Sepal.Length ~ Petal.Length + Petal.Width, iris)
predict(mod, list(Petal.Length = 3, Petal.Width = 0))


Regarding the problem with the example, the example data was not the
best for showing this because there are co-linear columns in the
anscombe data set.  Using iris we see that it does produce different
answers:

            

      
      
      
    

        
1
4.857262

            

      
      
      
    

        
1
5.53337

        
On Tue, Jan 19, 2010 at 12:05 PM, werner w <pensterfuzzer at yahoo.de> wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      werner w
    

    

      
      #
    

  


  

      
Thanks a lot for the answers! 

Somehow I thought there will be a facility to modify the formula. But
setting unwanted variables to zero will of course work and maybe is simple
enough. Thanks again!