[R] Coefficient of determination for generalized linear models

[eric.e.harper at us.abb.com]

   Asking the question usually makes one think about the workarounds J.


   I  reviewed  the  pscl source and found the code for pR2. Sourcing the
   following internal functions allowed me to compute the pseudo R2 measures.


   pR2Work <- function(llh,llhNull,n){

     McFadden <- 1 - llh/llhNull

     G2 <- -2*(llhNull-llh)

     r2ML <- 1 - exp(-G2/n)

     r2ML.max <- 1 - exp(llhNull*2/n)

     r2CU <- r2ML/r2ML.max

     out <- c(llh=llh,

              llhNull=llhNull,

              G2=G2,

              McFadden=McFadden,

              r2ML=r2ML,

              r2CU=r2CU)

     out

   }


   pR2.glm <- function(object,...){

     llh <- logLik(object)

     objectNull <- update(object, ~ 1)

     llhNull <- logLik(objectNull)

     n <- dim(object$model)[1]

     pR2Work(llh,llhNull,n)

   }


   Browse[1]> pR2.glm(fit[[2]])

                 llh           llhNull                G2         McFadden
   r2ML          r2CU

   -1658.5890626  -2874.0411606  2430.9041961     0.4229070     0.9472917
   0.9481926

   Browse[1]>


   Question 3: Is there a way to access the pR2.glm function from the library
   without using the source?

   ______________________________________________________________________

   From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On
   Behalf Of Eric E Harper/USABB/ABB
   Sent: Wednesday, April 09, 2008 10:47 AM
   To: r-help at r-project.org
   Cc: Emanuel Kolb/DECRC/ABB; Hans-Werner Borchers/DECRC/ABB
   Subject: [R] Coefficient of determination for generalized linear models


   Thanks in advance for your kind attention.
   I am using R to fit empirical data to generalized linear models. AIC (Akaike
   information criterion) is a measure of the goodness of fit returned by calls
   to glm(). I would also like to calculate the coefficient of determination
   R2, although there is no consensus about the exact definition for
   generalized models
   ([1]http://en.wikipedia.org/wiki/Coefficient_of_determination).
   I found a package “pscl” with a pR2 function that computes pseudo-R2
   measures for various GLMs. The arguments to the call are a fitted model
   object of class glm, polr, or mulitnom, and then ‘additional arguments to
   be
   passed to or from functions’.
   The example from the documentation works well.
   Browse[1]> data(admit)
   Browse[1]> require(MASS)
   Browse[1]> ## ordered probit model
   Browse[1]> op1 <- polr(score ~ gre.quant + gre.verbal + ap + pt + female,
   + Hess=TRUE,
   + data=admit,
   + method="probit")
   Browse[1]> pR2(op1)
   llh llhNull G2 McFadden r2ML
   r2CU
   -106.5088203 -151.0299826 89.0423245 0.2947836 0.5682989
   0.6032041
   Browse[1]>
   When I try with a glm object rather than polr, I get the following error:
   Browse[1]> class(fit[[2]])
   [1] "glm" "lm"
   Browse[1]> pR2(fit[[2]])
   Error in inherits(x, "data.frame") : object "ds" not found
   The ds object does exist in the environment, but I do not know how to pass
   it into pR2:
   Browse[1]> class(ds)
   [1] "data.frame"
   Browse[1]> pR2(fit[[2]], ds)
   Error in inherits(x, "data.frame") : object "ds" not found
   Browse[1]> pR2
   function (object, ...)
   {
   UseMethod("pR2")
   }
   Browse[1]>
   Question 1: How do I find the complete argument signature for pR2 in order
   to perhaps pass it the ds object?
   Question 2: If pR2 does not work with glm objects (for some unknown reason),
   is there another function I can use to calculate R-squared and adjusted
   R-squared for a generalized linear model?
   Best regards,
   \Eric
   References
   1. http://en.wikipedia.org/wiki/Coefficient_of_determination

   ______________________________________________
   R-help at r-project.org mailing list
   https://stat.ethz.ch/mailman/listinfo/r-help
   PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
   and provide commented, minimal, self-contained, reproducible code.