[R] pseudo-R by hand

[Achim Zeileis]

On Mon, 11 Apr 2011, Sacha Viquerat wrote:

> hello dear list! since we want to do a model analysis and some people would 
> like to see pseudo-R^2 values for different types of glm of a logistic 
> regression, i've decided to write a function that computes either nagelkerkes 
> normed pseudo-R or cox & snells pseudo-R. however, i am not clear as in the 
> decisive step, i need to calculate the log of (maximum likelihood estimates 
> of model divided by mle of null model). i am well aware of the functions 
> stats::mle and stats::logLik as well as of Design::lrm.

You can look at the pR2() function in the "pscl" package which provides 
various flavors of pseudo R-squared for "glm", "multinom", and "polr" 
objects. The idea is to extract the observed log-likelihood using 
logLik(), then update the model to obtain the null model only and call 
logLik() again. From the two log-likelihoods and the associated number of 
observations, the pseudo R-squared are computed using pR2Work(), see
pscl:::pR2.glm and pscl:::pR2Work.

hth,
Z

> however, I'm not sure 
> wheter mle helps me at all and I am uncertain about the logLik call I have 
> implemented:
>
> #cox&snell
> lambda<- -2*log((logLik(null.model)[1]/logLik(model)[1]))
> out<-1-exp(-lambda/n)
>
> #nagelkerke
> lambda<- -2*log( logLik(model)[1]/logLik(null.model)[1] )
> lambda2<- -2*log( logLik(model)[1] )
> out<-(1-exp(-lambda/n))/(1-exp(-lambda2/n))
>
> can anyone help me out?
>
> ______________________________________________
> 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.
>