Skip to content

frailty models in survreg() -- survival package (PR#2934)

1 message · Thomas Lumley

#
On Tue, 6 May 2003, Jerome Asselin wrote:

            
Looking at a wider context in the code

   pfun <- function(coef, theta, ndeath) {
        if (theta == 0)
            list(recenter = 0, penalty = 0, flag = TRUE)
        else {
            recenter <- log(mean(exp(coef)))
            coef <- coef - recenter
            nu <- 1/theta
            list(recenter = recenter, first = (exp(coef) - 1) *
                nu, second = exp(coef) * nu, penalty = -sum(coef) *
                nu, flag = FALSE)
        }
    }

so the recentering means the penalty is actually your penalty (2) -- not
surprising, as the code was written by Terry Therneau.
The log density penalty doesn't give maximum likelihood (which you would
get by integrating out the frailties). It gives the joint likelihood of
the data and the random effects.

For the gamma model, as you note, they are equivalent.  I believe that the
current state of knowledge is that the log density penalty generally gives
consistent estimates but is not equivalent to maximum likelihood. However,
I have not actually seen the arguments for consistency and it isn't
obvious to me how the argument would go.


	-thomas