The PHREG Procedure
The Penalized Partial Likelihood Approach for Fitting Frailty Models
Let 
be the vector of random components for the s clusters.
Gamma Frailty Model
Assume each 
has an independent and identically distributed gamma distribution with mean 1 and a common unknown variance 
; that is, is iid 
. The penalty function is
plus a function of . The penalized partial log likelihood is given by
where 
is the log of any of the partial likelihoods in the sections Partial Likelihood Function for the Cox Model and The Multiplicative Hazards Model.
The profile marginal log-likelihood of this shared frailty model (Therneau and Grambsch 2000, pp. 257–258) is
where 
is the number of events in the ith cluster.
The maximization of this approximate likelihood is a doubly iterative process that alternates between the following two steps:
For a provisional value of
, the best linear unbiased predictors (BLUP) of 
and 
are computed by maximizing the penalized partial log-likelihood 
. The marginal likelihood is evaluated. This constitutes the inner loop. A new value of
is obtained by the golden section search based on the marginal likelihood of all the previous iterations. This constitutes the outer loop.
The outer loop is iterated until the bracketing interval of is small.
Lognormal Frailty Model
With each 
having a zero-mean normal distribution and a common variance , the penalty function is
plus a function of . The penalized partial log likelihood is given by
where 
is the log of any of the partial likelihoods in the sections Partial Likelihood Function for the Cox Model and The Multiplicative Hazards Model.
For a given , let 
be the negative Hessian of the penalized partial log likelihood ; that is,
where 
, and 
.
The marginal log likelihood of this shared frailty model is
Using a Laplace approximation to the integral as in Breslow and Clayton (1993), an approximate marginal log likelihood (Ripatti and Palmgren 2000; Therneau and Grambsch 2000) is given by
The maximization of this approximate likelihood is a doubly iterative process that alternates between the following two steps:
For a provisional value of
, PROC PHREG computes the best linear unbiased predictors (BLUP) of and by maximizing the penalized partial log likelihood . This constitutes the inner loop. For
and fixed at the BLUP values, PROC PHREG estimates by maximizing the approximate marginal likelihood 
. This constitutes the outer loop.
The outer loop is iterated until the difference between two successive estimates of is small.
The ML estimate of is
The variance for 
is
The REML estimation of is obtained by replacing 
by 
.
The inverse of the final matrix is used as the variance estimate of 
.
The final BLUP estimates of the random components 
can be displayed by using the SOLUTION option in the RANDOM statement. Also displayed are estimates of the lognormal frailties, which are the exponentiated estimates of the BLUP estimates.
Wald-Type Tests for Penalized Models
Let 
be the negative Hessian of the partial log likelihood 
:
where 
, and 
. Write 
. The Wald-type chi-square statistic for testing 
is
Let be the negative Hessian of the penalized partial log likelihood at the ML estimate ; that is, 
. Let 
. Gray (1992) recommends the following generalized degrees of freedom for the Wald test:
See Therneau and Grambsch (2000, Section 5.8) for a discussion of this Wald-type test.
PROC PHREG uses the label "Adjusted DF" to represent this generalized degrees of freedom in the output.