Example 64.4 Firth’s Correction for Monotone Likelihood

In fitting the Cox regression model by maximizing the partial likelihood, the estimate of an explanatory variable X will be infinite if the value of X at each uncensored failure time is the largest of all the values of X in the risk set at that time (Tsiatis; 1981; Bryson and Johnson; 1981). You can exploit this information to artificially create a data set that has the condition of monotone likelihood for the Cox regression. The following DATA step modifies the Myeloma data in Example 64.1 to create a new explanatory variable, Contrived, which has the value 1 if the observed time is less than or equal to 65 and has the value 0 otherwise. The phenomenon of monotone likelihood will be demonstrated in the new data set Myeloma2.

   data Myeloma2;
      set Myeloma;
      Contrived= (Time <= 65);
      run;

For illustration purposes, consider a Cox model with three explanatory variables, one of which is the variable Contrived. The following statements invoke PROC PHREG to perform the Cox regression. The IPRINT option is specified in the MODEL statement to print the iteration history of the optimization.

   proc phreg data=Myeloma2;
      model Time*Vstatus(0)=LOGbun HGB Contrived / itprint;
      run;

The symptom of monotonity is demonstrated in Output 64.4.1. The log likelihood converges to the value of –136.56 while the coefficient for Contrived diverges. Although the Newton-Raphson optimization process did not fail, it is obvious that convergence is questionable. A close examination of the standard errors in the "Analyis of Maximum Likelihood Estimates" table reveals a very large value for the coefficient of Contrived. This is very typical of a diverged estimate.