Example 8.2 Using Multiple Runs of the EM Algorithm

Continuing the example from the section Getting Started: HAPLOTYPE Procedure, suppose you are concerned that the likelihood reached a local and not a global maximum. You can request that PROC HAPLOTYPE use several different sets of initial haplotype frequencies to ensure that you find a global maximum of the likelihood. The following code invokes the EM algorithm with five different sets of initial values, including the set used in the "Getting Started" example:

   proc haplotype data=markers prefix=SNP init=random seed=51220 
                  nstart=5;
      var m1-m8;
   run;
   

The NSTART=5 option requests that the EM algorithm be run three times with randomly generated initial frequencies, including once using the seed 51220 that was previously used, once using uniform initial frequencies, and once using haplotype frequencies given by the product of the allele frequencies. The two tables in Output 8.2.1 are from the run that produced the best log likelihood: