[R] numerical overflow/underflow in log-linear models

Jarrod Hadfield j.hadfield at sheffield.ac.uk
Mon Nov 6 17:57:54 CET 2006


Hi,

I've written some MCMC algorithms that repeatedly evaluate exp(Xb)  
where X is a design matrix and b a parameter vector.  For certain  
combinations of X and b I get numerical overflow, and would like a  
way of rescaling the columns of X so that the range of exp(Xb) is  
maximised without overflowing.  I can get approximate ranges for the  
parameter vector b using ML (X is fixed) and was wondering whether  
general solutions exist for this problem, and if they had been  
implemented in R.

Cheers,

Jarrod



More information about the R-help mailing list