To perform most regression types, the software must perform a true
matrix inversion to obtain results. The condition number provides
a measure of how well-suited a problem is for numerical computations.
A matrix with a low condition number is well-suited for
inversion; a matrix with a high condition number may cause problems
when being inverted.
The software calculates a reciprocal condition number for all
regression prior to proceeding to steps where inverse operations will
be performed. The reciprocal condition is compared to a threshold
value; if the reciprocal condition number falls below the threshold,
the calculation is stopped to avoid any numerical problems, including
possible software failure.
Method
The software uses the LAPACK function DGECON to compute the reciprocal
condition number. Most classes in the software make use of
1.0E-15 as the minimum reciprocal condition number to check for
numerical safety.