On Statistical Identification of Important Input Variables to Large Codes

The objective of “statistical screening” is to determine the effective input variables that contribute most to the total variation or uncertainty of the output of a complex computer code. This is accomplished by performing a relatively small number of computer runs with the code and performing statistical analyses on the results. There are two fundamentally different classes of statistical methods available, one based on overdetermined, the other on underdetermined, systems of equations. The basic features of both are described and compared, and a number of critical issues are discussed. Reference is made to a computer code system incorporating both methods and applied to numerical and physical problems.