Ill‐Conditioning in Regression Data

Norman R. Draper, Harry Smith · Wiley series in probability and statistics · 1998

Ill-conditioning is undesirable in regression analysis. It usually leads to unreliable estimates of the regression coefficients, which then have large variances and covari-ances. Whether the data are well-conditioned or ill-conditioned depends on the way the data relate to the specific model under consideration. Zero eigenvalues indicate that an exact relationship exists between the columns of X, and small eigenvalues indicate approximate relationships. It is also possible to use regression formulas to systematically transform the X matrix to a form in which all columns are orthogonal. The X matrix in a regression problem must be such that none of the columns can be expressed as a linear combination of the other columns. This effectively implies also that there must be at least as many rows not dependent on other rows, as there are parameters to estimate, or else a dependence will appear in the columns also.

Cited by 9 works. doi:10.1002/9781118625590.ch16

Authors