DIAGNOSTICS STATISTICS IN QSAR

The formal application of a Hansch analysis to a series of 3-quinuclid inyl benzylates (QNBs) led to a 'statistically significant' QSAR equat ion. In contrast, the application of the MASCA model has shown that th e design matrix is unsuitable for each QSAR analysis: one sample membe r is an outlier but not a high-leverage or influential point; another one is an influential point, a high-leverage point and an extra-carrie r point. The regressors of the design matrix are multicollinear withou t predictive model power. The result of such flagged observation and t his type of multicollinearity is a multiple cluster correlation. The Q NB series is a good example for 'sampling artifacts' where no practica lly important but artificial QSARs can be found.