Predictive Power of Robust Regression Applied to QSAR Studies

Abstract

AbstractRobust regression methods were applied to 56 sets of structure‐activity data collected from the literature. The methods included the Huber estimator, the sine estimator, the biweight estimator, and least‐squares with a cut‐off option. On the basis of the kurtosis (Ku) for the error distribution of ordinary least‐squares (OLS) regression, the data sets were classified into two groups: 26 sets with Ku > 0, i.e., with thick‐tailed error distribution (Group A) and 30 sets with Ku < 0 (Group B). The predictive powers of the robust methods and OLS were compared by means of the leave‐one‐out technique. The robust methods generally produced greater correlation coefficients between observed and predicted activities than OLS for the sets of Group A, but the differences were not significant in the majority of cases, e.g., 17 of the 26 sets by use of the sine estimator. For Group B, the predictive results of the robust methods were comparable or rather inferior to those of OLS. It may be concluded that OLS does as well as robust techniques in most cases, and that the adoption of the complicated robust techniques is not warranted in QSAR studies except under extreme circumstances.