Mager P P, Rothe H
Research Group for Pharmacochemistry, University, Leipzig.
Pharmazie. 1990 Oct;45(10):758-64.
Multicollinearity of physicochemical descriptors leads to serious consequences in quantitative structure-activity relationship (QSAR) analysis, such as incorrect estimators and test statistics of regression coefficients of the ordinary least-squares (OLS) model applied usually to QSARs. Beside the diagnosis of the known simple collinearity, principal component regression analysis (PCRA) also allows the diagnosis of various types of multicollinearity. Only if the absolute values of PCRA estimators are order statistics that decrease monotonically, the effects of multicollinearity can be circumvented. Otherwise, obscure phenomena may be observed, such as good data recognition but low predictive model power of a QSAR model.
物理化学描述符的多重共线性在定量构效关系（QSAR）分析中会导致严重后果，例如通常应用于QSAR的普通最小二乘法（OLS）模型的回归系数估计值和检验统计量不正确。除了诊断已知的简单共线性外，主成分回归分析（PCRA）还可以诊断各种类型的多重共线性。只有当PCRA估计值的绝对值是单调递减的顺序统计量时，才能规避多重共线性的影响。否则，可能会观察到模糊现象，例如数据识别良好但QSAR模型的预测模型能力较低。
