A reliable QSAR model for estimation of reactivities of radical addition reactions has been developed. The logarithm of the rate constant and the activation energy (reaction barrier) have been chosen to characterize the reactivity. Carbon-centered radicals of very different features and vinyl-type alkenes with diverse substituents served as reactants. The following chemometric methods were applied: multiple linear regression (MLR), principal component regression (PCR), locally weighted regression (LWR) and artificial neural networks (ANN). Principal component analysis of descriptor variables has shown which variables carry the same or similar information. We compared the different chemometric methods from the point of view of predictive power. The root mean square error of prediction (RMSEP) and the per cent relative prediction error (% Rel. RMSEP) were chosen to discriminate between methods. An RMSEP of 0⋅3 for logarithm of rate constant (approximately a factor of two in rate constants) can be achieved using properly built models. Activation energy can be predicted to within 3⋅3-3⋅6 kJ mol 71 , which is smaller than the error level in determining the activation energy of individual radical-alkene reactions. The 'linear methods' MLR and PCR failed as expected. They can be used to study the data structure and to detect outliers. The model expresses strong non-linearity. LWR and ANN, in their best configurations, are comparable in prediction. Although there is a casual relationship between activation energy and rate constants (the Arrhenius equation), the performances of LWR and ANN are reversed. (This can be accidental.) Rate constants are predicted slightly better than activation energy.