This paper proposes a coherent strategy for the statistical evaluation of 24-h ambulatory blood pressure recordings in controlled clinical trials. Because the usual cosinor method for the smoothing of individual profiles leads to unrealistic consequences, an alternative flexible approach has been developed. The new approach is based on a periodic spline model which is fitted by robust multivariate regression techniques. The model is used mainly to extract those circadian parameters that cannot be obtained from the original profiles due to artefacts or irregular sample points. In the comparison of groups of patients, the individually obtained parameters are then tested with exact non-parametric methods. For the prior selection of highly discriminating parameters, the method of classification trees is recommended.