The relation between change and initial value is of great interest in longitudinal studies. With variables containing random errors (short-term intra-individual variations and measurement errors) the directly computed relation is however, biased by the regression towards the mean phenomenon. Earlier proposed solutions of the problem are unsatisfactory. In this paper the regression towards the mean phenomenon is described and a new method is proposed by which the error caused by the regression towards the mean is avoided. The method is applied to a set of longitudinal blood pressure data. It is shown that the observed, biased relation in this case is significantly negative, while the correct relation obtained with this method is significantly positive. Since random errors are present in most biological variables, similar erroneous conclusions may easily be drawn also in other cases if the regression towards the mean phenomenon is not corrected for. In this analysis, random errors constitute 65–80% of the observed blood pressure change. To reduce this dominance, recommendations about study design for future studies of change/initial value relationships are given.