Prognostic factors can be dichotomous by nature, but are more often either polytomous or continuous. Continuous variables are almost always used in prognostic models after categorization. Categorizing a continuous variable has adverse consequences: it requires the definition of cutpoints and implies a threshold effect: the patients immediately above and below the cut off value are supposed to be less similar than the patients at each end of the same category. Choosing a single cutpoint will result in a loss of statistical power in most instances. But many practical or bio logical reasons lead researchers to use one or several cutpoints in order to define the therapeutic classification. Different methods have been proposed to select cutpoints. One is called the optimal cutpoint method (Hilsenbeck et al, 1992). It is based on the P value of repeated univariate analyses of the same sample. This method has been rejected (Hill, 1993;Altman et al, 1994) because it artificially decreases the P value. Another possibility is to divide the sample into equal parts (thirds or quintiles for instance) on the basis of the continuous variable under study, and to evaluate the risk in each category.Quantitative prognostic variables in childhood ALL include the white blood cell count, peripheral blast cell count, haemoglobin level, platelet count and age at diagnosis. The prognostic value of these variables was identified more than 30 years ago (Tivey, Summary Many cutpoints have been proposed to categorize continuous variables in childhood acute lymphoblastic leukaemia (white blood cell count, peripheral blast cell count, haemoglobin level, platelet count and age), and have been used to define therapeutic subgroups. This variation in the choice of cutpoints leads to a bias called the 'Will Rogers phenomenon'. The aim of this study was to analyse variations in the relative risk of relapse or death as a function of continuous prognostic variables in childhood ALL and to discuss the choice of cutpoints. We studied a population of 1545 children with ALL enrolled in three consecutive protocols named FRALLE 83, FRALLE 87 and FRALLE 89. We estimated the risk of relapse or death associated with different values of each continuous prognostic variable by dividing the sample into quintiles of the distribution of the variables. As regards age, a category of children under 1 year of age was distinguished and the rest of the population was divided into quintiles. The floated variance method was used to calculate the confidence interval of each relative risk, including the reference category. The relation between the quantitative prognostic factors and the risk was monotonic for each variable, except for age. For the white blood cell count (WBC), the relation is log linear. The risk associated with WBC values in the upper quintile was 1.9 times higher than that in the lower quintile. The peripheral blast cell count correlated strongly with WBC (correlation coefficient: 0.99). The risk increased with the haemoglobin level, and the risk in the up...