Abstract Background Measurement of obesity using self-reported anthropometric data usually involves underestimation of weight and/or overestimation of height. The dual aim of this study was, first, to ascertain and assess the validity of new cut-off points, for both overweight and obesity, using self-reported Body Mass Index furnished by women participants in breast cancer screening programmes, and second, to estimate and validate a predictive model that allows recalculate individual BMI based on self-reported data. Methods The study covered 2927 women enrolled at 7 breast cancer screening centres. At each centre, women were randomly selected in 2 samples, in a ratio of 2:1. The larger sample (n = 1951) was used to compare the values of measured and self-reported weight and height, to ascertain new overweight and obesity cut-off points with self-reported data, using ROC curves, and to estimate a predictive model of real BMI using a regression model. The second sample (n = 976) was used to validate the proposed cut-off points and the predictive model. Results Whereas reported prevalence of obesity was 19.8%, measured prevalence was 28.2%. The sensitivity and specificity of this classification would be maximised if the new cut-off points were 24.30 kg/m2 for overweight and 28.39 kg/m2 for obesity. The probability of classifying women correctly in their real weight categories on the basis of these points was 82.5% in the validation sample. Sensitivity and specificity for determining obesity using the new cut-off point in the validation sample were 90.0% and 92.3% respectively. The predictive model for real BMI included the self-reported BMI, age and educational level (university studies vs lower levels of education). This model succeeded in correctly classifying 90.5% of women according to BMI categories, but its performance was similar to that obtained with the new cut-off points. Conclusions Quantification of self-reported obesity entails a considerable underestimation of this problem, thereby questioning its validity. The new cut-off points established in this study and the predictive equation both allow for more accurate estimation of these prevalences.