Dichotomisation of Continuous Data: Review of Methods, Advantages, and Disadvantages

Background: In medical research, dichotomisation of continuous variables is a widespread use approach. However, it has been argued that dichotomisation might be waste of information. The aim of this paper is to review the main methods to dichotomise continuous data, to address practical issues around dichotomisation methods, and to investigate whether dichotomisation is always a bad idea. Methods: A total of 310 breast cancer patients were recruited. Information on 3 categorical and 1 continuous variable (age at diagnosis) was available. Missing data were imputed applying the Multivariable Imputation via Chained Equations (MICE) method. Then a minimum P-value method was applied to dichotomise the age variable. The Cox regression model was fitted to develop models in which dichotomised versus continuous version of the age variable plus other 3 variables were used. Results were compared in terms of discrimination ability, goodness of fit, and classification improvement. Results: For the age variable, an optimal split at 47 was found. This split was close to menopause age of women in Shiraz (48) so had biological interpretability. The stability of optimal split was confirmed in bootstrap study. Model in which dichotomised version of age was used showed higher discrimination ability and goodness of fit. Furthermore, dichotomised model assigned 14% of live patients into a more appropriate risk group. Discussion: Dichotomisation of continuous data is a contentious issue. We have shown that dichotomisation might improve performance of models when it has biological interpretation. More research is needed to understand situations in which dichotomisation might work.