OBJECTIVES: To select thresholds for predictive biomarkers using Cox regression.  METHODS: We used data from a Cuban trial designed to assess the efficacy of immunotherapy against the epidermal growth factor (EGF) to test our approach. The trial included 122 patients diagnosed with non-small-cell lung cancer (NSCLC) who had basal EGF concentration available. The EGF concentration was analysed as a predictor of immunotherapy success over the range of all possible values of the biomarker (w [a,b]). For each wi, patients with w > wi0 were selected and a Cox model adjusted to assess survival. We then identified the w0s with significant treatment results to find (a) the lowest biomarker threshold where the effect of treatment was significant and also to find (b) the biomarker threshold that reflected the highest difference between treatments. RESULTS: For NSCLC we observed that EGF concentration thresholds range from 870 pg/ml to 2000 pg/ml were significant. At the lowest threshold (870 pg/ml) the immunotherapy group showed a 6-month difference for the median survival (p = 0.022) whereas at the threshold that showed the maximum difference between treatments (EGF = 1750 pg/ml) the immunotherapy group presented a 10-month difference for the median survival (p = 0.004). CONCLUSIONS: The evaluation of p-values of the effect of treatment for each w0 [a,b] allows the selection of the thresholds where the treatment result is significant. Whereas the lowest threshold where the effect of treatment is significant allows the selection of patients that could be mostly benefited with the treatment, the selection of the threshold with the minimum p-value will reflect the higher difference between treatments.