We propose a predictive bound approach to determine the non-inferiority margin. This allows indirect statistical inference (i.e., p-values, confidence intervals and point estimates) on the effect of a new test drug against a putative placebo upon completion of the non-inferiority trial. We also develop a framework to evaluate the robustness of any non-inferiority margin approach against the constancy assumption, selection bias and variability of the historical data. With a proper choice of the predictive confidence level, the predictive bound approach ensures that the non-inferiority test remains interpretable even though the constancy assumption does not hold.