This paper concerns a series of simulations undertaken to examine the effects of two data features--number of cutpoints and true marker prognostic effect size--on three methods of p-value adjustment (asymptotic, P(acor); improved Bonferroni, P(bon); and empirical permutation, P(emp)). H(o) rejection rates for P(emp) and P(bon) are almost indistinguishable from those for an independent validation sample (P(vld)), while those of P(acor) are somewhat conservative, especially when the number of cutpoints is small. Analysis of a new breast cancer prognostic marker, heat shock protein 70, illustrates the methods. These results underscore many of the problems associated with data-derived cutpoints in general, and the need for p-value adjustment.