The paper discusses the use of distribution functions in the description of
the long-term behaviour of significant wave height data. It is shown by an
example that the usual goodness-of-fit diagnostics do not ensure correct p
redictions of return values based on arbitrary statistical models. The poss
ibility of defining populations of significant wave height and their associ
ated sampling reference periods is investigated, and it is concluded that s
amples associated with the same reference period are subject to a significa
nt statistical variability. A method for describing the long-term behaviour
of H-s that accounts for this variability is proposed. (C) 2000 Elsevier S
cience B.V. All rights reserved.