Estimating return levels of extreme wind speeds due to hurricanes pres
ents both practical and analytical difficulties. The practical difficu
lty of collecting data has been resolved in the past by modelling simu
lated data-we adopt such an approach in this paper also. The analytica
l difficulties concern the problem of estimating the probabilities of
events which are more extreme than those simulated. We follow common p
ractice here also, using standard extreme value models to describe ext
reme tail behaviour. We differ from previous analyses of hurricane dat
a in two respects. First, we use a model parameterisation which enable
s models fitted at different thresholds or at different sites to be ea
sily compared. Second, we use maximum likelihood as the method of infe
rence. This is found to produce results similar to those of previous s
tudies, but enables the development of a spatial analysis which exploi
ts similarities in the behaviour of the data from one site to another
in order to improve the precision of estimation, and facilitates predi
ction at coastline locations other than those with simulated data. (C)
1998 Elsevier Science Ltd. All rights reserved.