1: \begin{abstract}

2: We investigate the effect that the choice of measurement

3: scale has upon inference and extrapolation in extreme value analysis.

4: Separate analyses of variables from a single process on scales which are

5: linked by a nonlinear transformation may lead to discrepant conclusions

6: concerning the tail behavior of the process. We propose the use of a

7: Box--Cox power transformation incorporated as part of the inference

8: procedure to account parametrically for the uncertainty surrounding

9: the scale of extrapolation. This has the additional feature of

10: increasing the rate of convergence of the distribution tails to

11: an extreme value form in certain cases and thus reducing bias in

12: the model estimation. Inference without reparameterization is

13: practicably infeasible, so we explore a reparameterization which

14: exploits the asymptotic theory of normalizing constants required for

15: nondegenerate limit distributions. Inference is carried out in a Bayesian

16: setting, an advantage of this being the availability of posterior predictive

17: return levels. The methodology is illustrated on both simulated data and

18: significant wave height data from the North Sea.

19: \end{abstract}