The condition of the latest version of the ELMy H-mode database has been re-examined. It is shown that there is bias in the ordinary least squares regression for some of the variables. To address these shortcomings three different techniques are employed: (a) principal component regression, (b) an error in variables technique and (c) the selection of a better conditioned dataset with fewer variables. Scalings in terms of the dimensionless physics variables, as well as the standard set of engineering variables, are also derived. The new scalings give a very similar performance for existing scalings for ITER at the standard βn of 1.6, but a much improved performance at higher βn.