We determine the uncertainties on observables arising from the errors on the experimental data that are fitted in the global MRST2001 parton analysis. By diagonalizing the error matrix we produce sets of partons suitable for use within the framework of linear propagation of errors, which is the most convenient method for calculating the uncertainties. Despite the potential limitations of this approach we find that it can be made to work well in practice. This is confirmed by our alternative approach of using the more rigorous Lagrange multiplier method to determine the errors on physical quantities directly. As particular examples we determine the uncertainties on the predictions of the charged-current deep-inelastic structure functions, on the cross-sections for W production and for Higgs boson production via gluon–gluon fusion at the Tevatron and the LHC, on the ratio of W− to W+ production at the LHC and on the moments of the non-singlet quark distributions. We discuss the corresponding uncertainties on the parton distributions in the relevant x,Q2 domains. Finally, we briefly look at uncertainties related to the fit procedure, stressing their importance and using W, H and extractions of S(M2Z ) as examples. As a by-product of this last point we present a slightly updated set of parton distributions, MRST2002.