Theoretically, the normalized plasma pressure (beta) at which a neoclassical tearing mode (NTM) is triggered is expected to depend on normalized Larmor radius (rho*) and normalized collisionality (nu), and this has formed the basis for the way in which NTM onset scalings are quoted for many devices. However, new analyses of JET data show that such rho*-nu based scalings are non-predictive, with discharges largely following such scalings over the majority of their duration while in H-mode. Neural network techniques indicate that a key additional parameter to include is the sawtooth period, providing a better degree of predictability. Indeed, this parameter appears more important than rho* in predicting NTM onset. Analysis using cases where sawteeth are modified by localized heating and current drive indicates that it is the sawtooth that governs where it is along the NTM onset scaling trajectory that the NTM is triggered, rather than leading to departures from the scaling. Finally, exploring data from cross-device similarity experiments shows similar absolute values in NTM onset beta(N) across devices of differing size and rho* range. This suggests the possibility that a. simple rho*-based extrapolation for ITER may be inappropriate and that NTM threshold levels may be more directly related to the absolute value of beta, suggesting a higher beta threshold for ITER, at least if large sawteeth are avoided.