We developed a Bayesian framework to determine in a robust way the relation between velocity dispersions and chemical abundances in a sample of stars. Our modelling takes into account the uncertainties in the chemical and kinematic properties. We make use of RAVE DR5 radial velocities and abundances together with Gaia DR1 proper motions and parallaxes (when possible, otherwise UCAC4 data is used). We found that, in general, the velocity dispersions increase with decreasing [Fe/H] and increasing [Mg/Fe]. A possible decrease in velocity dispersion for stars with high [Mg/Fe] is a property of a negligible fraction of stars and hardly a robust result. At low [Fe/H] and high [Mg/Fe] the sample is incomplete, affected by biases, and likely not representative of the underlying stellar population.