A common problem in seismic tomography is to assess and quantify data uncertainties. The Bayesian approach to inverse problem by means of Markov Chain Monte Carlo (McMC) method samples relevant parts of the model space and provides an quantitative overview of the uncertainty of all model parameters. This method is very computing power intense and one important issue is to optimize the efficiency of the method. In this study, we investigate the difference between velocity-based and slowness-based McMC in refraction tomography. Whereas velocity in surface wave phase velocity inversions typically varies no more than by a factor of two, variations in refraction tomography can amount to a factor of ten, and the difference between slowness and velocity perturbations becomes more relevant. Because slowness is proportional to travel time, model perturbations need no arbitrary scaling relations. In our experiments, the associated perturbations are more uniform and show better mixing properties compared to velocity based McMC. We also investigate multivariate perturbations based on a projection of a single perturbation through the resolution matrix. Our tests show that these lead to higher acceptance ratios and/or greater step length.