Description: Estimating small-scale V P / V S variations at depth can be a powerful tool to infer lithology and hydration of a rock, with possible implications for frictional behavior. In principle, from the differential arrival times of P and S phases from a set of spatially clustered earthquakes, an estimate of the local V P / V S can be extracted, because the V P / V S is the scaling factor between the P and S differential times for each pair of earthquakes. We critically review the technique proposed by Lin and Shearer (2007) , in which the mean value over all stations is subtracted from the differential arrival times of each pair of events in order to make the method independent of a priori information on origin times. The demeaned differential P and S arrival times are plotted on a plane, and the V P / V S ratio is estimated by fitting the points on this plane. We tested the method by both theoretical analysis and numerical simulations of P and S travel times in several velocity models. We found that the method returns exact values of V P / V S only in the case of a medium with homogeneous V P / V S , whereas, when a V P / V S gradient is present, the estimates are biased as an effect of systematic differences between P and S takeoff angles. We demonstrated that this bias arises from the demeaning of the arrival times over the stations. In layered models with V P / V S decreasing with depth, we found that V P / V S is overestimated or underestimated, respectively, for takeoff angles larger or smaller than 90°. Moreover, we calculated analytically the dependence of this bias on the takeoff angles. Our simulations also showed that the difference between the calculated and the expected V P / V S is reduced for simple horizontally layered velocity structures (〈0.06), whereas it is 0.27 in a more realistic velocity model mimicking a subduction zone.