Let l(v) be the performance (mean waiting time, mean sojourn time etc.) in a queueing model governed by a parameter v and ∇vl(v) the associated sensitivity (gradient). Estimating l(v) and ∇vl(v) by simulation, one may want to run the system for a different reference parameter v0 and apply the likelihood ratio-score function method. This paper addresses the question of how to choose v0. The resulting variance is calculated explicitly in the case of the performance l(v) in the M/M/1 queue, and simulation results are given for more general models and the sensitivity ∇vl(v).