Testing autonomous vehicles is a challenging task, usually carried out using test scenarios, which are often derived manually from accident statistics and real traffic datasets, relying on expert knowledge and intuition to select the most relevant ones, a cumbersome task given the large size of these datasets. In this paper, we suggest a model-based approach to compare scenarios using quantitative measures computed by probabilistic model checking of user-defined temporal logic properties characterizing interesting event sequences. This approach facilitates the selection of the scenarios having the best tradeoff between coverage and overall testing cost (both for simulation and field testing). We illustrate the approach by comparing variations of scenarios, derived from frequent situations in accident statistics, using measures such as collision probability, arrival probability, and duration.