On the Need for a Statistical Foundation in Scenario-Based Testing of Autonomous Vehicles