Because data are collected for only fatal crashes, it is difficult to accurately measure seat belt and air bag effectiveness. The use of safety devices influences survival rates which in turn determine whether a crash is included in the sample, leading to sample selection bias. We propose a simple solution to the selection problem: limiting the sample to crashes in which someone in a different vehicle dies. Empirically, we find seat belts more effective and air bags to be less effective than previously found. The cost per life saved through seat belts is approximately $30,000, compared to $1.8 million for air bags.