Estimating the frequency of nuclear accidents

Abstract

We used Bayesian methods to compare the predictions of probabilistic risk assessment -- the theoretical tool used by the nuclear industry to predict the frequency of nuclear accidents -- with empirical data. The existing record of accidents with some simplifying assumptions regarding their probability distribution is sufficient to rule out the validity of the industry's analyses at a very high confidence level. We show that this conclusion is robust against any reasonable assumed variation of safety standards over time, and across regions. The debate on nuclear liability indicates that the industry has independently arrived at this conclusion. We pay special attention to the Indian situation, where we show that the existing operating experience provides insufficient data to make any reliable claims about the safety of future reactors. We briefly discuss some policy implications.