Clustering in the denominator: non-independence of starts in racing fatality studies

This really avoids your excellent question, but it may be worth taking stock of the likely correlation structure. A random intercept instantiates a compound symmetric correlation pattern that doesn’t respect the forward flow of time. A serial correlation pattern, e.g., a Markov process, may be more likely to hold. This comment would apply more to a non-death outcome perhaps. Related to death, this is a terminal event and is most elegantly handled as an absorbing state in a state transition model, at least in other settings I’m more familiar with.