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

I might be thinking of this wrong but my first thought would be to treat each start as an exposure and have n horses with different numbers of races so you’d get something like an exponential survival model? Something like death ~ XB + offset(starts)?