A comparison of survival models for assessing risk of racehorse fatality

Survival analysis was used to assess risk factors for fatal injuries on UK race courses. This allowed assessment of variation due to temporal horse-level effects, including previous racing intensity and historical distribution of race types, as well as race-level factors. Comparisons were made between measuring survival time as number of days and as number of races to injury from the first race. Two related models were presented for time as number of races to injury: a Cox regression model fitted using partial likelihood, with the Efron approximation to handling ties, and a discrete-time logit model fitted using maximum likelihood. The latter approach had the advantages of being computationally more efficient and enabling the testing of different functional forms for the dependence of hazard on time. Retrospective data were available from all race starts on the 59 courses in Britain from 1990 to the end of 1999, as analysed by . The analysis was conducted on the data for the 47,424 horses that had started racing in the UK: 538,895 starts with 1,228 fatal injuries. Horses starting racing abroad were excluded, but some included horses would have raced abroad at some stage during their racing career. The results for the selected models were broadly consistent with each other and with previously published studies. Steeplechase and hurdle races had a higher risk of fatal injury than flat races (relative hazards 1.5 and 1.7, respectively). Risk increased with the firmness of surface, age and race distance (reaching a plateau at 20 furlongs) and decreased with previous racing intensity (reaching a plateau after seven races run in the last 12 months). Horses running their first race of a new type were also found to be at higher risk (relative hazard 1.5). The main difference between the models for time as number of days and number of races concerned the role of age: age at race was identified as the more important factor in the latter model, whereas, age at first race was more significant in the former model.