Role of stability and triangulation-based methods to improve identification of risk factors for lameness in ewes

Lameness has serious impact on sheep health and profitability. In the UK, the largest-scale questionnaire investigating risk factors for lameness (107 questions and 1260 respondents) identified 20 significant variables using stepwise Poisson regression. It is now known that stepwise procedures with wide data can result in overfit models. This research reanalysed these data, using methods that minimise the likelihood of overfitting and therefore reduce the probability of identifying false positive variables. Poisson and log-normal regression models were built with six different variable selection methods, stability selection and triangulation. Six variables were selected in the final triangulated models associated with a reduced prevalence of lameness, fewer than the 20 variables selected in the original analysis. These six variables covered early treatment of individual sheep, treating sheep with any severity of lameness, avoiding routine foot trimming and avoiding footbathing to treat underrunning footrot. Early treatment of individual lame sheep had the highest population attributable fraction for reduction of lameness. Our results highlight the importance of addressing overfitting when fitting models to wide data and the usefulness of triangulating results across different model types. The results strengthen the evidence that the greatest reduction in lameness nationwide would be achieved if farmers treated the first lame sheep in a group rather than waiting until more become lame.