Abstract: The number of stillborn piglets per litter is a low-mean count trait that may exhibit overdispersion and zero inflation, and misspecification of the distributional assumption can lead to biased inference. This study systematically compared the applicability of six count models for this trait. Taking the number of stillborn piglets per litter as the target trait, a Monte Carlo simulation framework was built with four true data-generating mechanisms. Six models were fitted by maximum likelihood and evaluated using information criteria, cross-validated predictive likelihood, and convergence rates. The models were validated on public Danish farrowing records and more than 540,000 records from our farm. Model selection showed a clear mechanism-matching pattern: ZINB and Hurdle-NB failed to converge when overdispersion was absent. Both real datasets exhibited pronounced overdispersion and zero-inflation, but the best-performing model differed between datasets: ZINB had the best AIC and cross-validated predictive log-likelihood for the Danish data, whereas Hurdle-NB performed best on these two criteria for the farm data. For the fair pricing of a per-litter excess-loss compensation contract, Hurdle-NB deviated from the actual payout by only 1.5%, whereas Poisson underpriced the contract by 61.0%. NB is recommended as a robust baseline, and ZINB/Hurdle-NB as primary candidate models, with a distribution-aware selection strategy; the simple Poisson model should not be applied directly to stillborn piglet counts per litter.

Key words: pig, number of stillborn, zero-inflated model, negative binomial model, Monte Carlo simulation