The effect of sampling strategy on the estimated prevalence of welfare outcome measures on finishing pig farms

The effect of sampling strategy on the estimates of prevalence of some welfare outcome measures was investigated as part of a larger project examining the feasibility and benefits of incorporating some on-farm welfare outcome assessments into UK Pig Farm Assurance Schemes. All finishing pigs (>40 kg) on 6 farms that had pens of 25 pigs or less were assessed for dirtiness, body lesions, tail lesions, bursae, lameness, oral behaviour and pigs requiring hospitalisation. A statistical sampling methodology was applied to the data to generate 10,000 random samples for each measure, on each farm and for every possible number of sampled pens on that farm. The 95% confidence interval range, the proportion of samples outside a 'tolerance' range and the proportion of samples that would wrongly classify the farm with respect to a 'threshold' were calculated for each measure on each farm for every possible sample size. It was found that even large sample sizes were unable to accurately reflect the whole farm for measures with very low prevalences (lameness and pigs requiring hospitalisation). For the other measures, as the number of pens sampled increased, the confidence interval range and the proportion of samples outside the 'tolerance' range reduced. The differing inter-pen variation in prevalence on a farm gave rise to differences in accuracy of prevalence estimates between the farms. Those farms with a true farm prevalence close to the threshold required more pens to be sampled to achieve the same level of accuracy of classification as other farms with similar confidence intervals but a more extreme prevalence. The effect of sampling on the estimate of prevalence should be considered when an accurate prevalence of the whole population is required.