Sampling for welfare assessments: Individual cow, pen, and farm data with SAS code for dairy cow sampling strategies

Introduction. The aim of this study was to evaluate how sampling strategies currently used in 4 North American cattle welfare assessment programs affect the classification of dairy farms relative to various thresholds for lameness prevalence, injuries on the tarsal (hock) and carpal joints, and body condition score. Methods. All 12,375 lactating cows on 38 freestall dairy operations in the state of Paraná, Brazil were assessed between March and October 2016 for lameness, injuries on the tarsal (hock) and carpal joints, and body condition score (BCS). In addition, we collected farmer-reported numbers of dry cows, pre-weaned heifer calves, and weaned heifers ≤1 yr old. The number in the latter 2 groups was doubled to estimate the number of bred heifers. These categories were summed with the number of lactating cows to generate an estimate of total herd size. To test Objective 1, we evaluated 9 sampling strategies based on those used in existing North American dairy cattle welfare assessment programs. Desired precision (d) of 15, 10, or 5% was used in the sample size formula, applied to either all lactating cows, a high-producing pen of cows, or the total herd size (with lactating cows selected in proportion to their representation in the herd) on all n = 38 farms. To test Objective 2, all cows in the high-producing pen were compared against all lactating cows on all n = 38 farms. To test Objective 3, the sample size formula using d = 15, 10, or 5% was applied to all lactating cows either including or excluding the hospital pen on the n = 20 farms with this pen type. For Objectives 1 and 3, we selected cows randomly in 10,000 replicates using a simple random sampling technique in SAS, and the outcome measures were classification metrics (percent agreement, kappa, sensitivity, specificity, and positive and negative predictive value) for the true vs. estimated prevalence based on the samples. For Objective 2, the outcome measures were classification accuracy and kappa for the high-producing pen vs. all lactating cows and the linear regressions for the relationship between these populations. Results. As the number of sampled cows increased, so did classification performance. Assessing a single pen of high-producing cows resulted in farm classifications similar to all lactating cows, and the former might be a more efficient assessment strategy. Relative to selecting from all lactating cows, excluding cows in the hospital pen had little effect on farm classification