121 Statistical Analysis Method Counts for Sow Count Data Responses

Abstract

Abstract Several statistical analysis methods are typically employed to analyze sow reproductive count data. The research objective was to compare analysis methods of pig birth counts to determine their robustness in identifying simulated treatment differences. Counts of stillborn (SB), born alive (BA) and sow parity differences were simulated using descriptive statistics from a sow farm. Different scenarios were tested: 1) Effect of a 0.5, 1.0, 1.5, and 2.0 percentage point change in treatment difference in SB and BA and, 2) Replicates of 20 to 200 experimental units (EU) in increments of 20 sows; yielding 40 total scenarios. For each scenario, sow observations were simulated 1000 times over. Random sub-setting was used to create a random effect of parity in each dataset as follows: 20% Parity 1, 50% Parity 2–4, and 30% Parity 5+ sows. Each simulated scenario was analyzed as: 1) General linear model (GLM) with raw counts of number of SB or BA as the response variable, 2) GLM with the ratio of BA or SB to total born as the response variable, and 3) Generalized linear mixed model (GLMM) with a binomial distribution of SB or BA as events and total born as trials. Across the EU replicate range, gross performance of models was compared by measuring area under the curve (AUC) with EU as abscissa and the probability of the simulation being P < 0.05 as ordinate. Simulation results are provided in Table 1. The GLMM has elevated probability of detecting true treatment differences over both GLM models for SB and BA. For BA analysis, the GLM Model 1 the probability of detecting true differences is greatly reduced vs. the other two models. This research indicates that deploying GLMM in analyses is a more effective and improved method to detect true differences in sow count data.