Ontario Pork Carcass Appraisal Project Symposium

Appendix 3
Summaries of Statistical Analyses

Data Preparation and Checking

All data collected in the OPCAP, almost 1 million data items in total, were entered into a single data base. Data was examined at entry with checks for consistency with previous data for the animal in question and that values lay within sensible bounds. Prior to statistical analysis all traits were examined for severe deviations from normality, and outliers removed on a case by case basis; usually following examination of univariate or bivariate scatter plots. Allowing for the large number of observations, outliers generally had to be at least 3.5 s.d. from the mean and clearly outside the distribution as observed visually before being considered for elimination. Outliers were rare, accounting for 0 to 9 observations per trait.

A. Growth and Carcass Traits

Mixed model analyses were performed with sires, dams and herd of origin (within breed) as random effects, and breed, sex, PSS genotype (homozygote normal, heterozygote, or unknown; seven homozygote mutants were eliminated from the analyses) and all their two-way interactions, plus fill number (38 fills of animals into the test station) as fixed effects. No covariates were included for analyses of growth traits, carcass weights, dressing percentage and carcass index. Hot carcass weight was included as a covariate for all other traits. Least squares means are generally from the full model unless otherwise indicated. This allows consistency of interpretation. In all cases, however, models were reduced for each trait until all non-significant interactions were eliminated. If an interaction involving PSS remained in the model, animals with PSS genotype unknown were eliminated. Unless indicated, this had no more than trivial impact on estimates of effects. Analyzed in this way, the test of breed differences had 3 and 115 d.f., recognizing that herd of origin was nested within breed. This model provides a conservative test of breed differences, which was deemed appropriate given that the between herd variance component was highly significant for most traits.

B. Ultrasound Predictions of Carcass Composition

A variety of least-squares analyses of variance were performed for each carcass trait of interest, with varying combinations of ultrasound measurements included as covariates. A full model was explored to set an upper limit to the R² of prediction possible. This model had fill number, herd, ultrasound technician, breed, sex and breed*sex as fixed effects, with regressions on all within-site linear, quadratic and interaction terms of the ultrasound measurements being examined in that analysis, plus live weight at scanning. Analyses were performed across interpreters (in which case a fixed effect for interpreter was added), and for each interpreter separately. Reduced analyses started by eliminating fill number, herd and technician, and then reduced covariates to the most parsimonious set of linear, quadratic and interaction terms by stepwise elimination of least significant terms. Non-linear terms were then removed. Breed and sex was then removed, and the most parsimonious covariate model was found, including all linear, quadratic and interaction terms. Finally, all non-linear covariates were removed. Prediction equations would generally include breed and sex. Where breed and sex remained in the model, once suitable reduced models had been found, breed and sex by covariate interaction terms were added, to check homogeneity of regressions across breed and sex. In no case were significant breed or sex by covariate effects observed.

C. Taste Panel Data

Trait definition and scale of measurement for taste panel traits differed between Guelph and Lacombe and data from the two sites was analyzed separately. In both cases, fill number, herd, breed, sex, PSS genotype, breed*sex, breed*PSS and sex*PSS were initially included as fixed effects, with litter of birth and animal as random effects. Covariates included weight at slaughter, lean % of three primals, chemical lean in belly, chemical fat in belly, chemical lean in loin, chemical fat in loin and time from slaughter to tasting. Taste panellist was included as fixed for Lacombe data where a constant set of six panellists was used for all tasting sessions. Taste panellist was treated as a random error term for the Guelph data where panellists were less highly trained and varied between sittings. Non-significant interaction terms and covariates were then eliminated sequentially, along with fixed effects other than breed, sex and PSS status.

D. Boar Taint Data

Data for skatole, androstenone in fat and androstenone in salivary gland was analyzed on both the natural and Log₁₀ scale. Neither scale gave normal distributions. The Log scale gave the more normal distributions, but caused highly heterogeneous variances across different assay subclasses, which was not a problem with the untransformed data. Analyses included assay batch (three levels), fill number, within assay date, breed and PSS as fixed effects, with age at slaughter, average daily gain, lean % in three primals, and time of sample storage (linear and quadratic terms) as covariates. Additional analyses were run with age or weight at slaughter as fixed effects with categories <170 days, 170-180 days and >180 days, or <105 kg, 105-115 kg and >115 kg. In the latter case, some or all covariates were eliminated, except for storage time. Parsimonious models were found by sequential elimination of non-significant effects. Analyses were repeated with date of assay, herd and fill number as random effects, but had trivial impact on estimates of fixed effects and residual variances. Since time of storage of the tissue sample and assay batch had highly significant effects, all estimates were then corrected to day of slaughter and to last assay date (the largest group of samples) before being used in further analyses.

Corrected skatole and androstenone levels were then included singly and together as covariates in reanalyses of the taste panel data. Models were otherwise as at C above, except that only age at slaughter, time from slaughter to tasting and lean % in three primals were included as additional covariates, and sex was dropped since skatole and androstenone levels were recorded only for males.

E. Estimation of Genetic Parameters

Genetic variances and covariances were estimated using the multi-trait REML VCE package of Groeneveld (1996). Details of fixed effects, random effects and covariates included for each trait are given in the table below. An animal model was used with additive genetic relationships traced back two generations. There were 549 sires, 849 dams and 118 herds of origin. Parameters for the first seven traits were estimated in a single multiple-trait analysis. Variances and covariances among the quality traits, marbling, drip loss and colour, were estimated in a separate three trait analysis. Covariances between quality and other carcass traits came from three additional analyses including:

1. TLEAN, ADJFAT, MARB, COLOUR, DRIPL
2. LLEAN, LEA, PRBFAT, MARB, COLOUR, DRIPL
3. LONGMD, LONGFAT, MARB, COLOUR, DRIPL