Data source and variables
Whole carcass condemnation data were obtained from the Food Safety Decision Support System (FSDSS) database maintained by the Ontario Ministry of Agriculture, Food and Rural Affairs (OMAFRA). Data were extracted from the database for cattle animal classes: bulls, calves, cows, heifers and steers from January 1, 2001 to December 31, 2007. Missing geographical coordinates for abattoirs were approximated using postal codes and/or addresses with the address geocoding software GeoPinpoint Suite 6.4 (DMTI Spatial Inc., Markham, Ontario, Canada). Using the FSDSS database, the following information was extracted for each month: abattoir identification number, geographical coordinates of abattoir, year, season, number of weeks an abattoir was open each year, total number of whole carcasses condemned, total number of cattle processed each year, and animal class. Season was categorized by 3 month groupings as follows: winter (December - February), spring (March - May), summer (June - August) and fall (September - November). Animal class included five categories: bulls, cows, calves, heifers, and steers. Bulls were excluded from subsequent statistical analyses due to missing data and inconsistencies in the use of this classification. The number of weeks an abattoir was open each year was determined by the total number of weeks in which at least one bovine animal was processed. The total number of animals processed each year was calculated by adding the total number of condemned cattle and the total number of cattle fit for consumption each year. Linearity of continuous variables was assessed by plotting the log of the condemnation rate against the covariate using a lowess smoother. If there was no visible linear relationship between the outcome and the covariate, and the association could not be adequately modeled with a quadratic term, or transformation, then the variable was categorized.
Abattoir audit ratings were obtained for all abattoirs from the abattoir audit program administered through OMAFRA. The audit program assesses each facility's food safety performance and compliance with the Ontario Meat Inspection Act. Audits are conducted once a year and evaluate each premise on 14 food safety areas based on the Standards of Compliance relating to food safety, animal welfare and occupational health and safety with a letter grade given for each abattoir [18]. Annual OMAFRA audit ratings were obtained for all abattoirs in the audit program from 2001-2007. Abattoir audit ratings were classified according to the letter grade received from best to poorest as follows: AAA, AA, A, B or C and unrated for abattoirs that had missing data.
The price of cattle was obtained from the Ontario Cattlemen's Association market reports for 2001-2007. Prices were calculated to be the average price (in Canadian dollars) per lbs based on sales records from Ontario sales-yards. A price was assigned for each month and year by animal class. The most appropriate weight category was selected to represent each animal class based on an average animal at the time of slaughter.
The agricultural region where the abattoir was located was classified as: central, eastern, northern, southern or western Ontario using the Census Agricultural Region boundaries (Statistics Canada, Census Agricultural Regions, Census year 2001). The regional location of each abattoir was determined using the point-in-polygon technique with geographic information system software ArcGIS 9.2 (ESRI, Redlands, California, USA).
Travel distance between the animals' farm and the abattoir was estimated using data obtained from OMAFRA. Because farm location is not routinely recorded with condemnation data, a subset of cattle, in which a sample was sent for laboratory testing, were used to obtain geo-location information for the abattoir and farm. Like abattoir location, owner address information was geo-coded according to the owner postal code using geocoding software GeoPinpoint Canada. Distance from the farm to abattoir was calculated using the Haversine distance formula, which calculates the great-circle distances between two points on a sphere using their longitudes and latitudes [19].
Data from all sources were merged into one master dataset using Stata 10.1 (Stata Corp., College Station, Texas, USA).
Statistical analysis
To model and evaluate their association with monthly whole carcass cattle condemnation rates, the effect of year, season, annual audit rating, number of weeks open, number of cattle processed, census agricultural region, animal class and sales price of animal class were included in the model. All covariates were evaluated for statistical significance individually and then in a multivariable model using generalized estimating equations (GEE) to fit a negative binomial regression model with an exchangeable correlation structure to account for repeated measurements from each abattoir. Wald tests were performed on each covariate in the model to estimate the significance of each categorical variable as a group. Non-significant covariates (p ≥ 0.05) based on the Wald test were removed from the model. All excluded covariates were evaluated for their potential confounding effect by evaluating if their removal produced a 20% or greater change in the coefficient of significant variables in the model. Interactions between price and year, year and number of animals processed, year and animal class, as well as season and animal class were investigated. In addition, the covariates included in the model were then fitted using GEE with both Poisson and negative binomial distributions and the following correlation structures: exchangeable, first order autoregressive structure, second order autoregressive structure, non-stationary, and stationary. All resulting models were evaluated for how well the model fit the data using a quasi-log-likelihood under the independence model information criterion (QIC) statistic for model selection [20]. The model with the lowest QIC was selected as the final model. Robust standard errors were used for all GEE fitted models. All statistical analyses were performed using Stata 10.1.