A high relative frequency of a variable being included in the various models, and a consistent association with flock colonization across models (Table 4), may help indicate the true causal role of that factor, and hence the potential for producers to decrease risk on the farm by applying an appropriate intervention directed at that factor. Median flock size, followed by farm water source and the presence of other domestic livestock on the farm, were the factors that were included in most or all of the models. Spreading manure on the farm in the winter season was present in 60% of the models, the number of broiler houses was present in 50%, and storing manure on the farm at any time of year was present in one-third of the models. An all-in-all-out policy at the farm level (i.e. the practice of shipping all flocks on the farm within the span of a few days, with all houses remaining empty for a period of time) was not a significant predictor in any of the models. The direction of association was inconsistent for the presence of other commercial poultry on the farm and spreading manure on the farm in the summer season, and the statistical significance of the former was also inconsistent among models. In general, the models that employed a backward elimination approach had slightly smaller AIC's and a larger number of significant predictors than those using a forward approach. Backward selection has an advantage over forward and stepwise selection procedures in that negatively confounded sets of predictors are less likely to be omitted from the model . Thus, more emphasis could be placed on variables identified as significant in the backward-type models.
An increased risk of Campylobacter was associated with increasing median flock size on the farms. For example, as the average flock size increased by 5,000 birds, the risk of Campylobacter colonization increased by approximately 57% to 92% (i.e. 1.57 to 1.92 times). Our findings are in contrast with several studies [22–24, 26] that utilized multivariable logistic regression at the flock level, in which an association between flock size and Campylobacter status was not found. In a one-year study of 18 Swedish broiler farms, infection risk increased when the flock size was more than 25,000 birds . However, the authors noted that since only univariable associations were examined, their conclusions may have been confounded by farm size and management practices. To our knowledge, ours is the first study to examine the effect of the average flock size on the farm on the risk of Campylobacter colonization. It has been suggested that larger flocks require more water, feed, litter, air and personnel, all possible sources of the bacteria . Thus, in our study, increasing median flock size may be a surrogate for many other factors.
In our study, an official (municipal) water source was one in which the water was tested regularly for coliform bacteria by the municipality and treated if necessary, and an official treated water source was one in which the water was treated consistently with either ultraviolet (UV) light or a heat-cool method at the municipal level. We found that farms using official water sources had approximately one-third to half the risk of Campylobacter than farms using non-official untreated sources (the referent group). Similarly, farms using official treated water sources had roughly one-third the risk. These findings suggest that some flocks may have been exposed through contaminated water, as water has been identified as a suitable reservoir and medium for Campylobacter spp. . Several studies [23, 34–36] have found that there was no association between the occurrence of Campylobacter in flocks supplied with municipal (public) water compared to those supplied with well (private) water, however, in those studies, there was no distinction between the use of treated and untreated water sources. We found that farms using a non-official UV-treated water supply did not have a significantly different risk of Campylobacter than farms using non-official untreated water at the 5% level of significance, although in one model, non-official treated water did have a protective effect at a 10% level. Some researchers [25, 26] have found that water disinfection had a protective effect on the colonization of broilers with Campylobacter, although others [24, 36, 37] have not found such an association. The small number of farms using non-official treated water in our study, combined with potential confounding by other factors, may account for the wide range in p-values for this variable. Our results suggest that the use of municipal water (both treated and untreated) reduces the risk of Campylobacter colonization of broiler flocks, and that some potential also exists for decreasing risk through the practice of treating non-official water sources, depending in part on other management practices on the farm. It is possible that there may be other, more indirect factors contributing to the risk of colonization, such as animal density in the region. In addition, there may be complex relationships between access of livestock to the water source, type of water source (drilled versus upcoming wells), and the method of water treatment (UV versus heat-cool) that were not adequately addressed in this study. Dissection of these inter-relationships would require a study in a country or region with a larger number of farms.
The presence of other domestic livestock on the farm was associated with a decreased risk of Campylobacter colonization. Similar results were obtained when we assessed the effect of the presence of cattle, rather than the presence of other domestic livestock in general. These findings were unexpected and inconsistent with other studies as it has been suggested that other domestic livestock species (especially cattle) may act as reservoirs that potentially contaminate the farm environment thereby providing a continual source of bacteria to the birds . Several studies have shown that the presence of other animals on the farm (pigs, cattle, sheep, or fowl other than broilers) , (cattle) , (pigs, cattle, sheep and goats, or horses) , (laying hens, sheep, cattle, donkeys)  was associated with an increased risk of Campylobacter, although one recent Canadian study did not find such an association (cattle, sheep, goats, horses and/or pigs) . However, in one Norwegian study , the presence of other poultry or animals at the farm was not associated with increased colonization of flocks, rather, tending other poultry and tending pigs prior to entering the broiler house were independently associated with an increased risk. In our study, farms that did and did not keep other domestic livestock were similar with respect to the number of flocks raised and the number of houses, both surrogates of farm size. The distance between the broiler houses and the housing for the other livestock is quite variable among broiler farms in Iceland, with distances ranging from immediately adjacent to approximately 900 m apart. Additionally, consistent patterns among farms in the management of other species (e.g. manure management, assignment of workers dedicated to a specific species, etc.) were not observed during farm visits, although specific questions on such management practices were not included in our questionnaires. Our findings may reflect that Icelandic producers that raise domestic livestock in addition to broilers take precautions that prevent contamination of the broiler houses, such as increased efforts at biosecurity and sanitation practices.
An increased risk of Campylobacter was associated with increasing numbers of broiler houses on the farms. For each additional house on the farm, the risk of Campylobacter colonization increased by approximately 6% to 14%. Although we analysed this factor as a continuous variable, our finding is consistent with several other studies [12, 22, 24, 36]. There was a positive correlation (τb = 0.75, p < 0.001) between the number of houses on the farm and the number of flocks raised on the farm. To determine if the increased risk was indeed associated with increasing numbers of houses, rather than just increasing numbers of flocks, we included the number of flocks as an independent predictor in the models and found that the number of broiler houses remained statistically significant, while the number of flocks was not significant. Several houses on the same farm may lead to an increased risk of Campylobacter through the introduction of the bacteria into the house from the environment , possibly through the increased movement of farm workers between houses, or difficulty in maintaining strict hygiene or biosecurity practices. In general, broiler farm workers in Iceland are not specific to a house. However, on farms that have both breeder and broiler houses, workers are generally assigned to either the broiler or breeder houses and producers take precautions with any exceptions. Each broiler house has its own set of boots and clothing, and in most cases, there is a strict separation and physical barrier between the exterior personnel entry area (for removal of outside boots and clothing), and the inside clean area with dedicated broiler house boots, coveralls, and hand wash and disinfectant. However, given the increased risk associated with increasing numbers of houses, for new broiler farms, consideration should be given to limiting the number of houses built.
The practice of storing manure on the farm was associated with a decreased risk of colonization and was an unexpected finding. We considered that this protective effect may be a result of producers storing manure when there was not enough space to spread it in the immediate vicinity, however, a brief exploration of the interaction between manure storing and spreading showed that the two factors were independent (regardless of modelling approach). One possible theory for our finding is that manure stored in large piles (as is the practice in Iceland) may be subjected to a form of composting or fermentation, which may be detrimental to the survival of the organism. By contrast, spreading manure on fields in the winter season was associated with an increased risk of Campylobacter colonization, although it is unclear how this practice increases risk. The effect of spreading manure on fields in the summer season varied depending on the model. In the automated backward stepwise model, multicollinearity was a problem, thus, the protective effect may be a spurious result because of its strong positive association with spreading manure in the winter. There is very little information about these predictors in the literature, and it is uncertain whether these practices are unique to Iceland. In Senegalese broiler flocks, an elevated risk of Campylobacter colonization was associated with manure disposal inside the farm compared to disposal outside the farm, presumably through continual contamination of the environment , although the nature of disposal was not stated. Similar to our findings, in Québec, Canada, the presence of a manure heap ≤200 m from the broiler house (versus > 200 m) was associated with a decreased risk of colonization, although the authors considered that this unexpected finding was the result of confounding by farm size . Analysis of these risk factors in future studies, and studies that evaluate the survival of Campylobacter in manure under various environmental conditions, may substantially improve our understanding of the relationship between the farm environment and Campylobacter in broiler flocks.
A limitation of automated variable selection procedures is the potential for inclusion of strongly correlated variables in the model. In the automated backward stepwise model, the predictors "manure spread on fields in summer season" and "manure spread on fields in winter season" were both retained. The standard errors for these variables were slightly inflated (0.8 in this model compared to approximately 0.3 in other models) as a result of multicollinearity, therefore, the coefficients must not be over-interpreted. Notwithstanding this, these risk factors were significant in other models suggesting their importance in predicting the risk of Campylobacter on broiler farms in Iceland.
A second limitation of automated variable selection procedures is the inability to identify and evaluate potential confounding variables. In both automated forward models, the presence of other commercial poultry on the farm was associated with an increased risk of Campylobacter colonization. This finding is in agreement with one study , although other researchers [26, 34, 36] have not found an association. However, in both manual selection models, the presence of other commercial poultry was shown to be a confounder for most of the other predictors (including number of houses, farm water source, manure spreading and storing practices, and the presence of other livestock), and this accounts for the discrepancy between models. Sampling of sexually immature and parent breeder flocks in Iceland between May and July 2000, showed that up to 72% of faecal samples were positive for Campylobacter spp. , suggesting the potential for contamination of the farm environment from these other poultry. Our results show that after controlling for other farm-level factors, keeping other commercial poultry on the farm is not associated with the colonization of broiler flocks with Campylobacter in Iceland. However, in our classification of other poultry, we did not differentiate between those farms that raised turkey flocks and broiler flocks alternatively in the same house (with full cleaning and disinfection between flocks), from those farms that also have year round permanent breeder or egg layer flocks. With few exceptions, the latter tend to be constantly heavily contaminated Campylobacter reservoirs (based on sampling results of the on-going Icelandic surveillance program). Future studies should carefully classify other poultry on the farm in order to fully assess their impact on the risk of colonization of broiler flocks.
The variable "farm has an all-in-all-out policy" changed on one farm during the study. Since less than 80% of the flocks on this farm were subjected to this management practice, we deemed that the farm did not use an all-in-all-out system. In order to assess what effect this might have had on our models, we re-analysed the data using a repeated measures approach and found that the results were not affected. The repeated measures approach allowed the predictor to vary for different flocks raised on the same farm (i.e. flocks raised in the early part of the study were subjected to an all-in-all-out system, whereas, those in the latter part were not), and adjusted the standard errors to account for intragroup correlation. Regardless of statistical approach, this variable was one of the first to be removed in all of the backward elimination procedures and was not eligible for addition in any of the forward selection methods. Thus, in the Icelandic broiler industry, an all-in-all-out policy on the farm does not appear to be associated with Campylobacter colonization during the summer season. One possible explanation for this finding may be related to the changes in the broiler industry that took place following the epidemic in 1999 and the implicated role of fresh broiler chicken products. Broiler producers came under much pressure to reduce flock prevalence. A major emphasis was placed on heightened strict biosecurity rules on broiler farms, thorough cleaning and disinfection of houses between flocks, and pest control. Rigorous multi-step cleaning and disinfection of the live haul crates and trucks was also initiated. These initiatives began early in 2000. Freezing of products from all flock lots found positive on pre-slaughter sampling, and the price penalty to the producer for positive flock lots, ensured continued producer motivation to maintain high standards. This may have reduced the otherwise expected importance of an all-in-all-out system.
Fifteen percent of the farms in our study were excluded from the analysis due to missing data for one or more variables. In order to assess what effect this might have had on our results, we re-analysed the data using all 33 farms, excluding the four variables with missing data (an all-in-all-out policy, manure spreading in the summer season, manure spreading in the winter season, and manure storing). We found that, whether we used 33 farms or 28 farms in our models, our estimates for other domestic livestock on the farm, farm water source, and median flock size were consistent. However, when we used 33 farms, the presence of other commercial poultry and the number of houses did not remain in any of the backward elimination models. It was evident that there was confounding between the number of houses, the presence of other poultry, and manure spreading & storing practices on the farm. Therefore, by including manure management practices (and hence analysing data from fewer farms), we likely have better estimates for these potentially important risk factors for flock colonization at the farm level, and the impact of other variables appears stable.