The purpose of the Production Module is to estimate the prevalence of E. coli O157:H7-infected cattle entering US slaughter plants. The justification is that the prevalence of E. coli O157:H7 in slaughter cattle influences its occurrence on carcasses and ultimately in ground beef. At least three studies lend empirical support to the premise that infected cattle are a direct source of carcass contamination. Bonardi et al. (2001) and Chapman et al. (1993) reported an association between fecal positivity and carcass positivity at the level of the individual animal or carcass. Elder and colleagues (2000) found a correlation between combined fecal and hide prevalence of E. coli O157:H7 in groups of cattle (slaughter lots) and the prevalence on carcasses. An analysis of pulse-field gel electrophoresis profile isolates from the study by Elder et al. (2000) found high concordance between fecal and carcass isolates within slaughter lots (Barkocy-Gallagher et al., 2001). The common seasonal pattern, in temperate climates, of E. coli O157:H7 fecal prevalence in cattle (Hancock et al., 1997a; Heuvelink et al., 1998), retail meats (Chapman et al., 2001), and humans (Wallace et al., 2000) also lends credibility.
The utility of the estimates of fecal prevalence from the Production Module depends on the answers to two central questions that are discussed in detail below.
Is fecal prevalence alone an adequate measure of output for the Production Module?
Are the prevalence estimates in cull cows (called “breeding cattle”
in the Fod safety and Inspecton Service (FSIS) draft risk assessment) and feedlot animals defensible?
FECAL PREVALENCE AS THE SOLE OUTPUT OF THE PRODUCTION MODULE
The arguments against using fecal prevalence alone for risk assessment are related to the wide range of concentrations of E. coli O157:H7 in the feces of colonized cattle and the fact that E. coli O157:H7 occurs in locations other than feces.
On theoretical grounds, animals shedding 105 colony-forming units (CFU) of E. coli O157:H7 per gram of feces would cause much more contamination of meat than animals shedding, say, 102 CFU/g, but they are considered to contribute equally in a model that includes only prevalence. The issue might not be of concern if the distribution of pathogen concentrations were narrow or if one could assume a dependable relationship, at the group level, between prevalence and distribution of concentrations. However, the distribution clearly is not narrow. Experimentally infected animals often reach peak shedding concentrations over 105 CFU/g briefly and typically shed much lower numbers for longer periods (Cray and Moon, 1995; Sanderson et al., 1999). Some cattle that are naturally exposed to an infected animal never shed over 103 CFU/g (Besser et al., 2001). There is a paucity of data on shedding dynamics in field populations, but the seeming consequence of findings from challenge studies is that one would expect only a small fraction of positive animals in a group to be shedding E. coli O157:H7 at over 105 CFU/g on any given day. However, those few animals probably account for the large majority of total E. coli O157:H7 cells produced by the group. The disjunction between prevalence and quantity of E. coli O157:H7 shed has probably been magnified as tests have become more sensitive because, as documented by Sanderson et al. (1995) and Besser et al. (2001), the major impact of increased sensitivity of an assay is its ability to detect lower concentrations. For example, methods based on immunomagnetic separation (IMS) have allowed far better detection of animals shedding 102 CFU/g than older assays (Besser et al., 2001). But one animal shedding 105 CFU/g would yield the same number of E. coli O157:H7 cells as 1,000 animals shedding 102 CFU/g.
The use of fecal prevalence as the sole output of the Production Module requires the assumption that most carcass contamination with E. coli O157:H7 (or fecal bacteria in general) occurs directly from the gastrointestinal tracts of slaughtered animals. The draft defends that assumption by reference to a study showing little or no correlation between visible hide soiling and generic E. coli counts on carcasses (Van Donkersgoed et al.,
1997) and several studies suggesting that hide prevalence is lower than fecal prevalence. Admittedly, it is difficult to find studies that provide quantitative data on the source of bacteria contaminating carcasses, but circumstantial evidence suggests that the hide is a major source of carcass contamination (Castillo et al., 1998; Midgley and Desmarchelier, 2001). Visible soiling of hides and carcass contamination may correlate poorly because of the failure to distinguish between visible and microbiological hide soiling. That is, it may be that the level of microbiological contamination cannot be judged visually. The lower prevalence on hides than in feces may be due to the sampling of only a tiny fraction of the hide in the cited studies or to the lower sensitivity of IMS expected for an environmental sample. Only 450 cm2 of the hide was sampled, and sorbitol-negative bacteria resistant to cefixime and tellurite are more common in hide than in fecal samples and can interfere with identification of E. coli O157:H7. Moreover, although Elder et al. (2000) found 28% fecal prevalence versus 11% hide prevalence, the carcass prevalence was reported to be 43.4%. It seems unlikely that such a large percentage of carcasses were contaminated with feces directly from the rectum or with other gastrointestinal contents.
The use of fecal prevalence of E. coli O157:H7 as the output of the Production Module without consideration of the concentration of E. coli O157:H7 in feces or contamination of hides and hooves was seemingly necessitated by the paucity of data on anything beyond fecal prevalence. It appears that fecal prevalence is thus being used as a proxy variable to define several interrelated variables that are poorly understood and on which data are scarce. That is not an insurmountable problem, especially given the aforementioned studies demonstrating a correlation between fecal prevalence and carcass contamination. Use of fecal prevalence of E. coli O157:H7 alone does allow at least a crude assessment of the effect of farm-level interventions on the extent of ground-beef contamination, but it is possible to imagine that an intervention might reduce concentration, especially at the high end, and have little or no effect on prevalence. For example, if the reduction in prevalence resulted in fewer animals shedding 103 CFU/g or less but did not impact the prevalence of animals shedding >103, then the overall impact on amount of total number of E. coli O157:H7 cells shed by a group of animals would be negligible. Also, important sites for intervention, such as transport and confinement conditions or methods of hide removal, may be excluded from the draft risk assessment if fecal prevalence is the sole output of the Production Module. The exclusion of effects of preslaughter transport and lairage from the model may need re-examination in light of a report of increasing E. coli O157:H7 prevalence with increasing number of transit points (Cornell Collaborative Project, 1998). Although only
a small minority of cattle had multiple transit points, such uncommon occurrences may have a substantial effect on E. coli O157:H7 contamination of ground beef. Furthermore, a paper by Midgley and Desmarchelier (2001) documents the occurrence of a subtype of E. coli O157:H7 (and also shigatoxic O26:H11) on hides of animals sampled at slaughter that had never been observed on intensive sampling during the feeding period of these animals; this suggests that the subtype had been acquired during confinement at the slaughter plant. Although the sample size evaluated by Midgely and Desmarchelier was too small to allow a judgment of the quantitative impact of lairage, more recent papers by Small et al. (2002) and Avery et al. (2002) suggest that the lairage effect might be substantial. On the other hand, another recent paper (Barham et al., 2002) reported a significant decline in prevalence from the feedlot to the slaughter plant.
For those reasons, the committee recommends that the final risk assessment acknowledge forthrightly that fecal prevalence is being used as a proxy variable and that some carcass contamination is derived from hides.
DEFENSIBILITY OF PREVALENCE ESTIMATES FOR CULL COWS AND FEEDLOT ANIMALS
Pooling data from disparate studies that had differing assay methods and sampling designs is difficult, and the draft risk assessment does a generally credible job. However, several issues related to adjustment for imperfect sensitivity of tests and to estimation of within-herd prevalence, herd prevalence, and seasonal effects merit scrutiny.
Issues Related to Adjustment for Test Imperfections
Possible Specificity Problems
Theoretically, errors of two kinds can be made by a test: false negatives (imperfect sensitivity) and false positives (imperfect specificity). The draft risk assessment rightly devotes considerable attention to false negatives in its discussion of test sensitivity, but the potential for false positives is not raised. The most likely reasons for false positives are the incorrect identification of E. coli O157:H7 and cross contamination. While the former is not an evident problem in any of the studies cited, the latter is a potential problem, especially in studies that use IMS. Cross contamination of IMS samples is a problem that is to be expected in the absence of specific controls (PHLS, 2000). The two key issues for avoiding false positives are the use of blank tubes in each run and the capping of all tubes
except the one tube being immediately processed (PHLS, 2000). Unless the authors of a paper state (or, if asked, can state) that those procedures were adhered to and that blanks were uniformly negative, it must be assumed that false positives probably occurred. In any case, it seems unreasonable to go to considerable lengths to address imperfect test sensitivity while failing to note the possibility of imperfect test specificity. The committee suggests that the final risk assessment note that imperfect specificity was not assessed but may have had an impact on the output of the Production Module.
Estimating Prevalence of E. coli O157:H7 in Feces Requires Specification of Target Detection Limit
In contrast with the apparent assumption of Table 3-4 in the FSIS draft risk assessment, no procedure is capable of detecting one E. coli O157:H7 organism in a sample. As shown by Sanderson et al. (1995), every E. coli O157:H7 assay procedure has a 50% detection end point—the concentration at which half the positives are detected and half are missed—that is substantially greater than one organism per sample (or per gram). Every procedure (including IMS, in contrast with the draft risk assessment’s assumption of 100% sensitivity) has false negatives, even for occasional samples with concentrations of several powers of 10 above the 50% detection end point. It is also clear that as one collects more volume of feces and performs more replicate tests, the probability of detecting E. coli O157:H7 increases asymptotically in a fashion that has no obvious end point short of one cell per daily fecal output. If one organism per daily fecal output is considered an absurd boundary separating positive from negative results, one is left to define the reasonable boundary. Any such attempt will confront unresolvable arguments of whether it is colonization, infection, or simple shedding (including passive) that one wants to measure and whether a level of fecal shedding can even separate these states. The lack of any reasonable gold standard for E. coli O157:H7 fecal testing and the seemingly unfathomable problem of selecting an end point fecal concentration combine to make a convincing case for using the concentration of fecal shedding of E. coli O157:H7, rather than simple prevalence, as an output of the Production Module. That is so because the increasing numbers of animals shedding asymptotically lower levels of E. coli O157:H7 contribute little to the overall amount of E. coli O157:H7 shed into the environment (and ultimately onto carcasses). If one accounts for animals shedding E. coli O157:H7 at over 103 CFU/g, one has surely accounted for over 99% of total E. coli O157:H7 shedding. Although data on the concentration of fecal E. coli O157:H7 shedding are sparse, they should be sufficient to allow estimation of the relative effect of high-level shedders (say,
over 104 CFU/g) compared with low-level shedders. The committee recommends that the risk assessment provide an impact assessment of animals shedding E. coli O157:H7 at high and low levels. The committee notes that there is a paucity of information on this topic1 and suggests that the risk assessment highlight the need for more research.
Issues Related to Computation of Within-Herd Prevalence
Data from Juvenile Cattle May Have Been Included
Two of the within-herd prevalence estimates in Table 3-2 of the FSIS draft included juvenile animals (Besser et al., 1997; Hancock et al., 1994), whereas the intent was to estimate prevalence in adult animals. It is possible that only data from adults were extracted from the studies to obtain the tabulated values; if so, this should be made clear. The data on the distribution of within-herd prevalence shown in Figure 3-3 of the draft were collected exclusively from juvenile animals and should not be used to estimate the distribution of within-herd prevalence in adult animals; a number of studies have shown a higher prevalence of E. coli O157:H7 in juvenile than in adult cattle (Hancock et al., 1997a; Heuvelink et al., 1998). However, it is acceptable (even desirable) to include data on juveniles in determining herd prevalence, as was done in draft’s Table 3-1. The committee recommends that the risk assessment not use data on juvenile animals to estimate within-herd prevalence in adult animals.
Distribution of Within-Herd Prevalence Ignores Temporal Clustering Effect
Within a herd, the prevalence of E. coli O157:H7 seems to be unevenly distributed in time; bursts of prevalence occur periodically (Hancock et al., 1997a; Sargeant et al., 2000). Thus, a prevalence study in which each farm is sampled on a single occasion or a small number of occasions would be expected to find a wide distribution of within-herd prevalence even if long-term within-herd prevalence in the herd was identical. Indeed, a distribution much like those in Figures 3-5 and 3-10 of the FSIS draft risk assessment could be generated from multiple sampling visits to a single herd. Only by pooling data across many herd visits—as was done in the study by Hancock et al. (1997b), on which Figure 3-3 is based—can one obtain a reasonable picture of distribution of within-herd prevalence. However, it could be argued that the goal of risk assessment is best met by estimating the daily distribution of within-
herd prevalence (because cull cows are shipped on particular days rather than on average days). If that argument is accepted, a distribution based on pooling of multiple sampling visits, such as that portrayed in Figure 3-3, should not be used. The committee recommends that a decision be made as to whether distribution of within-herd prevalence by herd-day or by herd is more appropriate for the model and that only studies relevant to the chosen metric be used.
Estimates Are Biased by Use of Within-Herd Prevalence Data Only in Positive Herds
Use of data only on herds detected as positive (in contrast with herds actually positive), especially if samples per herd are relatively small, results in a biased estimate of within-herd prevalence and its distribution. The effect is most evident in studies with very small samples. If, for example, only five samples were collected per herd in a large number of herds, the herds detected as positive would be estimated to have a minimum of 20% prevalence even if the true within-herd prevalence were 1%. Although no studies with a within-herd sample size of five were included in Table 3-2 in the draft, the effect exists even with larger samples. It is evident in considering the extremely high estimates from Hancock et al. (1994) shown in Table 3-5 in the FSIS draft risk assessment (the estimate in Table 3-2 adjusted for sensitivity). If a very insensitive culture method was used, the sample size of 60 per herd in the Hancock et al. paper was not adequate to detect positive herds reliably. That expectation was confirmed by a follow-up study in which four of eight herds initially found to be negative were later found to be positive after more-intensive sampling (Hancock et al., 1997a). An accurate estimate of within-herd prevalence cannot be reasonably made from small samples unless one makes a priori assumptions about herd prevalence. That is, the denominator of the within-herd prevalence estimate would need to include not only the number of sampled animals from herds that were detected as positive, but also those from herds that were truly positive but, because of inadequate sample size, were not detected as positive. Practically speaking, that cannot be done unless one assumes 100% herd prevalence, which, on the basis of intensive sampling studies, is probably closer to reality than what has been depicted in the draft risk assessment. The committee recommends that the FSIS draft risk assessment either compute within-herd prevalence estimates as the total positives divided by the total sampled (herd status notwithstanding) or use a denominator based on the estimated herd prevalence, such as that depicted in Figure 3-2 of the draft.
Stage on Feed Was Not Considered for Feedlot Data
The largest study thus far published on E. coli O157:H7 in feedlot cattle (Hancock et al., 1997c) found a prevalence about three times higher in early-on-feed cattle (those that have recently entered the feed lot) than in late-on-feed cattle (those that will soon be slaughtered). In multivariate modeling of the data (Dargatz et al., 1997), the effect was found to exist in all regions sampled and for two independent sets of data assayed at two laboratories. If the goal is to estimate prevalence in slaughter cattle (late-on-feed), it is inadvisable to include data from cattle at all stages on feed. The committee recommends that the modelers adjust the estimate for prevalence in feedlot animals to that expected for preslaughter (late-on-feed) animals.
Prevalence Data on All Adults in Herd Might Not Yield a Good Estimate of Cows Soon to Be Culled
At least two studies (Garber et al., 1999; Rice et al., 1997) have reported higher prevalence in dairy cows identified for culling (but still on the farm) than in adult herdmates. Hence, the use of the within-herd prevalence in all adults may yield a biased estimate of prevalence of cull dairy cattle at the time of slaughter. The committee suggests that the risk assessment note as a possible weakness that prevalence estimates in cull cattle might be higher than those in all adult cattle.
Issues Related to Computation of Herd Prevalence
Elder et al. Data on Slaughter Lots Were Used Inappropriately
Data from Elder et al. (2000) were used in computing herd-prevalence in feedlots (Table 3-6 in the draft). The study did not sample individual feedlots, but rather slaughter lots—groups of 35–85 animals from a single source presented for slaughter on a particular day. It is not clear that all 29 of such lots sampled were from independent sources (that is, from 29 separate feedlots). If the Elder et al. data are derived from a relatively small number of separate feedlots, then the calculation of feedlot (herd) prevalence and the assumption that the sample is representative of feedlots in general become problematic. Beyond these simple mathematical issues lies a larger one that raises concerns regarding Elder and colleagues’ use of post exsanguination samples from slaughter lots to estimate feedlot prevalence. The committee identifies two questions for the consideration of the modelers:
Can one accurately determine if a feedlot is positive or negative for E. coli O157:H7 by taking a highly clustered sample of a relatively small number of the cattle in a feedlot?
Can one accurately make this estimate from a set of samples collected after exsanguination and hence after contact with pens and equipment of a slaughter plant that have been reported to be an important source of E. coli O157:H7 contamination (Avery et al., 2002; Small et al., 2002)?
The committee recommends that the risk assessment use only data from independent feedlots to estimate herd prevalence in feedlots.
Herd-Prevalence Estimates Erroneously Assume Homogeneous Prevalence over Time in Positive Herds
Computation of the herd-level sensitivity (Equation 3.3 in the draft) and thus herd prevalence appears to depend on an assumption of homogeneous within-herd prevalence. The strong temporal clustering observed for E. coli O157:H7 in a herd (Hancock et al., 1997a; Sargeant et al., 2000) invalidates the use of the equation. Consider that a herd with an average within-herd prevalence of, say, 5% will often, on any random sampling occasion, have much less than this prevalence and will occasionally have a much higher prevalence. In a study that uses only one or a few sampling points per herd, one will intercept many of the herds at a within-herd prevalence much below the long-term average, hence invalidating computations like those in Equation 3.3 (which is more appropriate for relatively stable measures, such as seroprevalence). The issue becomes more critical if some of the sampling visits are in winter, when underlying prevalence is lower (Hancock et al., 1994). It is worth noting that the lowest estimate of herd prevalence in the draft’s Table 3-4 was derived from the study with the fewest sampling points (N = 1 to 3; Garber et al., 1999) whereas studies that used more sampling visits per herd found much higher herd prevalences. Stated differently, the distribution of uncertainty in breeding-herd prevalence (Figure 3-2 in the draft risk assessment) was biased downward by use of an inappropriate means of computing herd sensitivity. The problem is more serious when one considers the feedlot data summarized in Table 3-6 in the draft, where the two studies reporting less than 100% herd prevalence involved only one sampling point per feedlot (Dargatz et al., 1997; Elder et al., 2000). On a theoretical basis it is difficult to imagine how any feedlot that receives cattle and feeds from many sources would not be at least intermittently positive for E. coli O157:H7. Inasmuch as the data in Table 3-6 are consistent with 100% herd prevalence in feedlots, one should probably use 100% as the estimate. In
Bayesian terms, there are good a priori reasons to believe that the herd prevalence in feedlots would be 100%, and, unless the data were inconsistent with this, 100% should be used. It is even possible to invoke this argument for breeding herds, but the lower incoming animal traffic and the failure, in some herds, to find any E. coli O157:H7 after collecting hundreds of samples at many times suggest that some small percentage of breeding herds remain free of E. coli O157:H7 over long periods.
The committee recommends that the risk assessment use an appropriate means of adjusting for herd sensitivity that incorporates effects of temporal clustering for breeding herds or base the estimate of herd prevalence only on studies in which breeding herds were sampled multiple times. For feedlots, the committee recommends that a 100% herd prevalence be used.
Issues Related to Estimation of Seasonal Effects on Prevalence
There are good reasons to provide a seasonal adjustment of the prevalence output from the Production Module. First, E. coli O157:H7-associated disease in humans is strongly seasonal (Wallace et al., 2000), and at least one study has attributed this to seasonal variation in prevalence of E. coli O157:H7 in retail beef and lamb meat (Chapman et al., 2001). Several longitudinal studies in dairy farms and feedlots have provided evidence of a marked seasonal effect (Hancock et al., 1997a; Heuvelink et al., 1998; Mechie et al., 1997; NAHMS, 2001), and some point-sampling studies corroborate it (Bonardi et al., 2001; Hancock et al., 1994; Van Donkersgoed et al., 1999). However, the seasonal prevalences estimated in the draft risk assessment are not the best ones possible.
Two technical mistakes were made in the data depicted in Table 3-5 in the draft:
Besser et al. (1997) was erroneously used instead of the companion paper, Hancock et al. (1997a). Since the former did not examine seasonal effect, the weighted averages are erroneously displayed as an adjusted 4.5% prevalence throughout the year (presumably on the assumption that if season was not mentioned there must not have been a seasonal effect). In Hancock et al. (1997a)—the companion paper on the same herds in which the effect was examined—a strong seasonal effect was reported.
In two of the studies (Hancock, 2001; Sargeant et al., 2000), the monthly estimates do not appear to have been adjusted for test sensitivity, although these adjustments were made in other studies.
Beyond those technical mistakes, the methods by which evidence from various longitudinal and point-prevalence studies were combined merit
close scrutiny. The problem is most evident in the handling of data from Hancock et al. (1994). In that study, 60 dairy herds were sampled one time each, but sampling was distributed roughly equally in each of the 12 calendar months—that is, about 300 animals in five herds were sampled in each calendar month. But no data (numerator or denominator) were included in Table 3-5 from herds or months in which positives were not found, presumably on the assumption that one would never observe a herd prevalence of zero on a sampling visit (say, in winter) to a truly positive herd. That assumption is not valid, as shown by Hancock et al. (1997a) and Sargeant et al. (2000). In both those studies, most sampling visits to positive herds were associated with completely negative results. If data from point-sampling studies (such as Garber et al., 1999; Hancock et al., 1994; Rice et al., 1997) are to be used in computing seasonal effect, the only reasonable way to treat them is as generally random surveys of the cattle population (ignoring herd, because herd status cannot be assessed accurately in such studies). Hence, one would not put the same prevalence in each month for point-sampling studies where no clear-cut seasonal effect was found. Rather, one would use the observed prevalence computed as the total positives found divided by the total samples tested, regardless of whether samples came from herds that were, by chance, found to contain positive animals or herds in which no positives were found. It is worth noting that when the data in Hancock et al. (1994) are treated in this manner, a seasonal effect is observed that is similar to that found in a year-long longitudinal study of a subset of the same herds (Hancock et al., 1997a).
Another possible approach that is worth considering is to separate the estimation of prevalence from the estimation of seasonal effect completely. A seasonal adjustment (multiplier) would be made to prevalence estimates from various studies in a manner similar to that for the adjustments for test sensitivity. That would allow one to restrict the computation of seasonal adjustment factors to longitudinal studies done over a year’s period—that is, the ones that are most appropriate to the unbiased estimation of seasonal effect. It would also potentially allow inclusion of high-quality longitudinal studies from areas outside the United States, such as that by Heuvelink et al. (1998). It is arguable that conditions peculiar to European cattle production might result in differences in overall average prevalence, but it seems more reasonable to assume that the magnitude of seasonal differences would be similar to that experienced by cattle herds in similar climatologic areas of the United States. Given the paucity of longitudinal studies in the United States, especially those using modern detection technology, the use of data from other countries is worth considering in the estimation of seasonal effects.
The committee recommends the following changes for estimating seasonal effects:
Use data from Hancock et al. (1997a) instead of Besser et al. (1997).
Adjust all monthly prevalence estimates for imperfect test sensitivity.
Either handle the data from multiple surveys as random surveys of the cattle population, thus using data on all cattle sampled in each month, or use only data from longitudinal studies to estimate seasonal adjustment factors.
Avery SM, Small A, Reid CA, Buncic S. 2002. Pulsed-field gel electrophoresis characterization of Shiga toxin-producing Escherichia coli O157 from hides of cattle at slaughter. Journal of Food Protection 65(7):1172–1176.
Barham AR, Barham BL, Johnson AK, Allen DM, Blanton JR, Miller MF. 2002. Effects of the transportation of beef cattle from the feedyard to the packing plant on prevalence levels of Escherichia coli O157 and Salmonella spp. Journal of Food Protection 65(2):280–283.
Barkocy-Gallagher GA, Arthur TM, Siragusa GR, Keen JE, Elder RO, Laegreid WW, Koohmaraie M. 2001. Genotypic analyses of Escherichia coli O157:H7 and O157 nonmotile isolates recovered from beef cattle and carcasses at processing plants in the Midwestern states of the United States. Applied Environmental Microbiology 67(9):3810–3818.
Besser TE, Hancock DD, Pritchett LC, McRae EM, Rice DH, Tarr PI. 1997. Duration of detection of fecal excretion of Escherichia coli O157:H7 in cattle. Journal of Infectious Diseases 175:726–729.
Besser TE, Richards BL, Rice DH, Hancock DD. 2001. Escherichia coli O157:H7 infection of calves: Infectious dose and direct contact transmission. Epidemiology and Infection 127(3):555–560.
Bonardi S, Maggi E, Pizzin G, Morabito S, Caprioli A. 2001. Faecal carriage of verocytotoxinproducing Escherichia coli O157 and carcass contamination in cattle at slaughter in north-ern Italy. International Journal of Food Microbiology 66(1-2):47–53.
Castillo A, Dickson JS, Clayton RP, Lucia LM, Acuff GR. 1998. Chemical dehairing of bovine skin to reduce pathogenic bacteria and bacteria of fecal origin. Journal of Food Protection 61(5):623–625.
Chapman PA, Siddons CA, Wright DJ, Norman P, Fox J, Crick E. 1993. Cattle as a possible source of verocytotoxin-producing Escherichia coli O157 infections in man. Epidemiology and Infection 111(3):439–447.
Chapman PA, Cerdan Malo AT, Ellin M, Ashton R, Harkin. 2001. Escherichia coli O157 in cattle and sheep at slaughter, on beef and lamb carcasses and in raw beef and lamb products in South Yorkshire, UK. International Journal of Food Microbiology 64(1-2):139–150.
Cornell Collaborative Project. 1998. Prevalence of E. coli O157:H7 Pre-slaughter/E. coli O157:H7 Associated Risk Factors. New York State Cull Cow Sub-Project. Final report.
Cray WC, Moon HW. 1995. Experimental infection of calves and adult cattle with E. coli O157:H7. Applied Environmental Microbiology 61:1586–1590.
Dargatz DA, Wells SJ, Thomas LA, Hancock DD, Garber LP. 1997. Factors associated with the presence of Escherichia coli O157 in feces in feedlot cattle. Journal of Food Protection 60:466–470.
Elder, RO, Keen JE, Siragusa GR, Barkocy-Gallagher GA, Koomaraie M, Laegreid WW. 2000. Correlation of enterohemorrhagic E. coli O157 prevalence in feces, hides, and carcasses of beef cattle during processing. Proceedings of the National Academy of Sciences 97(7):2999–3003.
Garber L, Wells S, Schroeder-Tucker L, Ferris K. 1999. Factors associated with the shedding of verotoxin-producing Escherichia coli O157 on dairy farms. Journal of Food Protection 62(4):307–312.
Hancock DD. 2001. Personal communication. Ongoing research project in collaboration with FDA-CVM.
Hancock DD, Besser TE, Kinsel ML, Tarr PI, Rice DH, Paros MG. 1994. The prevalence of Escherichia coli O157 in dairy and beef cattle in Washington State. Epidemiology and Infection 113:199–207.
Hancock DD, Besser TE, Rice DH, Herriott DE, Tarr PI. 1997a. A longitudinal study of Escherichia coli O157 in fourteen cattle herds. Epidemiology and Infection 118:193–195.
Hancock DD, Rice DH, Herriot DE, Besser TE, Ebel ED, Carpenter LV. 1997b. Effects of farm manure handling practices on Escherichia coli O157 prevalence in cattle. Journal of Food Protection 60(4):363–366.
Hancock DD, Rice DH, Thomas LA, Dargatz DA, Besser TE. 1997c. Epidemiology of Escherichia coli O157 in feedlot cattle. Journal of Food Protection 60(5):462–465.
Heuvelink AE, Van den Biggelaar FL, Zwartkruis-Nahuis J, Herbes RG, Huyben R, Nagelkerke N, Melchers WJ, Monnens LA, de Boer E. 1998. Occurrence of verocytotoxin-producing Escherichia coli O157 on Dutch dairy farms. Journal of Clinical Microbiology 36(12):3480– 3487.
Mechie SC, Chapman PA, Siddons CA. 1997. A fifteen month study of E. coli O157:H7 in a dairy herd. Epidemiology and Infection 118(1):17–25.
Midgley J, Desmarchelier P. 2001. Pre-slaughter handling of cattle and Shiga toxin-producing Escherichia coli (STEC). Letters in Applied Microbiology 32(5):307–311
NAHMS (National Animal Health Monitoring System). 2001. Escherichia coli O157 in United States Feedlots. USDA: Animal and Plant Health Inspection Service Bulletin #N345.1001.
PHLS (Public Health Laboratory Service). 2000. Laboratory cross contamination of Escherichia coli O157 in a sample of organic mushrooms. Public Health Laboratory Service (England and Wales), November.
Rice DH, Ebel ED, Hancock DD, Herriott DE, Carpenter LV. 1997. Escherichia coli O157 in cull dairy cows on farm and at slaughter. Journal of Food Protection 60(11):1386–1387.
Sanderson MW, Gay JM, Hancock DD, Gray CC, Fox LK, Besser TE. 1995. Sensitivity of bacteriologic culture for detection of Escherichia coli O157:H7 in bovine feces. Journal of Clinical Microbiology 33(10):2616–2619.
Sanderson MW, Besser TE, Gay JM, Gay CC, Hancock DD. 1999. Fecal Escherichia coli O157:H7 shedding patterns of orally inoculated calves. Veterinary Microbiology 69(3):199–205.
Sargeant JM, Gillespie JR, Oberst RD, Phebus RK, Hyatt DR, Bohra LK, Galland JC. 2000. Results of a longitudinal study of the prevalence of E. coli O157:H7 on cow-calf farms. American Journal of Veterinary Research 61(11):1375–1379.
Small A, Reid CA, Avery SM, Karabasil N, Crowley C, Buncic S. 2002. Potential for the spread of Escherichia coli O157, Salmonella, and Campylobacter in the lairage environment at abattoirs. Journal of Food Protection 65(6):931–936.
Van Donkersgoed J, Jericho KWF, Grogan H. 1997. Preslaughter hide status of cattle and the microbiology of carcasses. Journal of Food Protection 60(12):1502–1508.
Van Donkersgoed J, Graham JT, Gannon V. 1999. The prevalence of vertoxins, E. coli O157:H7 and Salmonella in the feces and rumen of cattle at processing. Canadian Veterinary Journal 40:332–338.
Wallace DJ, Van Gilder T, Shallow S, Fiorentino T, Segler SD, Smith KE, Shiferaw B, Etzel R, Garthright WE, Angulo FJ. 2000. Incidence of foodborne illnesses reported by the foodborne diseases active surveillance network (FoodNet)-1997. Journal of Food Protection 63(6):807–809.
Zhao T, Doyle MP, Shere J, Garber L. 1995. Prevalence of enterohemorrhagic Escherichia coli O157:H7 in a survey of dairy herds. Applied and Environmental Microbiology 61(4):1290–1293.