Attribution of human infections with Shiga toxin-producing Escherichia coli (STEC) to livestock sources and identification of source-specific risk factors, The Netherlands (2010–2014)

Summary
Shiga toxin-producing Escherichia coli (STEC) is a zoonotic pathogen of public health concern whose sources and transmission routes are difficult to trace. Using a com- bined source attribution and case–control analysis, we determined the relative contri- butions of four putative livestock sources (cattle, small ruminants, pigs, poultry) to human STEC infections and their associated dietary, animal contact, temporal and socio-econo-demographic risk factors in the Netherlands in 2010/2011–2014. Dutch source data were supplemented with those from other European countries with simi- lar STEC epidemiology. Human STEC infections were attributed to sources using both the modified Dutch model (mDM) and the modified Hald model (mHM) supplied with the same O-serotyping data. Cattle accounted for 48.6% (mDM) and 53.1% (mHM) of the 1,183 human cases attributed, followed by small ruminants (mDM: 23.5%; mHM: 25.4%), pigs (mDM: 12.5%; mHM: 5.7%) and poultry (mDM: 2.7%; mHM: 3.1%), whereas the sources of the remaining 12.8% of cases could not be attributed. Of the top five O-serotypes infecting humans, O157, O26, O91 and O103 were mainly at- tributed to cattle (61%–75%) and O146 to small ruminants (71%–77%). Significant risk factors for human STEC infection as a whole were the consumption of beef, raw/un- dercooked meat or cured meat/cold cuts. For cattle-attributed STEC infections, spe- cific risk factors were consuming raw meat spreads and beef. Consuming raw/ undercooked or minced meat were risk factors for STEC infections attributed to small ruminants. For STEC infections attributed to pigs, only consuming raw/undercooked meat was significant. Consuming minced meat, raw/undercooked meat or cured meat/ cold cuts were associated with poultry-attributed STEC infections. Consuming raw vegetables was protective for all STEC infections. We concluded that domestic rumi- nants account for approximately three-quarters of reported human STEC infections, whereas pigs and poultry play a minor role and that risk factors for human STEC infec- tion vary according to the attributed source.

1| INTRODUCTION
Shiga toxin-producing Escherichia coli (STEC) is a bacterial zoonotic agent whose infection in humans is associated with varying clinical manifestations, including diarrhoea, haemorrhagic colitis and (occa- sionally fatal) haemolytic uremic syndrome (HUS) (Karmali, Gannon, & Sargeant, 2010). Shiga toxin-producing Escherichia coli is a pathogen of public health concern given its recognized potential to cause large food- and waterborne outbreaks, as well as its association with HUS (Buchholz et al., 2011; Matsell & White, 2009; Nguyen et al., 2015), a leading cause of acute renal failure among children (Keir, 2015).While most STEC strains associated with human illness belong to serogroup O157, there are more than a hundred of non-O157 sero- types (Gould et al., 2013), some of which have been associated with large outbreaks of severe illness, such as the German O104:H4 out- break linked to contaminated fenugreek sprouts in 2011 (Buchholz et al., 2011), whereas others are associated with only mild or no illness at all (Coombes, Gilmour, & Goodman, 2011). Although this can be directly related to the virulence profile of the individual STEC strains, it also seems that there is no specific molecular pattern that would unambiguously enable hazard identification for any STEC strain (Franz et al., 2015). The occurrence of certain STEC strains among clinical cases may not only reflect differences in virulence, but also differences in the level of exposure to specific animal sources. However, the rela- tive contributions of different animal reservoirs to the human disease burden for STEC, as well as their transmission routes to humans, have not yet been quantified.

In 2012, 5748 confirmed STEC cases were reported by 28 European Union (EU) and European Economic Area (EEA) countries, resulting in an overall notification rate of 1.5 cases per 100,000 population (European Centre for Disease Prevention and Control, 2014). After Germany and the United Kingdom, the Netherlands had the third highest number of STEC cases reported in the EU/EEA during that year, that is, 1049 cases corresponding to 6.3 cases per 100,000 population (European Centre for Disease Prevention and Control, 2014), although approximately only half of these cases had clinical relevance (i.e., had an acute onset of symptoms or required hospitalization).Animals capable of maintaining STEC carriage in the absence of continuous exposure to STEC (reservoirs or amplifying hosts), as well as those that are frequently exposed to STEC, from the environment for instance, can serve as potential sources of human STEC infection (Persad & LeJeune, 2014). Although ruminants, and particularly cat- tle, are regarded as the main reservoir for STEC (Caprioli, Morabito, Brugere, & Oswald, 2005; Karmali et al., 2010), there is evidence for some (wild) birds, pigs, dogs and horses being significant spill-over hosts for STEC (i.e., animals that are susceptible to colonization by STEC but do not maintain such colonization in the absence of contin- uous exposure (Persad & LeJeune, 2014)). Implicitly, this means that there may be other epidemiologically relevant sources of human STEC infection beyond ruminants.To identify the main sources of bacterial zoonotic infections and to assess the impact of public health interventions, source attribution using the microbial subtyping approach is often the method of choice (Pires et al., 2009). Given the source specificity of certain subtypes of the pathogen in question and assuming a unidirectional transmission pathway from sources to humans (with humans representing the endpoint), the relative contribution of each source to human cases can be inferred probabilistically by comparing the pathogen subtype distributions in humans and sources.

While source attribu- tion studies for pathogens like Salmonella, Campylobacter and Listeria based on the microbial subtyping approach have been performed in many countries worldwide (Barco, Barrucci, Olsen, & Ricci, 2013; Boysen et al., 2014; David et al., 2013; Guo et al., 2011; de Knegt, Pires, & Hald, 2015; Levesque et al., 2013; Little, Pires, Gillespie, Grant, & Nichols, 2010; Mossong et al., 2016; Mughini-Gras & van Pelt, 2014; Mughini-Gras et al., 2012; Mughini-Gras, Barrucci, et al. 2014; Mughini-Gras, Enserink, et al., 2014; Mughini-Gras, Smid, et al., 2014; Mullner, Jones, et al., 2009; Mullner, Spencer, et al., 2009; Pires & Hald, 2010; Pires et al., 2009; Sheppard et al., 2009; Strachan et al., 2009; Wahlstrom, Andersson, Plym-Forshell, & Pires, 2011; Wilson et al., 2008), no comparable study on STEC has been performed so far. Moreover, to understand how STEC strains originating from specific animal sources infect humans (i.e., to uncover the underlying transmission routes), combined (source- assigned) case–control and source attribution analyses have been performed for Salmonella (Mughini-Gras, Enserink, et al., 2014) and Campylobacter (Mossong et al., 2016; Mughini-Gras et al., 2012) to bridge the gap between attributing human cases at the start of the transmission chain (i.e., reservoir level) and at the point of exposure (i.e., risk factors) to identify source-specific risk factors for human infection.Focussing on a country like the Netherlands, the aim of this study was to attribute human cases of STEC infection to four putative live- stock sources (i.e., cattle, small ruminants, pigs and poultry) and to de- termine how the STEC strains attributed to each specific source may infect humans.

2| MATERIALS AND METHODS
We used national surveillance data on human STEC infections that occurred in the general population of the Netherlands and were re- ported to the Dutch National Institute for Public Health and theSTEC infections from the enhanced surveillance in 2012 among 21 European countries. The results of this surveillance are available in the most recent (at the time of analysis) annual epidemiological report of food-and-waterborne diseases and zoonoses published in 2014 by the European Centre for Disease Prevention and Control (2014). The PSI is expressed as:detailed description of this surveillance system has been presented before (Friesema, Schotsborg, Heck, & Van Pelt, 2015). In total, 1183 O-serotyped STEC isolates from human cases (corresponding to 91 different O-serotypes) were obtained during this period. Concurrent (2011–2014) O-serotyped STEC isolates from four potential livestock sources of STEC, that is, cattle (n = 207), sheep and goat (i.e., small ruminants, n = 98), pigs (n = 109) and poultry (n = 30) were collected by the Netherlands’ Food and Consumer Product Safety Authority (NVWA) as part of national surveillance activities (i.e., official sam- pling) of foodborne pathogens on slaughterhouse and retail (car- casses and meat samples). The collection of multiple isolates by herd or retailer cannot be excluded, for example if samples from the same source were sampled at different sites (slaughterhouse and retail) or yielded isolates with different serotypes. Both human and animal iso- lates were serotyped as described elsewhere (Guinee, Agterberg, & Jansen, 1972; Orskov, Orskov, Jann, & Jann, 1977; Paton & Paton, 1998).To increase statistical power and serotype diversity, the Dutch livestock isolates were supplemented with others (440 from cattle, 239 from small ruminants, 28 from pigs and 61 from poultry) from 10 European countries by collating available data on reported O- serotyped STEC isolates per livestock source from the European Union Summary Reports on Trends and Sources of Zoonoses, Zoonotic Agents and Food-Borne Outbreaks, as published annually by the European Food Safety Authority (EFSA) from 2011 to 2014 (European Food Safety Authority and European Centre for Disease Prevention and Control, 2013, 2014, 2015a,b).

A summary of the isolates used in this study is given in Table 1. The non-Dutch isolates originated from official sampling strategies at the levels of slaughterhouse and retail like the Dutch ones. Selection of the non-Dutch isolates was based on the method proposed by (Smid et al., 2013). This method is based on the assumption that if the subtype frequency distribution of the human isolates of a zoonotic pathogen in one country resembles that of the human isolates in another country, then also those of their respec- tive animal sources can be used interchangeably for the purposes of source attribution (Smid et al., 2013). Using this method, we implicitly assumed that the epidemiology of STEC in the Netherlands, including the food consumption patterns and exposure pathways from livestockto humans, is similar to that in the countries from which the supple-where |Pk − Qk| is the absolute value of the difference in the relative frequency of the serotype k in group P compared to its frequency in group Q. Proportional similarity index values range from 0 to 1, with 0 indicating that the two distributions have no serotypes in common and 1 that they are equal. The countries listed in Table 1, whose live- stock isolates were used in the source attribution analysis, were those with the O-serotype frequency distributions among human infections showing the highest PSI values (>0.50) as compared to the O-serotype frequency distribution among human infections in the Netherlands. Thus, they were the countries assumed to have the livestock sources of STEC most similar to those in the Netherlands.In absence of a universally agreed source attribution approach for STEC, two commonly used frequency-matching models for source at- tribution of zoonotic pathogens were applied in a comparative fash- ion: the modified Dutch model (mDM) and the modified Hald model (mHM), with some minor adaptations as compared to the numerous applications of these two models described previously for Salmonella and Campylobacter (Mughini-Gras & van Pelt, 2014; Mughini-Gras, Barrucci, et al., 2014; Mughini-Gras, Enserink, et al., 2014; Mughini- Gras, Smid, et al., 2014; Mullner, Jones, et al., 2009; Mullner, Spencer, et al., 2009).

Both models were based on O-serotyping data of the human (the Netherlands only) and livestock STEC isolates summa- rized in Table 1, and veterinary data were pooled over years and countries to avoid data sparsity and increase certainty in the esti- mates. Moreover, given the evidenced importance of direct/indirect exposure to livestock, especially cattle, in the occurrence of human STEC infections in the Netherlands (Friesema, Van De Kassteele, de Jager, Heuvelink, & van Pelt 2011), both models did not include food consumption weights to allow also non-foodborne pathways to be re- flected in the attributions, as suggested elsewhere for Campylobacter (Mullner, Jones, et al., 2009; Mullner, Spencer, et al., 2009). Briefly, in the version of the mDM used here, the expected number of human STEC isolates with O-serotype i originating from source j, denoted as λij, is estimated by:mentary livestock data were taken and that the diversity of humanO-serotype frequency distributions among the different countries mainly reflects differences in the predominant livestock sources. As suggested by Smid et al. (2013), the proportional similarity index (PSI), a straightforward measure of the area of overlap between two fre- quency distributions of microbial subtypes, was used to measure the similarity between the O-serotype frequency distributions of humanwhere pij is the prevalence of O-serotype i in source j and ei is the observed frequency of human isolates of O-serotype i. pij is given by πj × rij, where πj is the overall prevalence of STEC in source j and rij is the relative frequency of O-serotype i in source j. Details of each param- eter estimation are reported in Table 2.

The mDM was implementedFrom both the mDM and mHM, we extracted the relative posterior probabilities (Pr) for each O-serotype to originate from each of the four animal sources.Of the 1183 human isolates, 151 (12.8%) were discarded because their O-serotypes had no one-to-one matching with any of the isolates from the considered four livestock sources, and therefore, they could not be attributed using the mDM or mHM. These human isolates were then assigned a priori to an “unknown” source. Therefore, the final data set comprised 1032 human isolates to be attributed. These isolates comprised 59 different O-serotypes plus a non-typeable (NT) group including those isolates with an undetermined O-serotype. All the 59 O-serotypes included in the analysis were therefore found in humans and in at least one of the considered four livestock sources. Another 29 O-serotypes were found only in the sources, but not in humans (the so-called non-pathogenic types). As suggested by Mullner, Jones, et al. (2009), these types were kept in the models to preserve the whole within-source relative frequency of O-serotypes.As the livestock isolates were not all from the same country and years as the human isolates and were supplemented with isolates from other European countries within a 4-year time frame, we assessedpotential temporal and geographical biases by quantifying the degree of similarity (as revealed by PSI) of the O-serotype frequency distribu- tions of the Dutch vs. non-Dutch isolates, as well as those from the contemporaneous (i.e., same year) vs. non-contemporaneous isolates as performed previously (Mughini-Gras, Smid, et al., 2014; Smid et al., 2013). We then calculated PSI values and 95% bootstrap confidence intervals (1000 iterations) between Dutch and non-Dutch source iso- lates overall and per livestock group, as well as PSI values between Dutch human isolates and isolates from each of the sources of either Dutch or non-Dutch origin.

Proportional similarity index values were also calculated between humans and each livestock source over the 4 years, from 2011 to 2014 (i.e., 16 comparisons per human-source- year combination). A linear trend was then fitted over the obtained PSI values to test whether increasing time interval between the human and livestock data sets corresponded to decreasing similarity in their O-serotype distributions.We investigated risk factors for human STEC infections caused by O-serotypes attributable to cattle, small ruminants, pigs and poultry. We did so by performing a so-called source-assigned case–control study using the same methodology as previously applied to Salmonella (Mughini-Gras, Enserink, et al., 2014). In brief, eight putative risk fac- tors investigated in a previous (2008–2012) Dutch case–control study for human STEC infection in general (Friesema, Schotsborg, et al., 2015) were retested for association with the occurrence of human STEC infections of probable cattle, small ruminant, pig or poultry ori- gin using Pr-weighted logistic regression models. A separate model for each source was built, with cases being weighted in each model based on the Prs from each specific source, and two analyses were per- formed using the Prs from the mDM and mHM, respectively. An equal probability weight of one was assigned to the controls, while the cases were weighted according to the Pr values per O-serotype to originate from each of the four livestock sources under study. Consequently, factors associated with STEC infection in the Pr-weighted models could be interpreted as source-specific risk factors (Mughini-Gras, Enserink, et al., 2014).The source-assigned case–control study included 342 STEC cases and 2260 controls, both enrolled in the Netherlands during 2010– 2014. Part of these cases and controls was already included in the previous Dutch case–control study for human STEC infection as a whole conducted in 2008–2012 (Friesema, Schotsborg, et al., 2015). The cases also represented a subset of the 1183 O-serotyped STEC cases included in the source attribution analysis (see “Data collection” section) and for which epidemiological information about exposure to potential risk factors was available through the voluntary completion of a standard questionnaire.

Questions contained in the questionnaire referred to the week prior to the onset of symptoms. A case was then defined as an O-serotyped STEC infection during the period 2010– 2014 in a person who was interviewed within 1 month after the onset of symptoms and who had not travelled abroad in the week before the onset of illness. Patients were excluded if they were known tobe secondary cases of a household cluster or if they were part of an outbreak (Friesema, Schotsborg, et al., 2015). The controls were en- rolled through a periodic survey of the Dutch population as described in Friesema, van Gageldonk-Lafeber, and van Pelt (2015). Briefly, in 2008, the RIVM started a periodic survey in the Dutch general popu- lation based on the thrice-yearly administration of an epidemiological questionnaire to a random, dynamic sample of the open population to obtain data from people to be used as control group for identifying risk factors for several notifiable gastrointestinal infections, and this has already been carried out for STEC (Friesema, Schotsborg, et al., 2015) and Listeria (Friesema, Kuiling, et al., 2015). Controls for the current study were selected if they participated to the survey between 2010 and 2014 and if they did not travel abroad in the week before com- pleting the questionnaire.The questionnaire for the controls asked similar questions as that for the cases. Thus, potential risk factors were selected from the vari- ables available from both cases and controls based on their biological plausibility and previous investigations (Friesema, Schotsborg, et al., 2015). Besides general demographics like age (≤5, 6–15, 16–45, 46–65, ≥66 years) and gender, urbanization degree of the residence postcode (urban: >2,500 addresses/km2; intermediate: 500–2,500 addresses/km2; rural <500 addresses/km2), socio-economic status (SES, expressed as a normalized score ranging from −4 to +4 based on income, employment and educational level per postcode area, with a higher score indicating lower SES, as provided by Netherlands Statistics), season (winter: December–February; spring: March–May; summer: June–August; autumn: September–November) and year (2010 to 2014), we assessed the association between being an STEC or an STEC-source-assigned case and the following eight variables:(i) consumption of raw and/or undercooked meat; (ii) consumptionof beef; (iii) consumption of minced meat; (iv) consumption of cured meat/cold cuts; (v) consumption of raw meat spreads (including also typical Dutch products like “filet americain,” a variation of the more popular steak tartare made with raw beef and various seasonings, and “ossenworst” a spreadable raw beef sausage); (vi) consumption of cheese made with raw milk; (vii) consumption of raw vegetables (including also sprouts); and (vii) contact with farm animals.

Missing values in these variables were handled using multiple imputation as described in Doorduyn, Van Den Brandhof, Van Duynhoven, Wannet, and Van Pelt (2006).Logistic regression models were fitted using a backward stepwise variable selection approach. Variables were dropped one by one if they showed a p-value ≥.05 and their exclusion from the models did not in- fluence the association of the other covariates; a change of >10% in the regression coefficients of the covariates was seen as a sign of confound- ing, so the variable in question was retained in the model regardless of its significance. Collinearities between independent variables were explored prior to regression analysis by examining their covariance ma- trix and selection between collinear variables was based on model fit as revealed by the Akaike information criterion (AIC). Associations were then expressed as adjusted odds ratios (ORs) and corresponding 95% confidence intervals (95% CIs). All final regression models showed an overall statistical significance (likelihood ratio chi-square test, p < .05),goodness of fit (Hosmer–Lemeshow test, p > .05) and negligible impact of influential observations as assessed by Pearson residual, deviance re- sidual and Pregibon leverage statistics. Age, gender, urbanization degree, SES, season and year were always included in the regression models to control for potential confounding effects. Plausible two-way interac- tions between the aforementioned eight putative risk factors and the variables age, gender, urbanization degree, SES and season were also tested. For comparison purposes, in addition to the Pr-weighted models for each source, an overall unweighted model was also developed to identify risk factors for human STEC infection as a whole. Regression models were also fitted with a random intercept at the postal code and/ or municipality levels to account for potential clustering of cases, but this did not improve the fit of the models; thus, only the results from the ordinary logistic regression models were reported. Statistical analyses were performed using STATA 14 (StataCorp, USA).

3| RESULTS
The attribution estimates from the mDM and from the mHM were very similar (Figure 1). Both models identified cattle as the primary source of human STEC infections, accounting for 48.6% (95% CI: 44.4%–52.6%) of the 1,183 human isolates in the mDM and for 53.1% (95% CI: 41.7%–63.6%) of these same isolates in the mHM. Small ru- minants were estimated as the second most important source: the mDM and the mHM attributed 23.5% (95% CI: 20.4%–26.8%) and 25.4% (95% CI: 17.2%–35.7%) of the human isolates to small rumi- nants, respectively. Pigs were estimated to account for 12.5% (95% CI: 9.0%–16.5%) of human infections by the mDM and for 5.7% (2.2%–11.1%) of them by the mHM. The least important source was estimated to be poultry by both the mDM (2.7%, 95% CI: 1.4%–5.1%) and the mHM (3.1%, 95% CI: 1.1%–6.9%). The source of the remain- ing 12.8% human isolates was unknown.The specific attributions of the most frequent O-serotypes found among the 1032 human infections attributed to livestock sources by the two models are reported in Table 3. Serotype O157 was strongly associated with cattle: 64.8% and 68.8% of the 321 human O157 iso- lates were attributed to cattle by the mDM and mHM, respectively. Also non-O157 serotypes were mainly associated with cattle, as 51.7% (mDM) and 56.1% (mHM) of human non-O157 isolates were attributed to cattle, but the contribution of small ruminants (30.1% and 32.0% in the mDM and mHM, respectively) was also substantial. Of the five most frequent non-O157 serotypes, O26, O91, O103 and O113 were predominantly associated with cattle, whereas O146 was mainly associated with small ruminants (Table 3).Table 4 shows the PSI values between the O-serotype frequency distributions of the Dutch and non-Dutch isolates from humans and livestock sources. Overall, Dutch human isolates appeared to be more similar to the non-Dutch (PSI 71.2%, 95% CI 63.1%–79.3%) than to the Dutch (PSI 39.8%, 95% CI 18.9%–60.7%) source isolates.

The PSI between Dutch and non-Dutch source isolates was 49.5% (95% CI 28.9%–70.2%), so the human cases were attributed based on a much comparable pool of non-local source isolates, indicating that pooling of data over countries was unlikely to introduce a significant bias. Of note, with the exception of cattle isolates, the PSI values between human isolates and the isolates of each source were similar regardless of whether these isolates were of Dutch or non-Dutch origin (Table 4). In Figure 2, the PSI values for the inter-annual comparisons ofO-serotype distributions between humans and each of the livestock sources are plotted. A linear function was fitted over these PSI values. No source had O-serotype distributions that appeared to be signifi- cantly less similar to the human ones as the time difference between the years of collection increased (cattle: slope −0.008, p = .483; small ruminants: slope 0.011, p = .516; pig: slope 0.009, p = .605; poultry:slope −0.007, p = .610). Proportional similarity index values for inter- annual comparisons of O-serotype distributions within sources and within humans were also computed, but none showed significant trends, thereby justifying pooling of data over the years.Significant risk factors for human STEC infection as a whole were the consumption of beef, raw/undercooked meat, and cured meat/cold cuts (Tables 5 and 6), whereas consuming raw vegetables was associ- ated with decreased STEC risk. Other significant risk factors for STEC infection in general were male gender, low SES, living in rural vs. urban areas, and summer and autumn seasons vs. the winter, as well as sum- mer, autumn and winter seasons vs. the spring. The odds of being a case were significantly greater in 2011, 2013 and 2014 as compared to 2010, and by including year as a continuous variable, there was a significant (p < .001) linear trend towards increased STEC risk over the study years (OR 1.21, 95% CI 1.09–1.34). Moreover, compared to children below 5 years of age, any of the other age groups above 15 years had a decreased risk of STEC infection (Tables 5 and 6).Looking at the Pr-weighted models revealing the specific factors associated with infection with O-serotypes originating from each source, both those based on the Prs from the mDM and those from the mHM gave essentially the same results (Tables 5 and 6). For the cattle-associated STEC infections, significant risk factors were the consumption of raw meat spreads and the consumption of beef. Consuming raw/undercooked meat or consuming minced meat were significant risk factors for infection with O-serotypes attributable to small ruminants.

For STEC infections of probable pig origin, only the consumption of raw/undercooked meat was a significant risk factor. Consuming minced meat, raw/undercooked meat, or cured meat/cold cuts, and living in an intermediate vs. urban area, were associated with increased risk of acquiring a STEC infection of probable poultry ori- gin. Conversely, consuming raw vegetables was a protective factor for infection with O-serotypes attributable to any source (Tables 5 and 6). In the Pr-weighted models, only a few deviations from the overall (unweighted) model were observed as regard to age and season, with small ruminant- and poultry-associated infections being significantly less likely to occur from 45 years of age onwards and no significant increase observed in the risk of poultry-associated STEC infections during summer or autumn as compared to winter, at least when using the Prs from the mDM. Male gender remained a significant risk factor only for pig-associated STEC infections, low SES for cattle- and pig- associated infections, and living in rural areas for cattle- and poultry- associated infections. The increasing inter-annual trend in STEC infection risk could be observed for all sources (Tables 5 and 6). In all models, no significant interactions were found.

4| DISCUSSION
The aim of this study was to quantify the relative contributions of cat- tle, small ruminants, pigs and poultry to the human disease burden ofSTEC infection and to determine risk factors for infection with STEC strains originating from these livestock sources in the Netherlands. Previous studies with this same aim focussed on Campylobacter (Mossong et al., 2016; Mughini-Gras et al., 2012) and Salmonella (Mughini-Gras, Enserink, et al., 2014). We then used quantitative risk modelling to attribute stochastically human STEC infections to sources based on O-serotyping data. Subsequently, a source-assigned case–control study allowed for the identification of source-specificrisk factors, which also provided an indication of the underlying trans- mission pathways.Cattle was estimated to be the most important source of human STEC infections in the Netherlands, accounting for 49%–53% of human cases, followed by small ruminants, pigs and poultry. This is in line with the recognized role of cattle as the main reservoir for STEC strains that are highly virulent to humans, such as those within the serogroup O157 (Caprioli et al., 2005), which was indeed stronglyTABLE 5 Adjusted odds ratios and 95% confidence intervals of the factors associated with human Shiga toxin-producing Escherichia coli infections attributable to each of the four animal reservoirs (according to the Prs from the modified Dutch model) and overall as given by the final multivariable logistic regression modelassociated with cattle in this study (65%–69%), as also supported by numerous investigations conducted in several countries during recent years (Friesema, Schotsborg, et al., 2015; Kassenborg et al., 2004; McPherson et al., 2009; Werber et al., 2007).

Yet, the fraction of human STEC infections attributable to cattle, while believed to be the largest, has never been quantified as carried out in this study, and this also applies to the fractions for the other sources. Shiga toxin-producing Escherichia coli, including O157 and other serogroups associated with human infections, are frequently isolated from sheep (Heuvelink et al., 1998; Urdahl, Beutin, Skjerve, Zimmermann, & Wasteson, 2003), goats (Pritchard, Willshaw, Bailey, Carson, & Cheasty, 2000) andfoods of animal origin from these species, including meat (Chapman, Siddons, Cerdan Malo, & Harkin, 2000) and milk (Rubini et al., 1999). Moreover, small ruminants may also have a relevant role in spreading STEC in the environment (Howie, Mukerjee, Cowden, Leith, & Reid, 2003). Therefore, small ruminants too are considered important res- ervoirs for human STEC infection, and this is supported by our results showing small ruminants to be the second most important source of human STEC infection in the Netherlands, accounting for 24%–26% of all cases. Of note is the relatively large attribution of non-O157 se- rotypes (30%–32%), particularly serotype O146 (71%–77%), to small ruminants, as this serotype is indeed predominantly found in sheepTABLE 6 Adjusted odds ratios and 95% confidence intervals of the factors associated with human Shiga toxin-producing Escherichia coli infections attributable to each of the four animal reservoirs (according to the Prs from the modified Hald model) and overall as given by the final multivariable logistic regression model(Beutin et al., 1997; Blanco et al., 2003; Frohlicher, Krause, Zweifel, Beutin, & Stephan, 2008).Shiga toxin-producing Escherichia coli has been isolated only sporadically from animals other than ruminants, and in most cases, it is unclear whether these animals can be deemed to be actual hosts for STEC or merely as spill-over hosts momentarily colo- nized by STEC after exposure to ruminant manure (Caprioli et al., 2005).

This seems to be particularly relevant for poultry, as its estimated contribution to human STEC infections was minimal in this study (around 3%), and no STEC is usually to be found in live chickens (Beutin, Geier, Steinruck, Zimmermann, & Scheutz, 1993; Heuvelink, Zwartkruis-Nahuis, van den Biggelaar, van Leeuwen, &de Boer, 1999), even though O157 strains have been isolated from retail poultry products (Leclercq & Mahillon, 2003). Interestingly, pigs accounted for a 6%–13% of human STEC infections, indicat- ing that they may be a non-negligible source of STEC associated with human illness. However, as pointed out before (Caprioli et al., 2005), the generally low faecal carriage of strains like O157 (<2%) in pigs in industrialized countries could be the result of accidental exposure of pigs to ruminant manure. Although the prevalence of STEC in pigs can be as high as 68%, the relatively low attributions to pigs are likely due to most STEC strains from pigs carrying the stx2e gene, which might limit disease potential for humans as other virulence genes might more be important for human disease thanstx2e (Fratamico, Bagi, Bush, & Solow, 2004; Tseng, Fratamico, Bagi, Manzinger, & Funk, 2015).The source-specific factors identified in our study were epidemi- ologically plausible according to the source in question. For instance, consumption of beef and raw meat spreads (which in the Netherlands are mostly produced from beef) was specific risk factors for infection with STEC strains attributed to cattle. Raw/undercooked meat (of any origin) was a risk factor for infections attributed to small ruminants, pigs and poultry, whose meats are usually consumed well cooked, so the consumption of, for instance, undercooked mutton, pork or chicken may be assumed to be a generally accidental event posing a risk for infection with STEC strains from these animals. Other as- sociations, such as the consumption of minced meat as a risk factor for small ruminant- and poultry-attributed STEC infections, as well as the consumption of cured meat/cold cuts for poultry-attributed infec- tions, are quite unclear.

Yet, they certainly deserve further investiga- tion, as in the Netherlands there are a number of specific products that can reflect such risks, for instance medium–rare hamburgers of lamb meat (often called “Greek hamburgers”) or poultry-based deli meats like smoked chicken or roasted turkey ham.In general, the risk factor analysis showed that risk factors forSTEC infection in the Netherlands vary according to the attributable reservoir, entailing that STEC may infect humans through different transmission routes depending on their original reservoirs. This is of importance since in the past years, remarkable changes in the epide- miology of human STEC infections have occurred. For instance, in ad- dition to relatively well-known foods of bovine origin like hamburgers, several outbreaks have been associated with low pH products, such as fermented salami, mayonnaise and yoghurt (Caprioli et al., 2005). In addition, outbreaks associated with environmental exposures are also reported (Howie et al., 2003), with spatial analyses indicating an increased risk for STEC associated with the magnitude of cattle farm- ing (Friesema et al., 2011), possibly due to dispersion of manure in the environment that can cause the contamination of different items, in- cluding vegetables. Indeed, vegetables fertilized with manure or con- taminated during harvesting or processing are a cause for concern and have been involved in several outbreaks (Franz & van Bruggen, 2008). An example is the outbreak of STEC O157 in the Netherlands and Iceland in 2007 linked to contaminated lettuce (Friesema et al., 2008) or again the outbreak of STEC O104:H4 in Germany in 2011 linked to fenugreek sprouts (Buchholz et al., 2011). We found consumption of raw vegetables to be a protective factor. Several studies on risk factors for STEC infection have found consuming raw vegetables to have a protective effect (Friesema, Schotsborg, et al., 2015; Locking et al., 2011; McPherson et al., 2009; Werber et al., 2007), and in the present study, raw vegetables were protective for infection with STEC strains of any origin. This finding is not limited to STEC, but the same has been found for other bacteria, including Salmonella (Mughini- Gras, Enserink, et al., 2014) and Campylobacter (Mughini-Gras et al., 2012).

A diet rich in vegetables may have genuinely beneficial effects on general health, meaning that the benefits of eating raw vegetables would somehow exceed the risk of acquiring STEC or other bacte- rial infections. However, it is also true that the controls returning thequestionnaire may just have been particularly motivated people with a generally healthier lifestyle than the cases (selection bias), providing an alternative explanation as to why raw vegetables had such a protective effect in this study. This study has several limitations. Although STEC cases over a pe- riod of 5 years were included in the source attribution analysis, only a subsample of them could be included in the case–control study as well. This is due to the lower number of cases that also agreed to be interviewed, which entails a tendency towards some overrepre- sentation of more severe illness. Moreover, as the cases were gen- erally identified by passive surveillance (so they were already likely to represent the more serious cases occurring in the population) and the controls were not tested for STEC carriage, the risk factors stud- ied here especially represent risk factors for serious STEC infection. Due to the study design and questionnaire limitations, analysis of more detailed risk factors was not possible. Moreover, the controls were known to slightly deviate from the general population of the Netherlands, with a small underrepresentation of men, young peo- ple and people living in large cities (Friesema, Gageldonk-Lafeber, et al., 2015), thereby supporting the choice of always adjusting the multivariable analyses for gender, age and urbanization degree. With regard to potential misclassification of the exposure, we did not in- vestigate private behaviours that can generate shame or stigma (e.g., sexual orientation, drug abuse, personal hygiene), and recall period was the same for both cases and controls, so we can reasonably as- sume that if misclassification occurred, it was mainly non-differential, so biasing the ORs towards the null hypothesis. While cases were ac- quired from a (passive) laboratory surveillance system, controls were acquired at random from the general population.

Many case–con- trol studies use similar sources of controls, often relying on random or sequential digit dialling from telephone directories in addition to general population registries. However, it is well recognized that ac- quiring cases and controls from different sources may lead to some selection bias, as even when controls are selected from the same catchment area of the surveillance system, there is no guarantee that, had the controls become ill, they would have sought medical at- tention and been reported as were the cases (Fullerton et al., 2012). Our study was based on countrywide laboratory surveillance data for STEC, and controls came from the same population and geographical area (i.e., catchment area of the surveillance system) as the cases, so we can assume that they had the same chance to be recognized as cases if they had developed the disease. Yet, as Fullerton et al. (2012) pointed out, there are factors affecting the identification process of the cases that have the potential to make the cases ascertained by surveillance systems substantially different from population-based controls.With a limited number of isolates and diversity of serotypestherein, not only our sample would have been less likely to be rep- resentative, but the source attribution models would have per- formed poorly (i.e., large uncertainty in the estimates, large fraction of non-attributable cases, failure of the models to converge). We have therefore used both local and non-local livestock data within a 4-year timeframe for the source attribution analysis. Inherent to thisapproach was the assumption that the non-local livestock data were also representative of the STEC strains circulating in the Netherlands and that there were no major differences in the occurrence of these strains over the study period. Although this may be considered a major limitation, the application of the method of Smid et al. (2013) provided a way to minimize potential biases. We also performed addi- tional analyses to assess the degree of (dis)similarity of livestock data over time and between local and non-local sources.

Altogether, these analyses confirmed that the non-local livestock data were more sim- ilar than they were different to the local sources, thereby acting as a good surrogate. Moreover, no significant temporal trends were found in the similarity of livestock data, indicating that temporal variation in their O-serotype distribution was unlikely to have introduced a sig- nificant bias in the analysis. We also found that human isolates were more similar to the non-Dutch cattle isolates than the Dutch cattle isolates. In general, foods and animals can be moved freely through- out the EU, provided that they meet EU’s standards for animal health and welfare, as well as safety, certification and proper use of foods thereof. According to Netherlands Statistics (www.cbs.nl), in the years 2010–2014, about 49%–52% of beef available for consumption in the Netherlands was imported (88%–91% of which from other EU coun- tries), providing a possible explanation for the high similarity between human and non-Dutch cattle isolates we found. Other limitations of this study concern the potential clustering of source isolates by herd or retailer, although including only different serotypes from a cluster did prevent sample inflation. In addition, there were no data available concerning other potentially relevant sources, including companion animals and wildlife, although the non-attributable fraction of 12.8% of cases suggests these other sources may be important to investigate in future studies.

In conclusion, this is the first combined source attribution and case–control analysis for STEC allowing for both the quantification of the fractions of human STEC infections attributable to four putative livestock sources and their associated risk factors. We showed that do- mestic ruminants account altogether for approximately three-quarters of reported human STEC infections in the Netherlands, whereas pigs and poultry play a minor role as sources of STEC. We also showed that risk factors for STEC infection may vary according to the attributable source, providing an approach for generating hypotheses on the trans- mission pathways for STEC, as its epidemiology has changed over the past years and a growing number of unusual vehicles are being associ- ated with human Brigimadlin infection.