Verotoxigenic E. coli (VTEC) is a highly infectious causative agent for severe gastrointestinal infection with haemorrhagic diarrhoea, which can lead to haemolytic uraemic syndrome (HUS) or thrombocytopaenic purpura (TTP) [1]. Most of the diagnosed VTEC infections in developed countries, including Finland, are of the serogroup O157, but an increasing number of cases are found with non-O157 serogroups O26, O111, O91, O103 and O145 [2, 3]. The majority of VTEC infections occur among young children, predominantly during the summer months [2–4]. VTEC infection is rare in Finland, with only 10-20 cases reported annually to the National Infectious Disease Register (NIDR) [5].

Several risk factors for VTEC infections have been identified from outbreak data, modelling studies, and descriptive and analytical epidemiological studies. Cattle are considered the main reservoir for VTEC [6–8]. Living, visiting or working in close proximity to a farm with cattle or having a household contact working in such a farm have been found to be important risk factors [9–11]. Cattle density or VTEC-positive cattle density have been established as important explanatory variables for VTEC infections [4, 12–16] or cases of HUS [16]. Also farm density per area of land [14] and animal manure application to soil are important [13]. Undercooked beef meat or meat products, water, vegetables and unpasteurized milk, ice cream, juice and cider can be sources of VTEC infections [17], and consumption of burgers and cold meat cuts are established risk factors [9, 18]. Pathogenic strains of VTEC have been detected in minced beef meat, beef meat cubes and cuts, raw cattle, sheep and goat milk and sprouts [11]. Lake water or contaminated municipal water systems have caused outbreaks of VTEC O157 [17, 19, 20] and consumption of untreated surface water is a risk factor for non-O157 [18].

Poisson distribution based regression models are standard models for count data, including for infectious disease surveillance data [21]. However, if the data has an excess of zeros, the assumptions of the Poisson distributed models are no longer valid. Models based on negative binomial distribution may slightly correct for this problem, but when there is a severe inflation of zeros, more specific models are needed. Zero-inflated and hurdle models have been developed to overcome this problem [21, 22]. With the use of a complementary log-log link function, the hurdle model reduces to a standard Poisson regression if the data is Poisson distributed [22]. These types of models have been applied recently for modelling sleeping sickness [23], cholera prevalence [24] and bacterial counts [25].

The aim of this study was to identify explanatory variables for VTEC infections reported to the NIDR in Finland between 1997 and 2006. VTEC infection is a rare disease in Finland and we used Poisson distribution based hurdle models with variable selection based on in-depth questionnaires and known risk factors. We chose the hurdle model based on its ability to handle an over-dispersion of zeros and applicability to scarce data. The incidence rate of VTEC infection at municipality level was used as an outcome variable and explanatory variables included various agricultural, socioeconomic and environmental factors. These explanatory variables also included several factors that have not been studied with VTEC infections previously. The null hypothesis that the populations in different municipalities in Finland are at the same risk of contracting VTEC infection was tested using a Poisson distribution based hurdle model.

VTEC infection has a low average annual incidence rate of 4.8 per million inhabitants in Finland. The cases described in the present study were predominantly children with typical symptoms of VTEC infection: Most cases had haemorrhagic diarrhoea and some cases of HUS were identified. Half of the cases were infected with O157 serogroup strains. Detailed microbiological analysis including genotyping was essential to distinguish between outbreak and non-outbreak strains. Approximately a third of cases had a household contact who was VTEC infection positive, indicating direct transmission of the infection. Many of the cases reported visits to farms or consuming undercooked meat or unpasteurized milk; however, it was difficult to assess the impact of the risk factors obtained from the in-depth questionnaires as the information was scarce. Bias due to under-reporting of the cases is considered to be low for VTEC; however, diagnostic differences may occur between health care districts.

Hurdle modelling has not yet been widely used on infectious disease surveillance data [23, 25, 31]. One reason for using the hurdle model in the present study was the fact that the variance was much higher than the mean in the observed number of cases per municipality. The data was also characterised by an excess of zeros. As the zero and the truncated Poisson parts of the model gave different sets of significant variables, this further suggested the need for this type of model. If the coefficients of the binary part and the truncated Poisson part are close to each other, the hurdle model may be replaced by a standard Poisson regression model [22]. From the technical point of view, the prior distributions used were as weakly informative as possible, which guaranteed the convergence of the simulations. The results from the simulations were robust, especially to changes in the variances of the spatial factors. We also recommend using a full-likelihood approach in the multivariable analysis to explain the mechanism for the missing values.

We identified the proportion of beef cattle to human population and the proportion of population with higher education as being associated with increased occurrence and incidence of VTEC infections. The proportion of the population with higher education is likely to represent an indirect indicator of consumption habits, possibly eating undercooked beef steak. In addition, the number of farms under cultivation per area, fresh water per area and the proportion of low income households with children were associated with increased incidence of the infection. These are likely to represent routes for the spread of the infection; the low income households may be an indirect indicator of consumption habits. The model without the spatial variable essentially gave the same results. Our results concur with a US study on ecological factors for gastrointestinal infections, where the proportion of the population living on a farm (positive association), low education (negative association) and living in the south (negative association) were the most important socioeconomic factors for acquiring a VTEC infection [32].

The finding of beef cattle as a strong explanatory variable was not unexpected as cattle have been found to be an important risk factor for VTEC infections in epidemiological and modelling studies [9–11, 13, 14]. The VTEC prevalence is low in cattle in the Nordic countries, including Finland [33, 34]. The overall cattle density [12] or dairy cattle or calves density [16] have also been found be significant, but these differences may reflect correlations between the cattle variables, variation in agricultural practices, and the ways in which data are collected. The significance of beef cattle over the other cattle variables may be due to a susceptibility of young animals to VTEC, raising the bull cattle in finishing units with multiple sources of animals, and the faecal contamination of the bulls and the environment during the finishing period [35]. The beef cattle density was an important factor both for the occurrence of the disease--confirming the role of cattle as a reservoir of the infection--and the incidence of the infection, indicating that contact with cattle may be an important route for spreading the infection. A relationship between human VTEC infections and VTEC positive cattle density has been found [14]. Also indirect cattle variables like the proportion of land where manure is applied has been found to be significant [13]; in the present study the proportion of cultivated land was also a significant risk factor for the incidence of the VTEC infections. Spreading animal manure on fields is common practice in Finland, with most of the manure coming from cattle [36].

The result for the proportion of fresh water surface per area is intriguing. Previous studies suggest drinking and swimming water as a source of VTEC infections and outbreaks [6, 18–20, 37]. However, two influential values had a strong effect for this variable in the present study. It would be interesting to examine the risk factors separately for water borne cases; this information was not widely available. There is recent evidence that contaminated water is an important route of infection in cattle [38] Campylobacter infections have been found to have a positive association with water pipe-length per population and a negative association with proportion of population receiving household water from a public supply [39]. It is noteworthy that the proportion of salted water was not a significant variable in the multivariable analysis. Finland has thousands of small lakes, but it also has a long sea coast along the western and southern borders. In a Swedish study, it was found that most of the VTEC cases lived along the coastline, near lakes or along rivers; however, the type of water was not specified [14]. VTEC prevalence studies have indicated that cases often have a private well or have experienced a water failure recently [12, 40].

The finding that the proportion of the population with a higher education and the proportion of low income households with children being associated with increased incidence of the infection is somewhat contradictory. However, we feel that these two variables may explain different phenomena. There is conflicting evidence about educational level as a risk factor for VTEC infections [32, 41]. However, it is known that people with higher education eat raw or undercooked ground beef more frequently [42], a known risk factor for VTEC infection [9]. The proportion of households where the main supporter of the household works in agriculture did not remain in the final model; however, there was a correlation between this and other agricultural variables. There may also be immunity: It has been found that in outbreaks, long-term exposure can lead to immunity [20]. In other studies, having a household member with an agricultural occupation was found to be a risk factor [9]. The most likely risk factor found for sporadic VTEC infection is having contact with a farm [11].

The study was limited by the low number of VTEC cases; therefore spatiotemporal analysis was not used. This was compensated for by using specific models designed for scarce data and performing a purely spatial analysis. The seasonal trends could not be evaluated properly due to the low number of cases. Another limitation of the study was the unavailability of data regarding consumption of beef or other food items per municipality. It has been found in Scotland that of all VTEC infections, 16% of cases had a significant food-related exposure, such as eating high-risk foods or having a contact working with meat [10]. Considering the VTEC infections in our study and the strong evidence of domestically acquired infections, transmission via imported food stuff would be plausible, based on the similar pheno- and genotypic characteristics of the non-sorbitol fermenting, sorbitol fermenting and non-O157 strains isolated in Finland and other European countries [7, 43, 44].

In conclusion, cattle are the main reservoir for VTEC infections as well as being also an important factor for the incidence of the disease. In addition, higher education per population was found to be associated both with the occurrence and the incidence of the disease, implying that food exposures may play a role. The role of fresh water and other environmentally related variables warrants further studies. Socioeconomic factors like low income and educational level expedite the incidence of VTEC infection. The recommendations for control measures for the prevention of VTEC infections given by the WHO include good hygiene in animal and slaughter process and food retail, consumer education on proper handling of foods, and up-to-date international and national legislation and regulation [45].

We enrolled 131 cases of VTEC with 74 VTEC O157 and 58 non-O157 strains isolated (one person had dual infections). With Poisson hurdle models we were able to distinguish between risk factors for the occurrence of the disease (zero part of the model) and incidence of the disease (zero truncated Poisson part of the model) for data characterised by an excess of zeros in the case counts. The number of bulls per human population and the proportion of population with a higher education were identified as being associated with increased occurrence and incidence of human VTEC O157 and non-O157 infections. The proportion of fresh water per area, the proportion of cultivated land per area and the proportion of low income households with children were associated with increased incidence of VTEC infections. We recommend using a Bayesian framework with imputation for the missing values of the explanatory variables in conjunction with the missingness equations.