We used data from van Loo et al. (8), which consisted of records of all index case-patients with nontypeable MRSA carriage (now referred to as LA-MRSA) from the Netherlands from its emergence in 2003 through September 2005, before the country adopted active surveillance of high risk populations. Information on case-patients and controls was obtained from a national MRSA surveillance program through the country’s Institute for Public Health and the Environment (www.rivm.nl/mrsa). Case-patients were LA-MRSA index patients, that is, the first in a cluster of persons who tested positive for LA-MRSA from a given reference laboratory. Each LA-MRSA case-patient was matched with 2 controls from the same laboratory. The controls were also index patients, but had tested positive for a typeable strain of MRSA (T-MRSA) instead of to LA-MRSA. Further details on subject selection can be found in the original article (8).

The same variables used in the 2007 study were used in this current study: contact with pigs, contact with cows, the probable source of MRSA, and whether one lived in a rural area. Probable source of MRSA was placed by the original authors into the following categories: healthcare setting, foreign source (such as travel to another country), other source, or an unknown source (8). Our study added municipality level variables of livestock and human population densities and location of residence of study participants. (Municipalities are administrative boundaries in the Netherlands that comprise provinces. In 2005, there were 498 municipalities.) To accomplish this, our inclusion criteria were residence in the Netherlands, availability of information on individual contact with livestock, and geographic information sufficient to support mapping each person to the 6 digit postal code of his or her residence.

We then assigned spatial coordinates to LA-MRSA case-patients and T-MRSA controls on the basis of their 6-digit postal code using the geographic information system software, ArcGIS version 9.3 (21). When this method was not sufficient, we used Google Earth (22).

We downloaded municipality level statistics of population; land area; and numbers of swine, veal calves, and cows in 2005 from CBS StatLine (23). Livestock densities and population density were calculated as the number of animals (pigs, cows, and veal calves) per hectare of land in a municipality. In ArcGIS, we determined in which municipality participants lived, and assigned to them municipality-level characteristics of animal and population densities. We determined counts of case-patients and controls for each of the municipalities using ArcGIS. This study was reviewed and approved by the Institutional Review Board at Johns Hopkins Bloomberg School of Public Health.

Summary statistics for all relevant variables were reported by using STATA version 10 (24). We explored the spatial variation in risk for LA-MRSA—or the concept that the risk or odds of acquiring LA-MRSA varies geographically—using spatial methods available in the R statistical package (25). We estimated spatial intensity of case-patients and controls for locations across the study area, defined respectively as the expected numbers of case-patients and controls per km². Spatially varying intensity provides an estimate of regions of high and low densities of case-patients and controls. Spatial intensity is often measured as weighted counts as described (26). We used the quartic kernel as our weighting scheme in this study. Using the intensity estimates, we calculated the spatial odds of LA-MRSA to compare the geographic variation of case-patients and controls across the study area. The spatial odds of LA-MRSA per km² compared with those for T-MRSA per km² were calculated as the ratio of estimated case-to-control intensities (26).

Contact with pigs, contact with cows, and rural (versus urban) residence were modeled as binary variables. Probable MRSA source was a categorical variable. We compared probable acquisition of MRSA from a foreign country, acquisition from another source, or acquisition from an unknown source with the referent group of healthcare-related acquisition.

We determined goodness of fit of the models using Akaike information criteria and the Hosmer-Lemeshow goodness-of-fit test. Likelihood ratio tests were used to compare multivariate nested models. The densities of livestock were right skewed; thus, we log-transformed the variables to create a more linear relationship between animal density and log odds of LA-MRSA. For ease of interpretation, instead of the 1 log increase in livestock densities, we used a doubling of livestock densities, which is calculated by raising 2 to the power of the β of the density coefficient in the logit model (27). Variograms were used to diagnose possible spatial variation in regression residuals, with inference on regression parameters adjusted accordingly.

In a separate but related analysis, we identified specific clusters of LA-MRSA. We used SaTScan version 9.0 (28) to conduct this cluster detection analysis with a Poisson model of counts of case-patients per municipality after adjusting for population size as described (28). SaTScan is a software package that is used to analyze spatial and temporal patterns in data. It uses a moving window method (in this application, over a set of contiguous municipalities) and determines the presence of a cluster on the basis of whether the estimated risk within a window is significantly greater than the estimated risk outside of the window. Statistical significance is based on the null hypothesis of Poisson constant risk (28). We created maps showing the identified clusters after adjustment of population density per municipality in ArcGIS. We made similar maps after further adjusting for pig, cow, and veal calf densities per municipality.

Descriptive statistics of the study population are shown in Table 1. From the total population used in the Van Loo analysis of 111 persons (35 case-patients, 76 controls) (8), 87 persons (27 case-patients, 60 controls) were included in our study after we excluded persons who lived outside of the Netherlands (n = 4), persons for whom spatial information was insufficient (n = 3), and persons for whom information about individual contact with livestock was lacking (n = 17).

Of the 87 subjects with complete case information, most of those who had contact with pigs (10/13, 76.9%) and cows (7/8, r 87.5%) were LA-MRSA case-patients. Three subjects had contact with both pigs and cows, 2 of whom were case-patients. Twelve of 27 persons without any direct contact with livestock were LA-MRSA positive (44.4%).

Specific locations of case-patients and controls are plotted against municipality level population (Figure 1, panel A), cow density (Figure 1, panel B), pig density (Figure 1, panel C) and veal calf density (Figure 1, panel D). Case-patients and controls had significant differences in human and livestock densities per municipality (Table 1).

Relatively high concentrations of controls are seen in general areas of high population density while higher spatial concentrations of case-patients are seen in the more agricultural areas of the country (Figure 2, panels A, B). We estimated spatial odds to give a visual assessment of the spatial variation in risk across the Netherlands (Figure 2, panel C). It is evident that the greatest differences in odds between case-patients and controls are in general areas of high pig density, as was originally reported by van Loo and colleagues (8). The elevated spatial odds in the northern part of the country are a spurious result because of small numbers of case-patients and controls and not something that we put forth as a valid result.

Univariate models results are reported in Table 2. Persons who have contact with pigs have 11.18 times the odds of carrying LA-MRSA (compared with odds of carrying T-MRSA) than persons without pig contact (95% CI 2.76–45.30; p<0.001). A similar relationship is seen when persons with and without contact with cows are compared (odds ratio [OR] 20.65; 95% CI 2.39–178.31; p <0.006). Living in a rural area rather than living in an urban area is associated with 11.2 times the odds for LA-MRSA compared with T-MRSA (95% CI 3.15–39.76; p<0.000). Carriage of MRSA from an unknown or “other” source, as compared to healthcare settings (a priori known to be associated with typeable MRSA), was significantly (p<0.05) associated with the odds of LA-MRSA carriage as compared to T-MRSA.

We found that when pig density per hectare is doubled within a municipality, the odds of acquiring LA-MRSA in a univariate model are increased by 29.5% over the odds of acquiring T-MRSA (p<0.003). Similarly, doubling cow and veal calf densities increases the odds of acquiring LA-MRSA compared with those for acquiring T-MRSA by 75.4% (p<0.001) and 19.8% (p<0.001), respectively.

When the inclusion criteria required by our study were used, 16% of original data were lost. We conducted sensitivity analyses by coding all persons with missing livestock contact information as either all having contact or as all not having contact to produce what might be bounds for high and low extremes. In all cases, livestock densities remained significant independent risk factors at the 0.05% level (Table 3).

Multivariate model results are reported in Table 4. Model 1 is based only on the original individual level variables from van Loo’s study (8): contact with pigs and cows, rural versus urban residence, and information on patient’s probable source of MRSA. In model 1, both contact with pigs and rural residence remain significant predictors as they were in the univariate models, even when adjusting for contact with cows and probable MRSA source. Odds for LA-MRSA compared with those for T-MRSA were 11.6 times higher for a foreign source of MRSA than they were with a healthcare source (95%CI 1.04–129.63; p<0.046). Similarly, the odds were 9.56 when persons with an unknown source were compared with those with a healthcare source (95% CI 1.76–51.93; p<0.009). Acquiring MRSA from another (other) source compared with healthcare acquisition also had increased odds, but this result was not significant (OR 4.3, 95% CI 0.55–33.56; p<0.164).

Models 2–4 build on model 1 (the base model) by adding in the logs of pig, cattle, and veal calf densities per municipality, respectively, with the same individual level variables used in model 1 (Table 4). Model 2 builds on model 1 by adding a term for the log of pig density. The odds ratio comparing LA-MRSA to T-MRSA for a 1 log increase in pig density per hectare after adjusting for the individual risk factors (the variables in model 1) for LA-MRSA was 1.37 (95% CI 1.01–1.87, p<0.041). A doubling of the pig density per municipality increases the odds of LA-MRSA carriage compared with T-MRSA carriage by 24.7% after adjustment for individual level risk factors. Model 3 builds on model 1 by incorporating the log of the cow density per municipality. Adjusting for the individual level predictors, a 1 log increase in cow density yields a 2.28 increase in odds for LA-MRSA compared with T-MRSA (95% CI 1.17–4.45, p<0.016). Here, a doubling of cow density in a municipality increased the odds of LA-MRSA compared with T-MRSA by 76.9%. The odds ratio of carrying LA-MRSA compared with those of carrying T-MRSA in model 4 for a 1 log increase in veal calf density after adjustment for individual variables was 1.37 (95% CI 1.08–1.72, p<0.009). Thus, a doubling of the veal calf density per municipality yields a 24.09% increase in the odds of carrying LA-MRSA compared with carrying T-MRSA.

The Hosmer-Lemeshaw goodness-of-fit tests indicate that all models fit the data sufficiently well. The Akaike information criteria and likelihood ratio values for models 2–4 indicate that adding area-level animal density variables improves the original model (model 1). Variograms of residuals from the 4 models did not reveal any substantial spatial variation.

Cluster detection analysis results indicate that after adjusting for the size of the population in a given municipality, 1 significant cluster of LA-MRSA cases (relative risk 5.2, p<0.014) was found when a maximum of 20% of the population at risk was designated as the maximum spatial cluster size. Figure 3 (panels A–C) shows the cluster detection results mapped on top of veal calf density, all cattle density, and pig density for visual identification of associations.

To test whether accounting for livestock density at the municipality level would eliminate the existence of this hot spot of LA-MRSA, we ran additional analyses in SaTScan with adjustment for the density of each animal population separately (28). These results indicated that adjusting for animal densities eliminated the presence of the cluster, further supporting the hypothesis that livestock densities per municipality are key risk factors for LA-MRSA carriage.