Outbreak-Related Disease Burden Associated with Consumption of Unpasteurized Cow’s Milk and Cheese, United States, 2009–2014

The growing popularity of unpasteurized milk in the United States raises public health concerns. We estimated outbreak-related illnesses and hospitalizations caused by the consumption of cow’s milk and cheese contaminated with Shiga toxin–producing Escherichia coli, Salmonella spp., Listeria monocytogenes, and Campylobacter spp. using a model relying on publicly available outbreak data. In the United States, outbreaks associated with dairy consumption cause, on average, 760 illnesses/year and 22 hospitalizations/year, mostly from Salmonella spp. and Campylobacter spp. Unpasteurized milk, consumed by only 3.2% of the population, and cheese, consumed by only 1.6% of the population, caused 96% of illnesses caused by contaminated dairy products. Unpasteurized dairy products thus cause 840 (95% CrI 611–1,158) times more illnesses and 45 (95% CrI 34–59) times more hospitalizations than pasteurized products. As consumption of unpasteurized dairy products grows, illnesses will increase steadily; a doubling in the consumption of unpasteurized milk or cheese could increase outbreak-related illnesses by 96%.

For α, the number of hospitalizations were directly obtained from the National Outbreak Reporting System (NORS) (2), while the number of illnesses was obtained after correction for pathogen-specific underreporting, under testing (i.e., the fact that samples are not collected from all suspected cases and not all samples are tested), and underdiagnosis (i.e., false negative). Sets of independent adjustment factors were sampled and combined as shown below to estimate illnesses: where αobs is the number of laboratory-confirmed cases as reported in NORS (2), γ is the underreporting factor, μ is the underdiagnosis factor, and ρ is the under-testing factor for a given pathogen. Another model structure was tested, where the adjusting factors were modeled using a hypergeometric process. However, a sensitivity analysis showed this did not affect the results, and thus the more parsimonious model structure shown in equation 2 was chosen. Means and credibility intervals for the adjustment factors and the data used for their calculation are shown in online Technical Appendix Table 2.

Estimation of the Underreporting Factor γ
We estimated the underreporting factor by comparing the total number of laboratory confirmed cases from dairy-related outbreaks (NODRcases) reported to NORS from 2009 through 2013 in the United States with the estimated number of laboratory-confirmed cases from outbreaks that were attributed to dairy consumption from FoodNet (NLCcases) for the same period: In doing so, we assumed that FoodNet surveillance population and reporting practices were representative of the overall United States. NODRcases was directly obtained from NORS.
NLCcases was derived from estimated numbers of laboratory-confirmed cases for the US population (NUScases), and adjusted to outbreak and dairy-related cases: NUScases was estimated by extrapolating the yearly incidence rates of laboratory-confirmed cases in the FoodNet population (RUScases) to the US population NresUS and summing them for where NresUS was calculated from the FoodNet study population (NFoodNet) and the proportion of the US population this study population represents (PFoodNet): For the 4 pathogens of interest, the incidence rates of laboratory-confirmed cases in the FoodNet population (RUScases) were given by: where NFoodNetcases were the total number of laboratory confirmed cases reported by FoodNet. This estimated number of laboratory-confirmed cases in the US derived from FoodNet data (NUScases) was then adjusted as described in equation 4, so as to only include the outbreakrelated cases attributable to dairy.
Assuming that proportions of laboratory-confirmed cases that are outbreak-related (PORcases) are pathogen-specific and do not change over time, PORcases were approximated using data from Scallan et al. (3): where Ncases was the total number of laboratory-confirmed cases, and Nob was the number of these cases that were outbreak related, as reported to FoodNet for 2004-2008.
The pathogen-specific estimates of the proportion of outbreak-related illnesses that are attributable to dairy (PDRcases) were derived from the study by Painter et al. (4): This assumes that the proportion of outbreak-related illnesses caused by dairy products were caused by milk or cheese (i.e., simple foods) during our study period.

Estimation of the Underdiagnosis Factor μ
The underdiagnosis factor used in equation 2, μ, accounts for the rate of false negatives using the test sensitivity described in Scallan et al. (3): where ~Pert (minimum, mode, maximum) (equation 11).

Estimation of the Under-testing Factor ρ
The under-testing factor in equation 2, ρ, accounts for the fact that in an outbreak investigation, samples are not collected from all suspected cases, and diagnostic tests are not conducted on all samples taken: where βobs is the number of estimated primary cases, and αobs is the number of laboratoryconfirmed cases (2,5). Because of the clustering of the cases by outbreak, the above estimation could potentially be biased.
In equation 1, the number of servings of a given dairy product and pasteurization status, Nserving, was calculated as: where Nresid is the total resident population in the United States (online Technical Appendix Table 3), Npers serv is the number of servings per person, and pcons is the proportion of the population of dairy consumers who consume milk or cheese of a given pasteurization status.
For example, pcons,milk,unpast, the proportion of the population of dairy consumers that consumes unpasteurized milk, is calculated as:  Table 5): Pc,state is given by with NPcons being the number of respondents that indicated that they consumed the product in the last 7 days and Nsurvey the FoodNet survey population in the given state.

Estimation of the Excess Risks Associated with the Consumption of Unpasteurized Milk and Cheese
The additional risk of outbreak-related illness and hospitalization for consumers of unpasteurized dairy products, compared with consumers of pasteurized ones, was estimated using 2 measures of excess risk (23). The risk difference measures the actual difference in the incidence rates of illness and hospitalization between consumers of unpasteurized dairy products (λu) and consumers of pasteurized ones (λp): The incidence rate ratio provides a relative comparison of the risks for illness and hospitalization between the 2 exposure groups: .

Impact of Hypothetical Changes in Consumption of Unpasteurized Milk or Cheese
A scenario analysis was performed for the year 2015 to assess the public health impact of hypothetical changes in consumption of unpasteurized dairy products. Six scenarios were considered: 10%, 20%, 50%, 100%, 200%, and 500% increases in the proportion of the US population consuming unpasteurized milk or cheese.
The number of annual outbreak-related illnesses associated with milk or cheese consumption, αpred, was estimated as As shown in equation 13, the number of servings of milk or cheese for 2015 requires the estimation of the total US resident population and the per capita consumption for that year. Using a simple linear regression, we predicted these 3 values using historical data on the US resident population from 1996 through 2014 (online Technical Appendix Table 3) and milk and cheese consumption per capita from 2006 through 2014 (online Technical Appendix Table 4). The variability in the 2015 predictions for these 3 values when considering parameter uncertainty was modeled using a standard prediction interval calculation: where y is the prediction for the year 2015, b0 is the regression intercept, βt is the slope for the year (i.e., the yearly growth or decline in y), xt is the predicted year (i.e. And the fraction of milk servings that are unpasteurized milk servings is the sum of PUnPcons,milk and ΔPUnPcons. The number of hospitalizations per year was modeled as a fraction of illnesses (αpred) where the uncertainty in the probability of hospitalization in case of illness is modeled using the conjugate prior: where αobshosp is the number of reported outbreak-related hospitalizations due to illnesses from a given pathogen.