For the purposes of attributing illnesses to foods, food vehicles must be grouped into suitable categories. Although the idea may seem simple, the need for a single food categorization scheme has emerged as a critical issue. At a general level, a list of major food commodities might include 11 categories: poultry, eggs, pork, beef, dairy, fish, mollusks, crustaceans, wild game, row crops (e.g., lettuce and corn), and tree crops (e.g., apples and oranges). Each of these commodities could be divided further, leading to such subcategories as broiler chickens and raw oysters. Additionally, classification could include level of processing (raw, fresh-cut, canned, frozen), origin (domestic, imported), and location of preparation (home, food processor, food service). When illnesses are linked to foods with multiple ingredients, such as soups or casseroles, the choices include whether to omit these cases from analysis, to categorize multiple-ingredient foods by "essential" ingredient, or to attribute illnesses by the proportion of individual ingredients.

A common food categorization scheme is essential if different sources of data are to be combined or compared. Because of lack of agreement in categorization, data from CDC, state health departments, and FDA and USDA and their constituent agencies are often not directly comparable. As a first step in any approach to food attribution, food categories need to be standardized across government agencies, with a scheme acceptable to industry, academia, and consumer groups.

Approaches to food attribution can be grouped into 2 broad categories, loosely designated as epidemiologic and microbiologic. Epidemiologic approaches are based on public health surveillance and include foodborne outbreak data and case-control studies. Microbiologic approaches rely on data on pathogen samples drawn from human, animal, and food sources and include pathogen subtyping, as used in Denmark's Salmonella Accounts (4) and microbial source tracking (MST) methods, as well as risk assessments of specific pathogens in specific foods.

Denmark has an integrated system in which data from public health surveillance and pathogen monitoring of foods and animals are routinely collected, collated, and analyzed by a single coordinating body. Cultures collected from infected persons, animals, and retail food sources are subtyped, allowing for the direct comparison of surveillance and monitoring data and the identification of public health outcomes by food source.

The regular monitoring of food sources is performed on farms, at slaughter, and in retail foods, although the emphasis is on primary production facilities (5). Every flock of egg-laying chickens is regularly tested for salmonellae, as are all flocks of broiler chickens, turkeys, and ducks. Finishing pigs are continually tested, dairy herds are routinely monitored, and poultry, pork, and beef are examined during slaughter processes. Imported meat and poultry products are monitored, as are wild animals, birds, and pets, and retail surveys are performed on raw meat, pork, poultry, shell eggs, fruits, and vegetables.

Salmonella isolates obtained from animal and food sources are subtyped (with serotyping, phage typing, and pulsed-field gel electrophoresis [PFGE]) and compared in a quantitative manner with isolates obtained from human infections (6). A prerequisite of the model is predominance of at least 1 "distinctive" Salmonella subtype in each main animal reservoir; human infections of distinctive subtypes are assumed to have originated from that reservoir. Human infections caused by Salmonella subtypes found in multiple animal reservoirs are attributed proportionally to the occurrence of the distinctive subtypes. Model results have been corroborated by case-control studies, outbreak reports, time-series analysis, and risk assessments (7). In the past 10 years, the Danish model has proven invaluable for identifying pathogen reservoirs in animal populations, tracking trends of human salmonellosis, and guiding interventions (4).

One weakness of the Danish method is that causation cannot be discerned for cases without distinctive Salmonella subtypes; thus, the proportional attribution of such cases across animal reservoirs may not necessarily be accurate. Also, vegetables, fruits, fish, pets, water, and other sources of infection are not directly included in the analysis, under the assumption that the original sources of bacterial infection are animal reservoirs. Furthermore, the model is currently focused on salmonellae and may not be applicable to other pathogens that do not meet certain prerequisites. For example, although extensive subtyping has also been performed on Campylobacter isolates, the homogeneous distribution of subtypes across reservoirs makes attribution difficult. Since the Danish model is focused on the major food-animal reservoirs, it cannot identify responsible foods at the point of consumption or at other points along the farm-to-fork continuum.

The United Kingdom uses an integrated systems approach to food safety that includes both epidemiologic and microbiologic methods, with responsibility for foodborne illness consolidated into a single government office. Annual reports on zoonoses, which combine surveillance data with data on food and animal monitoring (8), are produced. In addition, etiologic analyses of foodborne outbreaks, detailing illnesses by pathogen, food source, and additional risk factors, are performed. UK agencies also perform regular pathogen monitoring and subtyping of animals and retail food (8–10).

The United Kingdom has developed a method for estimating the relative risks associated with specific foods, dividing the number of cases due to a specific food (as derived from their outbreak database) by the estimated total servings of that food consumed in a year. The weaknesses of this system include the assumption that all hospitalizations and deaths are routed through general practitioners and reliance on outbreak data, which may not be representative of sporadic disease. However, the UK outbreak dataset is large, and the food vehicles implicated correlate with findings of local epidemiologic studies. Increasingly, data indicate that interventions guided by the system have been successful in reducing cases and risk for foodborne illness.

Reports of outbreak investigations provide the most comprehensive US data for determining the foods responsible for illnesses. The Foodborne Disease Outbreak Surveillance System contains data on >20,000 US foodborne disease outbreaks reported to CDC since 1973; these reports link specific foods to cases of human illness (11). CDC, the Center for Science in the Public Interest (CSPI), and the FSRC have estimated food attribution using these data (12–16).

Responsibility for investigating foodborne disease outbreaks resides with local and state health departments, which then report these data to CDC. Reported outbreaks represent only a small proportion of those that occur, and the degree of underreporting may vary geographically and temporally. For example, revision of the reporting process and provision of increased resources to CDC and state health departments from the National Food Safety Initiative were associated with a doubling of the number of outbreaks reported annually from 1996 to 1998.

Outbreak data have additional important limitations. Outbreaks that are large, associated with restaurants, have short incubation periods, or cause serious illness are more likely to be investigated and reported. Likewise, illnesses due to pathogens that are difficult to identify or rarely cause outbreaks are underrepresented. For example, the foods most frequently identified as the source of Campylobacter outbreaks differ markedly from those identified as sources in community studies of sporadic cases (17).

An approach used by CDC to estimate illness due to a particular food-pathogen combination is to count the number of outbreak-related illnesses due to a particular pathogen and to determine the proportion of these due to each food grouping. These proportions are then applied to estimates of incidence of that pathogen, as reported by CDC (1). This approach is also employed in the Foodborne Illness Risk Ranking Model (FIRRM), an analytic tool developed by FSRC to compare the public health impact of various pathogen-food combinations (14–16). As recent CDC outbreak data were not available when FIRRM was developed, FSRC relied on outbreak data compiled by CSPI. The CSPI database consists primarily (88%) of foodborne outbreak data compiled by CDC and now available on the Internet, but it also includes outbreaks not included in CDC data. CDC may not have received reports on these illnesses from state health departments or may have excluded reports from the database because they did not meet CDC criteria for foodborne outbreaks due to specific pathogens.

Improving food attribution from outbreaks will require improving both the quality and quantity of data. In particular, increased efforts are needed to obtain stool specimens from ill persons early in outbreaks to increase the fraction of outbreaks for which a pathogen is identified to >40% and to trace back foods implicated in outbreaks to their sources. CDC has launched an effort to improve the categorization of food items and ingredients, so outbreaks can be grouped in useful ways for regulatory agencies, industry, and consumers. CDC is also creating new analysis capabilities for the foodborne outbreak surveillance system that will provide data summaries for a variety of purposes, including food attribution.

FoodNet is an active surveillance program centered at CDC that tracks foodborne illnesses from 9 pathogens in 10 well-defined target populations (18–20). In FoodNet case-control studies, patients reporting through FoodNet are contacted for followup interviews and to complete questionnaires to estimate the proportion of illnesses associated with specific foods, food preparation, handling practices, and such behavior as pet ownership, farm visits, or international travel. FoodNet has performed case-control studies on a variety of pathogens, including Salmonella spp., Escherichia coli O157:H7, Campylobacter, Cryptosporidium, and Listeria monocytogenes (among others 17,21–24).

FoodNet case-control studies are of particular value for assessing food attribution of sporadic illness because they are population based. Because the diseases under investigation are rare in all population subgroups, rate ratios in the data closely approximate risk ratios in the population. Along with case exposure percentages, these risk ratio estimates may be used to calculate the "population attributable fraction," the proportion of new cases occurring during a given period in a particular at-risk population that was attributable to the effect of ≥1 exposures.

FoodNet case-control studies have limitations, primarily due to recall bias, long exposure windows, and immunity (20). First, patients and controls are limited in what they remember and can report in an interview, and the interview format itself has limitations. Second, the periods during which exposures are ascertained for FoodNet case-control studies tend to be long (5–7 days), so the likelihood of detecting a difference in exposure between cases and controls is limited by high exposure frequencies among both cases and controls. Further studies are needed to assess the consequence of using shorter exposure windows. Lastly, if a relatively common infection conveys durable immunity, an important segment of the population may be immune and therefore not susceptible to infection, making the demonstration of an association between exposures and risk for infection more difficult.

CDC's National Molecular Subtyping Network for Foodborne Disease Surveillance (PulseNet) links public health laboratories that use PFGE to routinely fingerprint suspected foodborne bacteria isolates (25). Results of PFGE subtyping of 5 bacteria (E. coli O157:H7, nontyphoidal Salmonella spp., Shigella spp., L. monocytogenes, and Campylobacter spp.) are stored in the electronic PulseNet database; bacterial strains in the database can be compared quickly and provide an early warning system for emerging outbreaks when related strains appear. PulseNet cannot currently be used for food attribution because it does not include isolates from sporadic cases of human illness or from food or animal sources.

MST refers to a specific application of microbial subtyping in which markers from an isolate are used to trace that isolate to an animal source, similar to what is done for salmonellae in Denmark. If different animal species carry unique, host-specific populations of microorganisms, a subtyped isolate drawn from an infected person could indicate that the isolate originated in 1 species as opposed to another. A large number of MST methods are being researched, most of which were originally developed to trace fecal bacteria in natural waters. Most approaches use genetic or phenotyping fingerprinting methods, although chemical markers, biomarkers, viruses, and bacteriophages are also used as indicators of animal source (26,27). Although MST techniques have potential, no single approach seems ideal for all pathogens and situations. US agencies have begun to research MST specifically for food attribution purposes; CVM, in particular, has investigated methods for Salmonella and Campylobacter spp (28). Although results to date are promising, they are only initial steps toward using MST methods to attribute illnesses to food animals. Similarly, data collected on the antimicrobial resistance of bacteria by researchers at CVM, CDC, the Agricultural Research Service (an agency of the USDA), and elsewhere may ultimately prove useful for food attribution purposes.

Risk assessments include food contamination data, food storage and consumption patterns, risk behavior, and dose-response functions to predict risks for illness from specific pathogens found in specific foods. If exposure estimates and dose-response functions are sufficiently accurate (a key consideration), risk assessments may produce excellent estimates of the true impact of illness.

Because risk assessments are so resource-intensive, they have been undertaken for only a limited number of pathogen-food combinations. The most comprehensive risk assessments for a single pathogen are those performed for L. monocytogenes by CFSAN and FSIS (29,30). This set of 23 individual risk assessments focused on ready-to-eat foods, including deli meats, dairy, produce, and seafood. Risk assessments have also been conducted on E. coli O157:H7 in ground beef (31), on Salmonella enterica serovar Enteritidis in shell eggs and egg products (32), and on Vibrio parahaemolyticus in molluscan shellfish (33).

For risk assessments to be used for food attribution, however, they need to be performed on most food items associated with a particular pathogen. Considering the 3-year duration of the L. monocytogenes risk assessments, performing comprehensive risk assessments on a sufficiently large number of pathogen-food combinations for systematic food attribution would be a colossal task.

The other major limitation of using risk assessments for food attribution is that they are inherently predictive. Unlike surveillance data, they do not measure observable public health effects, but rather estimate the impact on the basis of assumptions that are difficult to validate in a dynamic system, in particular dose-response functions, food storage and consumption patterns, and consumer behavior. Furthermore, risk assessments are ill suited for temporal analyses, since they are not routinely updated as new data become available. Risk assessments are most useful for food attribution purposes when compared with other estimates, such as those based on outbreak data or case-control studies.

Various US food safety agencies test for pathogen prevalence in foods through routine monitoring and case studies. FSIS monitoring, focused on the slaughter process, includes regular testing of raw meat and poultry for salmonellae (34), ground beef for E. coli O157:H7 (35), and ready-to-eat deli products for multiple pathogens (http://www.fsis.usda.gov/OPHS/rtetest/). FSIS and the Agricultural Research Service have examined pathogen prevalence in commercial food products, such as L. monocytogenes in frankfurters (36). The United States does not have a comprehensive program for monitoring live food animals. Food and animal monitoring data, when not associated with surveillance data, are not applicable for food attribution purposes.

When scientific or epidemiologic data are lacking, sparse, or highly uncertain, expert judgment may be used to fill gaps or combine conflicting estimates into a meaningful solution. Expert judgments derived through formal methods are increasingly used and recommended for assessing risk and the economic impact of regulations (37–39).

FSRC researchers administered an expert elicitation of experienced food safety researchers, public health scientists, and food safety authorities for use in the Foodborne Illness Risk Ranking Model (15). Produced with a standardized, vetted method, the survey asked respondents to estimate the percentage of 11 pathogens caused by each listed food category and included measures of respondent uncertainty and possible biases. Although data need to be analyzed further, initial results are promising and corroborate food attribution percentages derived from other means.

Expert elicitations are limited because they are based on perception, not on observable data. Results may be circular if experts rely on the same sources, or deceptive if experts are similarly misinformed or biased. Expert judgments are thus not an ideal source of food attribution data but may have utility if data are sparse or inconsistent and uncertainty is substantial.