“Human illness rates from Salmonella have not budged,” according to Dr. Elisabeth Hagen, Under Secretary for Food Safety at the United States Department of Agriculture, in a 2011 speech. Despite that observation, government agency websites and peer-reviewed papers continue to search for a correlation between annual figures for human salmonellosis and Salmonella prevalence in hazard analysis and critical control points, or HACCP, regulatory samples over the last 15 years.

Testing for correlations in annual Salmonella numbers has the problem that these figures typically include all data with no recognition of known sources of variability in the numbers. Both FoodNet and HACCP Salmonella data, for example, show distinct seasonal variation with large numerical and temporal differences between high and low points.

Sampling method not closely related to human risk  

Figure 1: Human salmonellosis cases by month shows the strong seasonal pattern of human Salmonella illnesses in the FoodNet surveillance area between 1996 and 2004, with disease curves becoming smoother as the population under surveillance increased from approximately 14 million to 44 million. Cases peak in July or August at a level about three times higher than the February lows, a seasonal pattern that has been observed in other countries as well, with the peaks occurring in warm months of the year. The top five human serotypes (Enteritidis, Typhimurium, Newport, Javiana and Heidelberg) cause about 50 percent of cases, with all other serotypes causing the remaining cases.

HACCP Salmonella testing started in 1998 with an approximate 50 percent reduction in Salmonella prevalence in chicken samples in that year when compared to the broiler chicken baseline sampling in 1994-95. Despite that apparent reduction in Salmonella carried by chickens, no obvious change in human salmonellosis can be seen in the graph in that time period. The method for sampling broiler chickens for Salmonella does not appear to be closely related to the risk of human illness from chicken.

Seasonal change in Salmonella less than increase in human cases  

FSIS has not released monthly chicken HACCP data to compare directly with the FoodNet pattern in the first graph, so averages have to be presented to make that comparison. Figure 2: Salmonella in humans and chickens shows the percentage of monthly Salmonella-positive samples from the first seven years of broiler chicken HACCP (1998-2004) plus FoodNet monthly data from the first graph averaged over the same seven years.

FoodNet averages show a peak in human cases in July with a low point in February. Chicken Salmonella prevalence peaks in September at 16.8 percent, versus a low of 9.8 percent in April and May. Salmonella prevalence in chicken peaks at about 1.7 times greater than the low figure, so the seasonal change in chicken is less than the three-fold increase in human cases.

Relationship between Salmonella in chickens, humans obscured  


Figure 3: Human-chicken Salmonella difference emphasizes periods during the year when the human and chicken Salmonella data are a poor match. After the July peak in human illness, Salmonella prevalence in chicken rises for the next two months while human cases are declining relatively sharply.

When human cases are increasing after their February low point, chicken positives are decreasing. If Salmonella in chicken is a major determinant of human illness, the curves for chicken should change concurrently or slightly ahead of the human curves, because of the lag time between in-plant sampling, distribution to consumers and incubation times before human illness is detected.

The complexity of seasonal variation cannot be summarized effectively with a single annual number for Salmonella in either human or chicken data. Correlations relying on analysis of a series of arbitrary annual numbers for both human and chicken Salmonella are deeply flawed and obscure any true relationship between the two patterns.

More noise than signal in Salmonella data  

There are several other problems with direct comparisons between the human and chicken data. Human illness statistics from FoodNet include serotypes that are rarely seen in chicken samples. Newport and Javiana, for instance, are among the top five serotypes causing human illness, but are rarely found in chicken HACCP samples. Kentucky is overwhelmingly predominant in chicken samples, accounting for about 42 percent of all HACCP isolates in 2004, but Kentucky made up less than 0.2 percent of human cases in the same year.

In the 1998 to 2004 period, Salmonella serotypes that each caused less than 1 percent of human illness made up between 43 percent and 53 percent of all chicken isolates. The substantial difference between serotypes seen in human illnesses and in HACCP samples indicates that serotypes occurring in meat and poultry are not all equally likely to cause human disease.

When data are presented and analyzed by correlation between yearly numbers for Salmonella observed in human illness and in chicken HACCP, with all seasons and serotypes averaged together, there is far more noise than signal in the resulting analysis. There is no doubt that some human salmonellosis comes from chicken, but there is no way to establish an actionable link between the two without a more detailed and complete analysis of the available data. Monthly HACCP Salmonella data should be available to allow a more sophisticated analysis than is possible with presently available information.

If the signal in the raw data cannot be detected without enhancement by log-linear Poisson regression, volume-weighted prevalence, moving averages or other statistical manipulation of the results, then there has not been much of an effect after 15 years of effort and hundreds of millions of dollars spent to control Salmonella in poultry.