A European Union-wide baseline survey on Listeria monocytogenes was carried out in 2010 and 2011 with the aim of estimating the European Union-level prevalence of Listeria monocytogenes in certain ready-to-eat (RTE) foods at retail. A total of 3 053 batches of packaged (not frozen) hot or cold smoked or gravad fish, 3 530 packaged heat-treated meat products and 3 452 soft or semi-soft cheeses were sampled from 3 632 retail outlets in 26 European Union Member States (MS) and one country not belonging to the European Union. Two fish product samples from the same batch were analysed upon arrival at the laboratory as well as at the end of shelf-life, whereas the meat products and the cheese samples were analysed at the end of shelf-life. All 13 088 food samples were examined for the presence of L. monocytogenes, in addition to the determination of the L. monocytogenes counts. European Union-level estimates of the prevalence and of the proportion of samples with L. monocytogenes counts for the fish samples at time of sampling and at the end of shelf-life as well as for cheese and meat samples at the end of shelf-life were presented in the Part A report, published by the European Food Safety Authority on 27 June 2013. The present report, Part B, provides the results of further statistical analysis of the baseline survey data.
The primary objective of the survey was to obtain valid EU-level estimates of prevalence and contamination levels of L. monocytogenes in the categories of surveyed RTE foods, by collecting and utilising comparable data from all MS, through harmonised sampling schemes. An important characteristic of the data from this baseline survey that greatly affected their further statistical analysis, which is the subject of the present report, is that, even though a large number of samples were obtained during the survey, the variety of the obtained samples was very large and the number of L. monocytogenes-contaminated samples and the number of samples with counts exceeding 100 colony forming units (cfu)/g were small. This affected especially the attempts at data analyses aiming at identifying factors related to the prevalence of contaminated foods and at developing predictive models for the microbial growth of L. monocytogenes under various storage conditions. In the former case, problems due to sparseness were evident during the model-building process, while no factor models are presented for soft or semi-soft cheese samples, owing to the very low number of cheese samples that were found to be contaminated with L. monocytogenes in the baseline survey. In the latter case, after extensive analysis of the available eligible data from pairs of fish product samples for the development of predictive models for the microbial growth of L. monocytogenes, it was concluded that, given the limitations of the available information and of the nature and characteristics of the collected baseline survey data, these data were not appropriate for the development of satisfactorily accurate predictive models. The above-mentioned issues limited the number and strength of conclusions that were possible from the further analysis of the data of the baseline survey.
Multiple-factor models were used to examine the statistical association between several factors, on which information was gathered during the baseline survey, and two outcomes:
- Prevalence: a food sample was considered contaminated if L. monocytogenes was detected by at least one of either the detection or the enumeration methods (i.e. a sample was regarded as contaminated if either the detection test result was positive and/or the enumeration test result was positive, i.e. having a count of at least 10 cfu/g).
- Proportion of samples with an L. monocytogenes count that exceeded the level of 100 cfu/g.
Generalized Estimating Equations (GEE) methodology was used in order to account for the hierarchical nature of the baseline survey data. Six models were constructed: four models for fish product samples (for prevalence and for proportion of samples with an L. monocytogenes count that exceeded the level of 100 cfu/g, at time of sampling and at end of shelf-life) and two models for meat product samples (for prevalence and for proportion of samples with an L. monocytogenes count that exceeded the level of 100 cfu/g, at end of shelf-life). The results of the models are presented in the form of odds ratios (ORs) with corresponding 95 % confidence intervals (CIs) and P-values, estimated each time for a specific category (or level) of a factor, compared with another category (or level) of the same factor.
To facilitate the implementation, interpretation and feasibility of statistical models, some additional variables were defined, and some existing categorical variables were redefined by collapsing some of their categories into new ones. The large number of factors affecting the modelled outcomes (prevalence and proportion of samples with an L. monocytogenes count > 100 cfu/g) in combination with the, frequently, very large variability in the characteristics of the food items and the great imbalance in the distribution of the food items among the levels of several factors, made the analysis difficult, owing to sparseness problems. This was greatly exacerbated by the small number of baseline survey samples that were positive for the examined outcomes, which led, on many occasions, to very unequal distribution of the data among the categories defined by the combinations of the levels of all factors and of the modelled outcomes in each model, including, on occasion, combinations with zero frequencies. These problems were evident during the model-building process and also resulted in instability of the effect estimates of some factors during the sensitivity analysis. While some of the associations between the modelled outcomes and the examined factors were stable during sensitivity analysis, others were unstable with ORs and/or P-values of the same factor fluctuating importantly among different analyses. Care should be exercised when formulating statements about those factors that were unstable across different models. Therefore, the discussion of results focuses mainly on the factors which were significantly associated with the modelled outcomes, and exhibited consistent and stable associations in the presented models and the corresponding sensitivity analyses.
The variable 'Subtype of the fish product' reflected the different types of processing that the fish products had undergone. The OR of being contaminated with L. monocytogenes for 'Hot smoked fish' and for 'Unknown smoked fish' (fish which may have been hot or cold smoked) compared with 'Cold smoked fish' was significantly lower than 1, meaning that the odds of a sample being contaminated with L. monocytogenes were significantly lower for 'Hot smoked fish' and for 'Unknown smoked fish' than for 'Cold smoked fish', both at time of sampling and at the end of shelf-life. Concerning the multiple-factor models for the proportion of samples with an L. monocytogenes count exceeding 100 cfu/g, the variable 'Subtype of the fish product' was not included in either of the two models (for time of sampling and for the end of shelf-life), as there was no significant association between this variable and that outcome. Concerning the 'Number of antimicrobial preservatives and/or acidity regulators (AP/AR)' the OR of being contaminated with L. monocytogenes for samples with 'Two or more AP/AR' were over seven times those of samples with 'No reported AP/AR', both at the time of sampling and at the end of shelf-life. Conversely, whilst not statistically significant, samples with 'One AP/AR' had lower odds of being contaminated with L. monocytogenes than samples with 'No reported AP/AR' both at sampling and at end of shelf-life. Furthermore, no significant association with the 'Number of antimicrobial preservatives and/or acidity regulators (AP/AR)' emerged in the model for the proportion of samples with L. monocytogenes counts in excess of 100 cfu/g. At the time of sampling, the odds of being contaminated with L. monocytogenes in 'Sliced fish' were 1.59 times the odds in 'Not sliced' fish (P-value = 0.04). While the odds of contamination in 'Sliced fish' remained higher than in 'Not sliced' fish products, at the end of shelf-life, that result was not statistically significant. The OR of having an L. monocytogenes count above 100 cfu/g for a 'Sliced' fish sample compared with a 'Not sliced' sample was 2.79 (P-value = 0.07) at the time-point of sampling and 2.55 (P-value = 0.03) at the end of shelf-life. In addition to this consistent and frequently significant association across the four models, this finding appeared to be robust in sensitivity analysis using weighted analysis. Several other factors were included in the final multiple-factor models for the fish samples, for at least one of the two outcomes; however, the results were not always stable in sensitivity analysis.
Concerning the models for the two outcomes for the meat products at the end of shelf-life, the most stable associations with the outcome were found for the following factors: 'Type of the meat product', 'Possible slicing', 'Animal species of the origin of the meat product' and 'Remaining shelf-life'. The OR of a sample being contaminated with L. monocytogenes for 'Pâté' compared to 'Cold, cooked meat product' was 2.91 (P-value = 0.005). However, the odds of being contaminated with L. monocytogenes for 'Sausage' samples were not statistically significantly different from the corresponding odds for 'Cold, cooked meat product' (OR = 0.97, P-value = 0.93). Furthermore, the odds of a sample being contaminated with L. monocytogenes for 'Sliced' meat products were 2.13 times the odds for 'Not sliced' meat products with a P-value of 0.07, while the odds of a 'Sliced' meat product sample having an L. monocytogenes count above 100 cfu/g were 2.61 times the odds for a 'Not sliced' meat product but were not significantly different (P-value = 0.36). Concerning the variable 'Animal species of the origin of the meat product' the OR of a meat product sample having an L. monocytogenes count above 100 cfu/g for 'All other species' compared with 'Avian species' was 0.35 (P-value = 0.04). Finally, the corresponding OR for a meat product sample compared with a sample whose 'Remaining shelf-life' is one day shorter was 1.010 (95 % confidence interval (CI): 1.005, 1.016) which was statistically significantly higher than 1 (P-value = 0.0002). Based on the results of the multiple-factor analysis it can be recommended that food business operators producing cold smoked fish, pâté or sliced ready-to-eat smoked or gravad fish and heat-treated meat products might actively reconsider food safety management systems and their ongoing verification, in particular with increased attention to environmental L. monocytogenes sampling in the area of the slicing process, in order to ensure effective control of L. monocytogenes in their products.
The final Term of Reference (ToR) for the work presented in this report, Part B, required the development of predictive models for compliance with L. monocytogenes food safety criteria in foods. Commission Regulation 2073/2005 mentions two microbiological criteria applicable for RTE foods at different stages. The criterion with which compliance might usefully be considered at the retail stage is the requirement for RTE foods not to harbour L. monocytogenes counts in excess of 100 cfu/g at the end of shelf-life. The fundamental requirement to predict compliance from this prevalence survey, therefore, involves consideration of what a survey of single-unit samples (n = 1) might represent for the surveyed population of RTE foods if a multiple-unit sample approach (n = 5) had been followed. In statistical terms, the probability of compliance for this exercise was defined as the probability that no individual unit, out of n = 5 units constituting a sample taken from a population of RTE foods, exceeds the level of 100 cfu/g, at the end of shelf-life. The estimation of this probability is based on an estimate of the proportion of samples with L. monocytogenes counts exceeding the level of 100 cfu/g obtained from a single-unit sample survey in the same population of RTE foods.
A statistical model was developed for this purpose and is presented in the current report. An illustration of the application and results of the developed model is provided by using the baseline survey data for fish, cheese and meat product samples. This method may have some utility when, for example, a Competent Authority has carried out a prevalence survey in a population of RTE foods, based upon a representative sampling plan, and wishes to make some assessment of compliance within that population of RTE foods. It has to be noted that the potential utility of such a statistical method would not alter the obligation on food business operators, which explicitly remains in Commission Regulation 2073/2005, to analyse n = 5 samples, in order to demonstrate compliance. The statistical methodology developed and applied in this report should not be seen as a way to facilitate demonstration of compliance by food business operators using fewer than five sample units.