Quantitative risk assessments (QRAs) are done for specific purposes and provide numerical risk estimates to answer questions that were posed by the risk managers who originally commissioned the assessment. In the seafood area there have been three QRAs:
The United States risk assessments were very large, taking more than one year to prepare and then moving to a 1-2 year review period of public comment. The L. monocytogenes risk assessment involved more than 30 people arranged in six teams, each of which was assigned specific tasks; more than 50 additional participants were acknowledged for their assistance. It must be stressed that this QRA involved a range of foods, not just seafoods, but the QRA of V. parahaemolyticus in oysters also involved more than 20 people who received information from scientists at more than 20 institutions in the United States and internationally. The Swedish QRA had two authors and acknowledged the help of two collaborators.
In a QRA, it is vital to define what you want the work to achieve, and to do this right at the beginning. This is called a Statement of Purpose. In the United States, the risk managers stipulated that, for V. parahaemolyticus in oysters, the risk assessors:
1. produce a mathematical model of the risk of illness incurred by consumers of raw oysters containing pathogenic V. parahaemolyticus;
2. provide the regulators with information to assist with reviewing current regulations to ensure that they protect public health by evaluating:
For L. monocytogenes, the Statement of Purpose was to examine available scientific data systematically in order to estimate the relative risks of serious illness and death that might be associated with consumption of different types of ready-to-eat foods that might be contaminated with L. monocytogenes. The work produced mathematical models to predict contamination at the retail level and in the home, and different consumer groups were included in the assessment. The result was predicted rates of listeriosis from various foods for various at-risk groups.
In Sweden, Lindqvist and Westöö (2000) set the objective to develop a QRA for estimating the exposure and risk of acquiring listeriosis from consumption of packaged smoked or gravad salmon and rainbow trout.
In the seafood industry, the process is usually stretched out from harvesting, storing prior to processing, processing in the seafood plant, storing/distributing, retailing and consumption. Whatever the seafood product you are considering, the hazard may change throughout the process, either in prevalence or in concentration. We need to chart these changes often by making a process flow diagram and then mathematically measure or estimate changes in the hazard at each stage. In risk assessment this is called "modelling". Usually modellers try to make a "farm-to-fork" model that takes in changes to the hazard all along the harvest-process-consumption route. This part of the risk assessment is best done by people who understand the industrial process and combined with microbiologists who understand the hazard and how it reacts to changes, particularly to changes in temperature and time.
When the model of the system has been set, data must be gathered (exposure assessment). Ideally, there would be time to carry out experiments that give you exactly the data you need but, almost always, there are not sufficient resources or time to do this. So you need to investigate all sources of existing data and try to incorporate them into the model. This is where the modeller on your team takes the data and constructs mathematical relationships that describe changes in the hazard throughout the process. The modeller will encounter a number of problems, the most common being variability and uncertainty.
This occurs because of the diversity in any population, and it cannot be reduced, no matter how much the property is studied. To illustrate, let us use height as an example. In any population there is variability in height. We could do a survey by measuring how tall people are, and we would find most adults are 160-175 cm tall but that some are 220 cm while others are 120 cm. This is an example of variability within a population.
This is due to our (the risk assessor's) lack of knowledge about a parameter and our inability to measure it. Uncertainty can be reduced if we study the characteristic. Using the same example of peoples' height, we could do a national survey and measure everyone. Then there would be no uncertainty.
The risk is never fixed - it varies according to a range of parameters. For example, take the risk of dying in an air crash. For the vast majority of people on this earth the risk is zero because they never fly but, among those many millions who do fly, the risk varies according to how often they fly (likelihood), the airline (some have more crashes than others), the weather conditions (many crashes occur in bad weather) and the country (some have better systems than others). So estimating the risk is difficult because there is a distribution of risk from very low, through average to very high. Often the best estimate of distribution is minimum, most likely (average) and maximum value. For example, we might say the bacterial levels of shrimp landed aboard a trawler ranged from 10/g to 10 000/g, with the most likely count being 100/g.
Modellers generally use simulation or stochastic modelling in which data are inserted into a spreadsheet. Computer software is then used to analyse the data. Each analysis is called an iteration where a value is selected from the distribution describing each variable range, more or less at random, but according to the probability distribution of that variable (more likely values are run more frequently than minimum or maximum values). A large number of iterations is run (10 000 is a popular number) and collated; the technique is called Monte Carlo simulation. The result is a distribution frequency of possible outcomes, which forms the basis of the risk estimate.
The way you estimate the risk in a QRA is usually set by the statement of purpose. For example, Lindqvist and Westöö (2000) estimated the risk of acquiring listeriosis, and so risk estimates included the number of cases per annum and risk of becoming ill on a per serving basis. The researchers used two models and so had two estimates for each output. In the United States, the relative risk of acquiring listeriosis from a range of foods was the estimate, with pâtés, smoked seafoods, soft cheeses and delicatessen meats being the four most likely to cause the illness. For V. parahaemolyticus in oysters the single most important factor related to risk of illness was temperature - of air and water (seasonality). The model predicted nationwide illnesses of 4 750 per annum with a range of 1 000 to 16 000 cases. The model also indicated that risk of illness was reduced if product temperature could be lowered soon after harvest.
When you have the risk estimates it is a good idea to do a reality check to see that the model is not predicting something that will seem absurd. For example, suppose you are estimating the number of cases of listeriosis caused by consumption of smoked fish and the model predicts the most likely scenario of 1 million cases each year. If your country statistics on illness and death state that there are 1 000 such cases each year, you know there is something wrong either with the model or with the inputs. You have more work to do!
As the software grinds through the iterations it also keeps a record of which factors have the biggest effect on risk estimate. This allows you to do sensitivity or importance analysis to identify those factors most influencing risk - either reducing or increasing it. This analysis then points risk managers to those areas where process control can be increased.
Risk assessments range in complexity from qualitative, through semi-quantitative to quantitative. As assessments become more complex, they also become more expensive and take longer to complete. So before you begin a risk assessment be sure you know exactly what you want or you may end up using resources unnecessarily.