Abstract

The RANS team at Fera works with various government agencies, regulatory bodies and software developers in the UK and EU to try and develop and apply improved methods for probabilistic risk assessments across a range of policy areas. My work has been focused on quantifying uncertainty and variability within risk assessment for dietary exposure to pesticides. For risk managers it is clearly important to separate uncertainty from variability, so they can decide whether they should be collecting more data to reduce uncertainty or taking steps to reduce inherent variability by changing the system.

2-dimensional Monte Carlo (2DMC) algorithms have been used as the standard approach to separately represent uncertainty and variability, but these are frequently used incorrectly. Lognormal distributions are often assumed for the variability distribution as standard without justification. Also, many sources of uncertainty are ignored, or treated using ad-hoc methods. Problems include:

· Limited residue measurements from random market surveys

· Extremely high proportions of non-detects or missing residues

· Composite measurements from a small set of food units treated as a batch/field mean residue

· Measurement uncertainty of residues

· Limited dietary survey used to infer whole population eating habits (over a lifetime)

· Pesticide residues can come from multiple sources, e.g. a child might eat an apple and a pear on any given day, both of which were treated with pesticides

· Lognormal often fits badly, but what is the ‘right’ distribution?

I will talk about how Bayesian methods have been applied in a range of applications and list some open research problems. Computational aspects will also be mentioned.