DEALING WITH UNCERTAINTY IN RISK ASSESSMENT

This paper is a commentary on Hattis' three laws of risk assessment. T he first law, that ''application of standard statistical techniques to a single data set will nearly always reveal only a trivial proportion of the overall uncertainty in the parameter value'' is illustrated bo th by examining the relevance of animal models to man and by a retrosp ective view of exposure conditions whose importance has only recently been recognized to be important. The second law, that ''any estimate o f the uncertainty of a parameter value will always itself be more unce rtain than the estimate of the parameter value,'' is examined in terms of a model addressing multiple levels of uncertainty, e.g., the ''unc ertainty in the uncertainty''. A argument is made that the number of t erms needed for convergence of this uncertainty hierarchy depends on h ow far from the central tendency of the risk distribution one goes. Th e further out the ''tail'' of the distribution, the more terms in the uncertainty hierarchy are needed for convergence. The third law, that ''nearly all parameter distributions look lognormal, as long as you do n't look too closely,'' is illustrated with a number of examples. Seve ral reasons are put forward as to why risk variables appear so frequen tly to be lognormal. Recognition of the lognormal character of variabl e distributions can provide insight into the proper form for the assoc iated uncertainty distributions.