The definition of domains (collection of strata for analysis) that correspond to pesticide use areas and the differential sampling of CWSs based on a measure of the potential concentration in the water to be tested ("vulnerability") is a good stratification choice for the single pesticide. If a short list of priority pesticides can be established and the statistic of primary interest is the 95th percentile of the annual mean concentration for CWSs, then the use of strata defined by "vulnerability" or modeled predictions of concentrations of pesticides is a more optimal strategy compared to simply sampling at random from the CWSs within each use area domain. Optimal definition of strata boundaries and allocation of the total sample size to the defined strata will depend on the shape of the distribution of CWS annual mean concentrations including the proportion of CWSs that test below the LOD for the compound and the distributional skewness and thickness of its upper tail. Since these distributions will vary from one pesticide to another and possibly from one geographic region to the next, care should be taken to not rely totally on a priori predictions of the form of these distributions. It is important to emphasize that while oversampling strata of CWSs of expected higher concentration of pesticides can be efficient, the stratified sampling should provide representation from the full set of strata that have been defined for the pesticide use area.
Conversely, any stratification plan that ignores existing information (i.e., use areas, intensity of application) for the major pesticides or fails to discriminate these pesticides in developing the "vulnerability" of the CWS is highly inefficient from the standpoint of estimating the distribution of concentration for a set of pesticides that should have a priority in the Agency's initial research.
While stratifying by vulnerability is important, random sampling has a feature that allows an inference about the statistical population (a.k.a. "domain" of interest) under the assumption of representativeness. While it may sometimes be inefficient, it is typically reliable. Even though the Panel did raise the advantages of random sampling, the advantage of stratification is that it can focus limited empirical effort on events where regulation can make a difference. Thus, the Panel concluded that stratifying by vulnerability would be much more useful than stratifying by geography.
Therefore, the Panel raised the question - How should vulnerability be defined? A very crude system could be used that calls vulnerable any small or medium CWS in a higher pesticide use intensity region. (A county that is in the top quartile of use for any compound is said to be in a "higher pesticide use intensity area".) It is unclear how this might work in practice. The chance that a watershed is in the top quartile of at least one of 25 chemicals is a Bonferroni problem. Of course it depends on the statistical associations among the target chemicals, but it would be surprising if it is not very close to one. This would mean that virtually all small and medium-sized CWSs are vulnerable.
The Agency's interest appears to be in defining vulnerability in a more refined way. It was claimed that stratifying by vulnerability could give the study design more statistical power to address its goals. But where are the power calculations that suggest this? The Agency has not