commonly used in pollutant load estimation, either with ratio or regression approaches, and it might be adapted in this context to adjust the mean concentration. Its utility would be primarily in the flowing water systems.
A primary statistical objective of the proposed program is to estimate the national distribution of average annual concentrations of pesticides at the level of the individual CWSs. These data will in turn be used to assess the chronic exposure risks of the populations served by these CWS. The annual sampling program within each selected CWS should be optimized for the estimation of the annual mean concentration for that CWS. Based on the empirical evidence presented, this suggests that the number and temporal distribution of the water samples extracted and analyzed for each site should be tailored to the type of water source, the environmental fate and persistence of the pesticide, and any other observable factors that are known to effect the distribution of the pesticide's concentration over the annual period. The empirical data provide basic guidance on the nature of the annual sampling program for specific pesticides and water source types, e.g., the greater temporal stability in concentrations for reservoirs and lakes as opposed to rivers and streams and the illustrated longer-term persistence of atrazine vs. bromoxynil. The Agency suggests the use of prior simulations of annual concentration distributions for individual pesticides to guide the number and timing of water samples for each selected CWS. This is a very reasonable approach if these simulations are able to reasonably capture the shape of the true concentration distribution; however, care should be taken not to adopt sampling strategies that are not robust against nontrivial departures of the actual from the simulated distributions. This a particular concern for any simulation-based plan that highly concentrates the water sampling within a very narrow time frame of spiking concentration. A disadvantage to developing a sampling plan for water samples from a CWS that is adapted to the simulated distribution of a particular pesticide (e.g. atrazine) is that it may be completely inefficient for other pesticides (e.g. bromoxynil).
Finally, depending on permanent and seasonally-adjusted water treatment practices at the individual CWSs, a sampling plan that is optimized to simulated temporal distributions in concentrations for pesticides in raw water samples may not be optimal for treated, "finished" water outputs.
The Panel was concerned about the inferences in this discussion. One SAP member suggested that in part the Agency was asking about whether (and then how) sample sizes should be differentially allocated when there are different measurement errors associated with data collection. In particular, suppose that measurement error for population A is five times larger than that for population B. How should we apportion the available empirical effort to get estimates of the respective means that have roughly the same precision?
The measurement errors might be different in the two populations for reasons that have nothing to do with any systematic difference between them. For instance, suppose that the laboratory that made the chemical determinations for A had inaccurate protocols and reported error ranges that are 5 times wider than those for B, but not biased either way. Another source of