In 1996 the U.S. Congress instructed the U.S. EPA to study the toxicity of mixtures of chemicals in drinking water, with a particular focus on potential synergistic and antagonistic interactions. Assessing the net toxicity of every potential mixture of chemicals is a daunting task; as a complementary approach, the U.S. EPA and the ATSDR propose to estimate mixtures' toxicity based on the toxicities and interactions of component chemicals. Both approaches require substantial toxicological information. However, large compilations of water-quality data have recently become available from federal and state agencies. This thesis demonstrates that by using environmental data to supplement available toxicological data in a decision-analytic framework, regulators may set priorities for future research and regulation and examine the assumptions and sensitivities of proposed toxicity models.

National data on the occurrence of binary and ternary mixtures of arsenic, cadmium, and manganese are used to parameterize a modified hazard index model for 129,000 water sources. Hazard index scores are computed for these mixtures' potential effects on the cardiovascular, hematological, neurological, and renal systems under several chemical interaction scenarios. National and regional statistics are presented to describe the distribution of hazard estimates for drinking-water sources used by the U.S. population. These results provide a framework to evaluate the value of collecting additional toxicological information for each mixture and target organ. For example, according to the model, the levels of arsenic and cadmium found in U.S. drinking water are unlikely to have synergistic cardiovascular effects, but the same mixture's potential for synergistic neurological effects merits further study. Similar analysis could be used to prioritize toxicological studies based on their potential to reduce scientific and regulatory uncertainty.

These findings are not complete risk assessments; the contributions of this thesis are primarily methodological. The text explores several issues in the use of environmental data in risk assessment, including the use of recent semi-parametric statistical methods for the analysis of censored and incomplete chemical data, and spatial extrapolation from environmental samples to population exposure estimates. The analysis demonstrates that the benefits and costs of regulation may accrue to strikingly different regions and populations for individual versus mixture assessments, even for small groups of chemicals. Significant recommendations include the promotion of exposure assessment to an early stage of risk assessment; the development of new data reporting standards to reduce censoring and improve spatial coverage in environmental data; the development of alternative toxicity models describing a variety of interaction mechanisms such as equitoxicity, potentiation, and inhibition; and the use of empirical, data-driven methods to explore the implications of these alternative risk models for complex mixtures of many chemicals.