Hazard data provide information about the presence and quantity of contaminants in environmental media. A hazard has the potential for harmful effects, but its presence alone may not be sufficient to produce an adverse effect in a population. Most environmental data collection by federal and state environmental agencies is mandated by legislation such as the Clean Air Act (1970), Clean Water Act (1977), and Safe Drinking Water Act (1974), and data are used for regulatory purposes. Examples of environmental data include the Toxic Release Inventory (TRI; http://www.epa.gov/tri/), criteria air pollutant data (ozone, sulfur dioxide, particulate matter), pesticide exposures, and the Safe Drinking Water Information System (SDWIS; http://www.epa.gov/enviro/html/sdwis/). Data analysis is often limited to comparison of each data point with an environmental standard or guideline, and regulatory action is triggered if the data point exceeds the standard. Standards are not available for all environmental contaminants, and those that do exist are extrapolated primarily from toxicologic studies. Standards are developed for regulatory purposes and include uncertainty and safety factors that are meant to be protective of public health; exceeding the standard does not predict health outcomes in the population.

Hazard data alone are not measures of individual exposure; in some cases, certain types of hazard data; for example, monitoring contaminants in drinking water, may contribute to the characterization of population exposure. Table 1 lists limitations in the use of hazard data for EPHTN.

Exposure data are the essential link between environmental hazards and health outcomes. Optimally, exposure data include biomarkers such as a compound or its metabolite(s) in a biologic sample (e.g., blood, urine, hair, fat). Levels of agents in blood or urine are not necessarily proportional to environmental concentrations. Additionally, factors such as exposure level, route of exposure, frequency and duration of exposure, baseline health status, behavioral factors, and genetics influence the internal dose of the compound in an individual.

Exposure data represent the largest data gap for EPHTN. Ideally, exposure data would be available at the individual level, but very few of these types exist. Childhood lead levels are one of the few nationwide exposure data sets. The CDC initiative to monitor exposure of the general population to 116 environmental chemicals through biomonitoring as a part of the National Health and Nutrition Examination Survey (NHANES) is providing essential baseline exposure data (CDC 2003b). However, there are no other systematic biomonitoring programs that provide nationwide exposure data. Data from sporadic, single-event biomonitoring that occurs in response to specific incidents or investigations are not widely available. Another limitation is the rapid metabolism and clearance of many compounds of concern. Because of the limited availability of exposure data, these parameters are often estimated rather than measured. Exposure modeling and assessments, as used in regulatory risk assessment, are meant to be protective of public health and therefore use conservative assumptions that overestimate actual exposure (e.g., consumption of 2 L/day of water for 70 years; U.S. Environmental Protection Agency 1989). Multiple routes of exposure and chemical mixtures are difficult to estimate. Individual factors that modify exposure, such as age, race, gender, time near the hazard, behavioral factors, and genetics, are often not incorporated into the exposure assessment models. Inexact exposure estimates limit the ability to ascribe health outcomes to a contaminant.

The lack of exposure data further complicates issues because the level of exposure influences the appropriate health end point. At high exposure levels, health effects are observed within a relatively short time frame and are often documented in the scientific literature. Low-dose chronic exposures present greater challenges: health outcomes may not be known; effects of low-dose exposures are not observable for years or decades; effects may require repeated exposures over time; multiple agents may need to interact simultaneously or sequentially; and the effect may occur only in sensitive subpopulations or those with existing health conditions. Table 1 lists additional limitations in using exposure data for environmental public health tracking.

Adverse health events such as morbidity and mortality are the outcomes of interest in efforts to associate environmental hazards and exposure with effects in humans. Of particular interest are asthma, birth defects, cancer, and sequelae of lead and pesticide poisoning/exposure. Health events may be identified and evaluated at the individual level, as is the case in traditional epidemiologic studies, or may be aggregated for populations, such as those available from local, regional, and national surveys. The uses and limitations of population-based aggregate data can differ from those of individual-level data, and this affects the use of each type of data.

EPHTN efforts focus on the use of existing (secondary) data rather than on the generation of new data (primary). Sources of such data include local, regional, and national health surveys such as the NHANES and the Behavioral Risk Factor Surveillance System, and health events and reportable disease registries such as those for cancer—the National Cancer Institute’s (NCI) geographic information system (GIS) web site (NCI 2004b)—and birth defects. Other sources include state vital records, hospital discharge data, emergency room data, and health insurance statistics. Several potential limitations must be considered when using secondary HOD, including those related to evolving or changing diagnostic criteria, misclassification, generalizability, measurement error, and completeness. Secondary HOD are also limited because many diseases/conditions are not reportable, resulting in incomplete information for many outcomes of interest. Further, confidentiality may prohibit access to the data at the level of detail that might be needed for correlation with an exposure, for example, address of an individual. Table 1 lists additional limitations in using HOD for EPHTN.

The association of a disease with a risk factor in epidemiologic research requires more than observing corresponding fluctuations between the two factors. For example, examining the effect of air pollution as measured by criteria pollutants on the incidence of asthma must consider related factors. Other factors may include residence, proximity to known asthma-causing sources, socioeconomic status, age, race, and adherence to treatment regimens that may be related to incidence and hazard/exposure. These characteristics related to the factors of interest must also be considered and controlled. In epidemiologic research, strict control is not feasible, in the sense that these characteristics are not assigned by the researcher. Instead, important factors are observed, measured, and controlled in the analysis. Information regarding many of the standard covariates is routinely collected and made available through the U.S. census and surveys, including information regarding race, age, and sex distributions, as well as a variety of variables describing socioeconomic status, employment, and housing statistics, among others. These data are aggregated by specified units (e.g., state, county, census tract) and have the same limitations identified above for HOD. The decennial census data have limitations with respect to how current or complete the data are for a given time period. Migration is problematic, especially in estimating populations for small areas. Nevertheless, these data provide important information regarding population characteristics necessary for establishing denominator estimates and associating hazard/exposure with health outcomes.

Examining time trends in HOD is helpful in identifying disease clusters. An example is the observation of birth defects (rare limb malformations) within a short time among women who took thalidomide for morning sickness. This cluster helped to identify thalidomide as a human teratogen (Taussig 1962). Trends identified through hazard surveillance are also important for characterizing background and changes in environmental contamination. Seasonal patterns are also identified through surveillance, for example, deaths from heat waves in the summer and pneumonia and influenza deaths in the winter season. Statistical approaches to the detection of clusters of disease include cell-count methods that compare observed with expected counts of events (Knox 1964; Openshaw et al. 1987, 1988), adjacency methods that examine whether areas of high rates of disease are likely to be adjacent to other high-rate areas (Moran 1948), and distance or nearest-neighbor methods that compare physical distances between cases to expected distance (Besag and Newell 1991; Cuzick and Edwards 1990; Mantel 1967). Alexander and Boyle (1996) provide extensions to the basic methods. Other statistical methods can be used to determine interarrival times between rare disease cases or to model seasonal patterns using time series methods, autoregression methods, and joinpoint regression (Kim et al. 2000). These trends are informative; however, confounding may prohibit correct interpretation. An example is an observation in prostate cancer incidence: A study was conducted to examine the prostate cancer incidence in a plant manufacturing triazine (MacLennan et al. 2002). Subsequent investigation of prostate cancer incidence indicated that the increase may have reflected increased prostate-specific antigen (PSA) screening rates rather than increased incidence of disease.

Environmental and health data have very few commonalities; in many cases, the only commonalities of the data are the general geographic area and time frame. GIS is a useful tool for examining each type of data and one of the few methods that can be used to compare disparate health and environmental data. These include the “brush and link” exploratory data tools such as GeoDa (Anselin et al. 2004) and identification of space–time patterns as available in SaTScan software (SaTScan 2004) and the TerraSeer Space Time Intelligence System (TerraSeer 2004). The common denominator is the geocode for each type of data. Issues arise in reconciling the point of sampling and the location (and often lack of location) of individuals with the health outcome. Nonetheless, simple mapping of each data set and overlay of the maps can assist in targeting geographic areas. GIS software often comes with an array of tools to produce maps with impressive visual displays of data, overlay of disparate environmental and HOD, and conduct trend analysis. Because of this, caution is necessary to ensure that the maps are not misleading. This can occur when health and hazard data are mapped independently and there is confounding in the data that may lead the investigator to an inappropriate conclusion. For example, the potential confounding due to the latency period and ecologic fallacy must be considered in interpretation. Additionally, the residential address of a cancer case at diagnosis may not be a good proxy indicator for the lifetime environmental exposure given the mobility of the U.S. population. Geocoding methods need to be developed that allow better cumulative estimation of exposure that individuals might experience during their lifetimes. These should be available for regular or irregular points in residence or occupational histories.

Disease mapping summarizes the spatial and temporal–spatial variation in risk. When compared with spatial and temporal–spatial variation in exposure/hazard, one may get some disease etiology clues. The National Cancer Institute’s Cancer Mortality Atlas (NCI 2004a) and State Cancer Profiles (NCI 2004c) show county-specific maps that have been instrumental in identifying cancer disease patterns with the location of major industries by means of ecologic studies—nasal cancer in areas with furniture manufacturing (Brinton et al. 1977), lung cancer in counties with petrochemical manufacturing (Blot and Fraumeni 1976), bladder cancer where chemical industries were located (Hoover et al. 1975), and oral cancer in regions where snuff use was common (Blot and Fraumeni 1977). Comparing county-level health rates from maps requires that rates be adjusted for potentially confounding variables such as age, race, sex, and socioeconomic status. Elliott et al. (2000) discuss problems in the comparison of indirectly adjusted rates [standard mortality ratios (SMRs)] and directly adjusted rates [comparative mortality figures (CMFs)]. CMFs may be unstable if the stratum-specific rates are based on small numbers; however, they are unbiased estimates of the relative risk to the standard population. SMRs may be biased if proportionality assumptions are not met. Neither of these rates (SMRs, CMFs) adjusts for the overdispersion typically found in this type of data. Nevertheless, Breslow and Day (1987) observe that the use of SMRs or CMFs often leads to the same results. A clearer exposition of which summary measure to use would be a welcome addition to the literature. Maps of rates can be very deceptive if they are composed of different-sized units; for example, large units are overwhelming, and small units are often lost in interpretation. The microplot map (NCI 2004c) shows 95% confidence intervals for the units, and their intervals may overlap even though a choropleth map indicates a difference. Additionally, maps at different scales may give different patterns, further obscuring the interpretation. Observed rates are quite variable, especially when the expected rates are small, so maps of county-specific rates are smoothed by a variety of means to enhance the visualization of regional patterns of distribution. Smoothing methods need to consider and adjust for differences in the population size (denominator) for different geographic units. Methods to compute and visualize the spatial and temporal–spatial variability in disease/mortality controlling for such covariates as age, race, sex, and deprivation are currently available, as are extensions of the method to smooth the data and model heterogeneity and clustering of the area-specific effects in both the traditional approach and Bayesian framework (Banerjee et al. 2004; Wakefield and Elliott 1999).

These temporal, spatial, cluster, and exploratory methods have standard software to analyze the data. These should become standard procedures in health departments. Problems in interpretation will arise because of geographic resolution of disease and environmental data, data quality, confounding, and ecologic bias. Replication of the analyses in widely differing settings should provide scientists with background information to make the judgment as to whether to abandon the examination of a particular health outcome–hazard link, replicate the study, and/or proceed to an ecologic study if significant effects are apparent or sufficient evidence to warrant further investigation exists (see “Descriptive analysis” in Figure 1).

Ecologic studies describe the coexistence of risk factors with disease among and within populations. Aggregate exposure data are correlated with aggregate health data for each unit of observation, usually defined by geographic or administrative boundaries (e.g., city, county, state). Rates of exposure and rates of disease within each unit are known, but the exposure status of diseased individuals is not known. Analysis of ecologic studies can be conducted visually by interpreting the slope of a line plot of the exposure rate by the disease rate for each unit or using the correlation coefficient, r, as a measure of association. In ecologic trend analyses, changes in rates of exposure are correlated with changes in rates of disease over time. The most salient limitation of these studies is ecologic bias, resulting from spuriously ascribing aggregate-level observations to individuals. These studies are useful for generating hypotheses, for assessing the impact of community-level interventions, or for initial evaluation of suspected associations, although more rigorous studies are necessary to support a causal relationship between exposure and disease.

These studies model the interrelationships of hazard, exposure, and health over time and space. The objective is to relate environmental variables to disease and control for other factors such as life style (Banerjee et al. 2004; Wakefield and Elliott 1999). Poisson regression provides the framework for modeling the rates for rare diseases, whereas binomial regression or survival analysis is suitable when disease is more common. Counts or rates of events are described as a function of exposure, space, time, demographics, and other variables. The models are hierarchical when hazard data are used and are thus subject to ecologic bias. Spatial correlation in the data should be anticipated. For example, screening for PSA is likely to be correlated among counties when the screening initiative is directed toward a large area. Overdispersed data are an additional problem. The methods for estimating parameters relating hazard to health outcomes in these models are by means of likelihood methods and by Bayesian methods. Advantages of the likelihood methods include the readily available software and, in some cases, the incorporation of dispersion common in these data. Disadvantages include the difficulty of specifying complex covariance structures in the data, the unreliability of SMRs based on sparse data, and the problems of estimating variances of the relative risks. Bayesian methods applied to the data require the assumption of a prior distribution on these parameters from which a posterior distribution can be used to provide point (mean, median) and interval (95% confidence intervals) estimates for the parameters of interest. These estimates are obtained by means of Markov chain Monte Carlo methods using WinBUGS software (MRC Biostatistics Unit, Cambridge, UK; http://www.mrc-bsu.cam.ac.uk/bugs/welcome.shtml). Advantages of the Bayesian methods include the use of random effects to represent overdispersion, the availability of WinBUGS software for computation, and the provision of smoothed estimates of SMRs based on sparse data. Disadvantages of the Bayesian methods include the sensitivity of the estimates to the choice of the prior, and problems of convergence in complicated models (Elliott et al. 2000). Wakefield and Morris (1999) extend their analyses to accommodate spatial correlation and overdispersed data in the Bayesian analysis of the relationship between heart disease mortality and magnesium and calcium in water. Another example using Bayesian methods is a study measuring ozone levels at 10 fixed sites and emergency department visits—total visits and those for asthma by ZIP code—to assess the effect of ozone levels on pediatric emergency department asthma visits. In this example, the data for ozone and emergency department visits are “misaligned,” and before the data can be analyzed, a surface (kriging) for ozone from the site data must be developed so that an ozone level can be estimated for each ZIP code. Carlin demonstrates the use of hierarchical Bayesian models in solving this and other misalignment problems (Banerjee et al. 2004)

Multilevel models estimating hazard and health outcome effects and controlling for potential confounders and covariates provide hypothesis-generating information; however, they will likely require more refined hazard and disease data. Statistical methods are available but are not trivial to run and interpret, and the potential for ecologic bias remains. Replication of these studies in several areas should be guided by the descriptive analyses. Judgment of the scientific community is necessary to decide whether to halt studies at this level or, if significant effects and consensus of the scientific community indicate, then to design an in-depth study (etiologic research) (“Ecologic analyses” in Figure 1).