Estimation of exposure to toxic releases using spatial interaction modeling
© Conley; licensee BioMed Central Ltd. 2011
Received: 14 September 2010
Accepted: 21 March 2011
Published: 21 March 2011
The United States Environmental Protection Agency's Toxic Release Inventory (TRI) data are frequently used to estimate a community's exposure to pollution. However, this estimation process often uses underdeveloped geographic theory. Spatial interaction modeling provides a more realistic approach to this estimation process. This paper uses four sets of data: lung cancer age-adjusted mortality rates from the years 1990 through 2006 inclusive from the National Cancer Institute's Surveillance Epidemiology and End Results (SEER) database, TRI releases of carcinogens from 1987 to 1996, covariates associated with lung cancer, and the EPA's Risk-Screening Environmental Indicators (RSEI) model.
The impact of the volume of carcinogenic TRI releases on each county's lung cancer mortality rates was calculated using six spatial interaction functions (containment, buffer, power decay, exponential decay, quadratic decay, and RSEI estimates) and evaluated with four multivariate regression methods (linear, generalized linear, spatial lag, and spatial error). Akaike Information Criterion values and P values of spatial interaction terms were computed. The impacts calculated from the interaction models were also mapped. Buffer and quadratic interaction functions had the lowest AIC values (22298 and 22525 respectively), although the gains from including the spatial interaction terms were diminished with spatial error and spatial lag regression.
The use of different methods for estimating the spatial risk posed by pollution from TRI sites can give different results about the impact of those sites on health outcomes. The most reliable estimates did not always come from the most complex methods.
Environmental pollution data such as that collected by the United States Environmental Protection Agency's Toxic Release Inventory (TRI) have been used extensively for studies in environmental justice and medical geography . These studies involved estimating an individual's or a community's exposure to pollution using the spatial information contained in the TRI database. Despite the use of this spatial information, the geographical theory used to guide the estimation of location-based exposure to pollution has frequently been limited to basic containment and buffer analysis, especially at the national scale. The aim of this research is to improve the spatial analysis of TRI data by incorporating distance decay effects derived from spatial interaction modeling in order to provide a more realistic approach to the estimation of location-based exposure to pollution, particularly airborne pollution. This is achieved by using several different functions for calculating this exposure and comparing the results when they are used in multivariate regression analyses with lung cancer mortality rates.
The different methods for estimating the risk at a location are evaluated because, while many studies have explored and demonstrated a link between environmental pollution and a variety of adverse societal and medical effects , understanding the nature of this relationship is equally important. As the variety of methods used to estimate these impacts attests, the nature of this relationship is not as well understood as the existence of the relationship. The form of this relationship greatly impacts the answers to questions that may arise from the discovery of a relationship, such as the extent to which rural counties experience adverse impacts from urban polluters. A visual cartographic comparison of some approaches has been explored by McMaster et al , although they do not make the statistical comparison carried out here.
Spatial analyses of toxic pollution data, whether for environmental justice or for medical geography, have typically used a simple spatial estimate of exposure. The exposure has been recorded as a binary variable (exposed or not exposed) either through spatial containment such that a person is exposed if they live in the same census tract or county as an industrial site [1, 3–8], or a spatial buffer such that a person is exposed if they live within a threshold distance (e.g., 1 mile) of an industrial site [5, 9–12]. Variations on the latter use multiple buffers to approximate decreasing risk with increased distance, or select a small number of neighborhoods at increasing distances which can be treated as samples from multiple buffers. This enables the study to reflect decreased exposure as the distance increases [1, 9, 13–15]. To provide a better measure of the impact of sites on a census tract, four studies [16–19] use a raster grid that can account for whether a site is in the center of the tract, or near an edge, and whether any sites are just over the border in neighboring tracts. These raster grids reflect the density of TRI sites around each raster cell, although the density is calculated using a small buffer, such as the density of sites within a one mile radius of the cell. A gradual decay of impact as distance increases is still lacking.
To address these simplifying assumptions, Dent et al  have proposed using a GIS to combine atmospheric modeling with the release data and health outcomes and provide a detailed analysis of the potential effects and risks associated with TRI releases. Morello-Frosch et al [23, 24] and Fisher et al  similarly incorporate atmospheric modeling in their analysis. These models are typically used for local, rather than national-scale analysis. The United States Environmental Protection Agency's Risk-Screening Environmental Indicators (RSEI) Model uses principles of atmospheric modeling to derive a level of risk across the entire United States . This has been used by Abel  and Downey and Hawkins , and is used in this resesarch.
Prior work summary
Containment & multiple buffers
Containment & buffer
Containment & plume modeling ("census tract containing the site and its plume", p. 148)
Multiple buffers, RSEI
Containment & buffer (both distance boundary and areal apportionment)
Neighborhoods of increasing distance
Multiple distance-based rasters & distance to nearest TRI facility
Multiple distance-based rasters
Spatial interaction modeling
In this research, I use a spatial interaction modeling approach that is more flexible than the binary approaches commonly used in spatial analysis of TRI data, yet is fast enough and generic enough to apply to the thousands or millions of release sites involved in a national scale study. Spatial interaction modeling was developed in economic geography to estimate the level of economic interaction between two towns [29–32]. The underlying assumptions are analogous to the physics theory of gravity. Just as two objects in space exert a stronger gravitational pull on each other as they increase in size and move closer to each other, two towns are expected to have a stronger level of economic interaction as the towns increase in size and as the distance between them decreases. These broad trends are applicable in many fields within geography, even though the specific functional form from physics (3), may not be as useful as other functional forms.
These models are used to estimate the effect of each TRI site on each county. As the toxic release volume increases, the impact of that site on the county increases. Likewise, the impact of nearby sites is assumed to be greater than the impact of more distant sites, as from Tobler's First Law of Geography .
The models in economic geography, such as those in Sen and Smith , give equations with a third term which in this work corresponds to the population of the county and a related positive constant β, such that the power model becomes and the exponential model becomes k ij = t i α p β i exp(- θ d ij ). Because I use age-adjusted rates of lung cancer rather than unadjusted counts for the dependent variable, these population terms are set to 1 and effectively removed from the equations.
Four sets of data are used in this paper. The first are lung cancer age-adjusted mortality rates from the National Cancer Institute's Surveillance Epidemiology and End Results (SEER) database . These rates are from the years 1990 through 2006 inclusive. The second are TRI releases from 1987 to 1996. The years chosen for the TRI databases reflect a lag time between chronic exposure to toxic chemicals and the development of lung cancer. All data are temporally aggregated to the entire time series, rather than evaluating year-by-year temporal lags. The third are risk estimates computed by the EPA's RSEI program to be used as a basis for comparison against the spatial interaction estimates The RSEI data are the risk-related results calculated from airborne releases of chemicals that are flagged as carcinogenic and have a non-zero Inhalation Unit Risk.
The final dataset, the covariates, come from multiple sources. One source is the United States Census Area Resource File . Thun et al  show variable risks for age, sex, and racial categories, so census data for the proportion of the population which is male and the proportion of the population which is non-white are included. Hendryx et al  note that lung cancer mortality is impacted by socioeconomic factors and access to health care, so additional covariates include the percent of the population with a less than high school education, the percent of the population with a college education, the percent of families below the poverty level, the unemployment rate, and the number of physicians per 1000 residents. Because smoking is the most significant risk factor for lung cancer , I also include the smoking rate of the county based on the BRFSS survey data from 2003 to 2006. Different regions of the United States have different rates of lung cancer , so the covariate data also includes spatial indicator variables recording whether a county is in the American south, northeast, Midwest, or western region, and whether a county is part of Appalachia, a regional designation from the Appalachian Regional commission which overlaps parts of the northeast, Midwest and southern regions. Due to a lack of data, information regarding personal movements is not included, although analysis comparing place of birth with place of death may partially account for this .
The Modifiable Areal Unit Problem  introduces difficulties into the interpretation at the county scale, especially in the larger counties in the western United States where the county centroid may be tens of miles away from the county's population center. Additionally, in these larger counties, the risk may vary within the county, and this variance is masked by calculating the risk at the county scale. However, some of the covariate data (e.g., the BRFSS-derived smoking rate) is not available at a finer scale, necessitating a county-level analysis. In the research presented here, the impacts of the Modifiable Areal Unit Problem and large county sizes are expected to have a similar impact across all models because all tests use the same spatial scale.
An examination using a synthetic dataset was considered, but the results of such a test would minimize the AIC in the situation reflecting the way the dataset is constructed (e.g., the impact falls off according to an exponential distance decay function), which may or may not correspond to a real world situation. Therefore, actual, rather than synthetic, data are used in this research.
The calculated impacts and covariates are then used in multivariate regression models calculated with the R software package : ordinary least squares (OLS) linear regression, general linear model regression (GLM), spatial lag regression, and spatial error regression. The latter two incorporate spatial dependence in the regression model and are detailed below. There is spatial autocorrelation in the response variable (Moran's I = 0.69, p < 0.01), demonstrating the existence of spatial dependence and suggesting the applicability of the spatial regression techniques. Geographically Weighted Regression , which can vary the regression coefficients across the study area was not applied because it is unlikely that the nature of the relationship between toxic releases and mortality changes across the country.
This situation is not strictly one of evaluating a single function against a null hypothesis of zero impact from toxic emissions, but is rather evaluating many different functions against both each other and the null hypothesis. This makes the task more akin to model parameter optimization than traditional statistical hypothesis testing. It is considered here that, of the different functions and their parameterizations, the most appropriate representation is the one that minimizes the Akaike Information Criterion (AIC) of a regression test in which the modeled risk is one of the independent variables. In all the regression tests used in this comparison, the remainder of the independent variables are the demographic, behavioral and regional covariates, and the dependent variable is lung cancer mortality.
Experiments testing many parameterizations of the buffer, Cutter, power, and exponential functions were used to guide the results given here. These parameterizations are for the contiguous United States, and were evaluated on which parameterizations gave the lowest AIC values when combined with the covariate data in an OLS regression using the lung cancer mortality rate as the dependent variable. The same tests are conducted for the containment and RSEI approaches to risk estimation for comparison. AIC values were also computed for generalized linear model regressions, although none of the generalized linear model regressions produced lower AIC values than linear regression. As a result, the generalized linear model regressions are not further discussed. Preliminary experiments (not presented) demonstrated that the parameterizations that perform well for OLS regression also typically perform well for the spatial lag and spatial error regressions. The tests presented here evaluate α and θ values of 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, and 5.0, and distance thresholds of 5, 10, 15,... 500 miles (8, 16, 24,... 805 km) for the buffer and Cutter functions. After the parameterizations were found, spatial lag regression and spatial error regression models were computed. Also, linear regression models for each of the rural-urban continuum codes from the United States Department of Agriculture. Lastly, maps of proportion of the TRI impact for each county that originated in release sites located in urban areas were produced. This allows an examination of whether the impact of sites in urban areas is limited to those cities, or whether it extends far into the surrounding rural areas, and demonstrates that some environmental justice questions are not robust to the choice of risk model.
Spatial Regression Methods
500 miles (804 km)
500 miles (804 km)
Akaike Info. Criterion
probability of TRI term
Akaike Info. Criterion
probability of TRI term
Akaike Info. Criterion
probability of TRI term
OLS Regression results
< 2E -16
% no high sch.
% in poverty
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
< 2E -16
OLS regression R-squared values by rural-urban code
Definition of each rural-urban code
Counties in metropolitan areas
Central counties of metropolitan areas of 1 million population or more.
Fringe counties of metropolitan areas of 1 million population or more.
Counties in metropolitan areas of 250,000 to 1 million population.
Counties in metropolitan areas of fewer than 250,000 population.
Counties not in metropolitan areas
Urban population of 20,000 or more, adjacent to a metropolitan area.
Urban population of 20,000 or more, not adjacent to a metropolitan area.
Urban population of 2,500 to 19,999, adjacent to a metropolitan area.
Urban population of 2,500 to 19,999, not adjacent to a metropolitan area.
Completely rural or less than 2,500 urban population, adjacent to a metropolitan area.
Completely rural or less than 2,500 urban population, not adjacent to a metropolitan area.
The buffer and Cutter functions outperform the containment, power and exponential functions (Table 3). This improvement is notable both for the R-squared values of the OLS regressions and the Akaike Information Criterion (AIC) for all regressions. Because the AIC penalizes regression models with more parameters, lower values are preferred. These functions also outperformed the RSEI risk-related results values, which was not the expected outcome. Other RSEI products, the hazard, modeled hazard, and modeled hazard*population product were also tested, but all performed worse than the buffer and Cutter functions. However, the power and exponential functions from the spatial interaction literature did not perform much better than leaving out the toxicity term, and occasionally even increased the AIC value, which may result from the coarse resolution of the county-level dataset. These two functions may yet be useful at a finer scale.
The improvement of the buffer and Cutter functions over the RSEI data demonstrates that despite the difficulties posed by the Modifiable Areal Unit Problem and the size of large western counties obscuring variation of risk within the county, these spatial interaction approaches may still be an accurate reflection of the risks posed by TRI facilities. It should be cautioned that while this work demonstrates a relationship between TRI facilities and lung cancer, it does not yet indicate a causal link, nor does it indicate that the best-fitting risk estimation method, a large buffer around the TRI site, has the strongest causal relationship with lung cancer mortality.
Additionally, AIC values are better for the spatial regression techniques compared to the OLS regression values. However, including the TRI term in the spatial regression techniques does not lead to as much improvement over the base case of no interaction term. Even so, the improvement in the buffer model gives the spatial error regression of the buffer model the lowest AIC value. Both the Moran's I spatial autocorrelation statistic given above and the lower AIC values for the spatial regression techniques indicate that there is spatial dependence in lung cancer mortality. Moreover, this spatial dependence is not accounted for by the independent variables. This dependence is most likely the result of one or more additional spatial processes affecting lung cancer that are not accounted for in these data, rather than a simple diffusion or contagion process of lung cancer itself. The limited improvement from adding the TRI impacts strengthens the suggestion that there remain geographic processes affecting lung cancer that are not accounted for in these datasets. While it is not executed in this study, GWR may also reveal further evidence of confounding processes by revealing interactions with modeled covariates via non-stationary regression coefficients.
As with the different regression methods, the buffer and Cutter functions have the best R-squared values across the entire range of rural-urban continuum codes (Table 5). Also, category 5, defined as counties containing a larger town (more than 20,000 residents) but which are not adjacent to a metropolitan area, has much higher R-squared values than the other rural-urban codes across all models. It is not yet clear why this would be the case.
These results suggest that changing the method used to estimate risk will change the representation of the spatial impacts of the TRI sites on public health. As others have noted, the scope and scale of analysis can substantially impact the results , so researchers should be cautious when generalizing these findings at a county scale and national scope to more local scales and scopes. Nonetheless, researchers using the TRI dataset to estimate the health risks from pollution should carefully consider the method used to estimate the risk, as the most sophisticated model used here, the RSEI data, did not provide the lowest AIC values.
The maps in figure 3 display the percent of the TRI impact on each county from sources in urban areas calculated using the functions that performed best in the earlier results. As estimated by these models, the potential effects of pollution from urban TRI releases extend far beyond the limits of the urban areas. However, the extent varies depending on which function is used and how it is parameterized, highlighting the importance of using an appropriate function. In the power and exponential maps (figure 3a and 3b), the impacts from urban release sites are more limited to urban areas and the nearby rural communities. In both the buffer and Cutter maps (figure 3c and 3d), rural areas in the northeastern and southwestern United States, have between 75 and 100% of their estimated TRI impact from release sites in urban areas. These extended effects of urban areas are related to the large radii used in the distance decay functions. Additional work is needed to examine the environmental toxicology to determine whether the chemicals being released could travel such large distances or whether these models are simply capturing spatial dependence of the outcome that is induced by a confounding spatial process.
Future work will investigate the parameterization choices of the functions. This ad hoc approach to parameterization-examining different possibilities of the α, θ and threshold parameters-is not the ideal approach. Statistical approaches to finding the optimal α and θ parameters can be incorporated to improve the spatial interaction models that are generated [29, 30, 32]. A geostatistical approach can be applied to determine the decay function form and parameters. A correlogram plot comparing the distance between two counties and the difference between their mortality rates or their residuals from a regression function could be used to parameterize the function. Additionally, subsets of the correlogram could be examined separately to investigate anisotropy and non-stationarity. However, with both the ad hoc approach in this paper and a statistical model-fitting approach, using the data to optimize the parameters and then using those parameters to analyze the same data introduces circularity into the model-fitting process that would best be avoided.
A more theoretically sound approach would be to vary the α and θ parameters based on the properties of the toxic chemicals that are released. Varying α is similar to methods used somewhat frequently to account for the different toxicity of the chemicals released [2, 8, 11, 20–22], although the studies cited here use multiplicative rather than exponential modifiers (α t i instead of t i α ). In each case, higher values of α correspond to more toxic chemicals. Different studies have made this adjustment using different references, including American Conference of Governmental Industrial Hygienists Threshold Limit Values [3, 22], a chronic toxicity index , an inhalation unit risk , a lifetime cancer risk , and the RSEI model [9, 27]. Similarly, θ and T can be varied to reflect differences in airborne transport of the chemicals. If a chemical travels more easily and farther, lower values of θ and higher values of T can be used. These parameters can also be varied based on the direction from the release site to the affected community, thus incorporating anisotropy.
Ongoing work includes the refinement of at-risk population estimates using the LandScan USA population dataset  which can explore variation missed by county-level populations unable to capture fine-scale risks. For example, if a chemical is only present in the atmosphere within a mile of the release site, any county-by-county analysis will be problematic because the spatial resolution of county-level data are coarser than a square mile. The LandScan dataset provides population estimates at a 3 arc-second resolution (roughly 90 meters). This can then provide improved estimates of the number of people within one mile of the release site instead of assigning the impact of a release site on the county as if everybody lived at the centroid of the county. This approach will have stronger effects on the power and exponential models because they have more rapid decreases in the impact as one travels farther from the release site (figure 2). This ongoing work also incorporates the adjustments given above varying the parameters to account for properties of the chemicals released and local climatic conditions to account for prevailing wind directions.
The research in this paper demonstrates that the use of simple containment techniques for estimating the spatial risk posed by pollution from TRI sites as well as the RSEI risk-related results can give misleading results about the impact of those sites on health outcomes. This is done through a comparison of multivariate regression results using inputs of six different functions for estimating the impact of a release site on a county: containment, buffering, the quadratic distance decay function proposed by Cutter et al , an inverse power distance decay function, an exponential distance decay function, and the RSEI risk-related results. The buffer and Cutter approaches consistently performed the best among these methods. The effects of this function choice are also demonstrated through mapping the percent of the overall impact that comes from urban TRI sites for all models except containment. As refinements to the parameterization process are made, the utility of more theoretically sound spatial interaction models will improve further.
Support for this report was provided by the Office of Rural Health Policy, Health Resources and Services Administration, PHS Grant No. 1 U1CRH10664-01-00. The author also acknowledges Dr. Michael Hendryx for support and comments on an earlier draft of this manuscript. Lastly, the author thanks two anonymous referees for their valuable comments and suggestions.
- Pastor M Jr, Sadd JL, Morello-Frosch R: Waiting to Inhale: The Demographics of Toxic Air Release Facilities in 21 st -Century California. Soc Sci Quart 2004, 85: 420–440.View Article
- McMaster RB, Leitner H, Sheppard E: GIS-based Environmental Equity and Risk Assessment: Methodological Problems and Prospects. Cart and Geog Info Sys 1997, 24: 172–189.View Article
- Bowen WM, Salling MJ, Haynes KE, Cyran EJ: Toward Environmental Justice: Spatial Equity in Ohio and Cleveland. Annals of the Assn of Am Geog 1995, 85: 641–663.View Article
- Cohen MJ: The spatial distribution of toxic chemical emissions: Implications for nonmetropolitan areas. Soc & Nat Res 1997, 10: 17–39.View Article
- Croen LA, Shaw GM, Sanbonmatsu L, Selvin S, Buffler PA: Maternal Residential Proximity to Hazardous Waste Sites and Risk for Selected Congenital Malformations. Epidemiology 1997, 8: 347–354.PubMedView Article
- Cutter SL, Solecki WD: Setting environmental justice in space and place: Acute and chronic airborne toxic releases in the southeastern United States. Urb Geog 1996, 17: 380–399.View Article
- Daniels G, Friedman S: Spatial Inequality and the Distribution of Industrial Toxic Releases: Evidence from the 1990 TRI. Soc Sci Quart 1999, 80: 244–262.
- Shaw GM, Schulman J, Frisch JD, Cummins SK, Harris JA: Congenital Malformations and Birthweight in Areas with Potential Environmental Contamination. Arch of Env Health 1992, 47: 147–154.View Article
- Abel TD: Skewed Riskscapes and Environmental Injustice: A Case Study of Metropolitan St. Louis. Env Mgmt 2008, 42: 232–248.View Article
- Gragg RD III, Christaldi RA, Leong S, Cooper M: The location and community demographics of targeted environmental hazardous sites in Florida. J Land Use & Env Law 1996, 12: 1–44.
- Kearney G, Kiros GE: A spatial evaluation of socio demographics surrounding National Priorities List sites in Florida using a distance-based approach. Int J Health Geog 2009, 8: 33.View Article
- Neumann CM, Forman DL, Rothlein JE: Hazard Screening of Chemical Releases and Environmental Equity Analysis of Populations Proximate to Toxic Release Inventory Facilities in Oregon. Env Health Persp 1998, 106: 217–226.View Article
- Berry M, Bove F: Birth Weight Reduction Associated with Residence near a Hazardous Waste Landfill. Env Health Persp 1997, 105: 856–861.View Article
- Knox EG, Gilman EA: Hazard proximities of childhood cancers in Great Britain from 1953–80. J Epid Comm Health 1997, 51: 151–159.View Article
- Nordström S, Beckman L, Nordenson I: Occupational and environmental risks in and around a smelter in northern Sweden: I. Variations in birth weight. Hereditas 1978, 88: 43–46.PubMedView Article
- Downey L: Spatial Measurement, Geography, and Urban Racial Inequality. Social Forces 2003, 81: 937–952.View Article
- Mennis J: Using Geographic Information Systems to Create and Analyze Statistical Surfaces of Population and Risk for Environmental Justice Analysis. Soc Sci Quart 2002, 83: 281–297.View Article
- Mohai P, Saha R: Reassessing racial and socioeconomic disparities in environmental justice research. Demography 2006, 43: 383–399.PubMedView Article
- Mennis JL, Jordan L: The Distribution of Environmental Equity: Exploring Spatial Nonstationarity in Multivariate Models of Air Toxic Releases. Annals of the Assn of Am Geog 2005, 95: 249–268.View Article
- Cutter SL, Hodgson ME, Dow K: Subsidized Inequities: The Spatial Patterning of Environmental Risks and Federally Assisted Housing. Urb Geog 2001, 22: 29–53.View Article
- Fisher JB, Kelly M, Romm J: Scales of environmental justice: Combining GIS and spatial analysis for air toxics in West Oakland, California. Health & Place 2006, 12: 701–714.View Article
- Horvath A, Hendrickson CT, Lave LB, McMichael FC, Wu TS: Toxic Emissions Indices for Green Design and Inventory. Env Sci & Tech 1996, 29: 86A-90A.View Article
- Morello-Frosch R, Pastor M Jr, Sadd J: Environmental justice and Southern California's 'riskscape'. Urb Affairs Rev 2001, 36: 551–578.View Article
- Morello-Frosch R, Pastor M Jr, Porras C, Sadd J: Environmental Justice and Regional Inequality in Southern California: Implications for Future Research. Env Health Persp 2002, 110: 149–154.
- Pastor M Jr, Morello-Frosch R, Sadd JL: Breathless: Schools, Air Toxics, and Environmental Justice in California. The Policy Studies J 2006, 34: 337–362.View Article
- Dent AL, Fowler DA, Kaplan BM, Zarus GM, Henriques WD: Using GIS to Study the Health Impact of Air Emissions. Drug and Chem Toxic 2000, 23: 161–178.View Article
- Downey L, Hawkins B: Single-Mother Families and Air Pollution: A National Study. Soc Sci Quart 2008, 89: 523–536.View Article
- United States Environmental Protection Agency: Risk-Screening Environmental Indicators (RSEI) Model. [http://www.epa.gov/opptintr/rsei/]
- Batty M, Mackie S: The calibration of gravity, entropy, and related models of spatial interaction. Env and Planning 1972, 4: 205–233.View Article
- Baxter M: Estimation and inference in spatial interaction models. Prog Human Geog 1983, 7: 40–59.
- Sen A, Smith TE: Gravity Models of Spatial Interaction Behavior. New York, NY: Springer; 1995.
- Sheppard ES: The distance-decay gravity model debate. In Spatial Statistics and Models. Edited by: Gaile GL, Willmott CJ. Boston, MA: D. Reidel Publishing Company; 1984:367–388.
- Tobler WR: A computer movie simulating urban growth in the Detroit region. Econ Geog 1970, 46: 234–240.View Article
- SEER. Surveillance, Epidemiology, and End Results (SEER) Program: SEER*Stat Database: Incidence - SEER 13 Regs Limited-Use. Nov 2009 Sub (1992–2007) <Katrina/Rita Population Adjustment>. [http://www.seer.cancer.gov] National Cancer Institute, Cancer Statistics Branch, released April 2010, based on the November 2009 submission;
- ARF: Area Resource File. Rockville, MD: 2006 U.S. Department of Health and Human Services, Health Resources and Services Administration, Bureau of Health Professions; 2005.
- Thun MJ, Henley SJ, Burns D, Jemal A, Shanks TG, Calle EE: Lung Cancer Death Rates in Lifelong Nonsmokers. J of Natl Cancer Inst 2006, 98: 691–699.View Article
- Blot WJ, Fraumeni JF Jr: Geographic Patterns of Lung Cancer: Industrial Correlations. Am J Epidem 1976, 103: 539–550.
- Hendryx M, O'Donnell K, Horn K: Lung Cancer Mortality Is Elevated in Coal-Mining Areas of Appalachia. Lung Cancer 2008, 62: 1–7.PubMedView Article
- Sabel CE, Boyle PJ, Löytönen M, Gatrell AC, Jokelainen M, Flowerdew R, Maasilta P: Spatial Clustering of Amyotrophic Lateral Sclerosis in Finland at Place of Birth and Place of Death. Am J Epidem 2003, 157: 898–905.View Article
- O'Sullivan D, Unwin DJ: Geographic Information Analysis. 2nd edition. Hoboken, NJ, USA: John Wiley & Sons, Inc; 2010.View Article
- Luo J, Hendryx M, Ducataman A: Association between Six Environmental Chemicals and Lung Cancer Incidence in the United States. J Rural Health in review
- The R Project: The R Project for Statistical Computing. [http://www.r-project.org/]
- Ward MD, Gleditsch KS: Spatial Regression Models. Los Angeles, CA, USA: Sage Publications; 2008.
- Robert Haining: Spatial Data Analysis: Theory and Practice. Cambridge, UK: Cambridge University Press; 2003.
- Smirnov O, Anselin L: Fast maximum likelihood estimation of very large spatial autoregressive models: a characteristic polynomial approach. Comput Stat & Data Anal 2001, 35: 301–319.View Article
- Rogerson PA: Statistical Methods for Geography: A Student's Guide. 2nd edition. Los Angeles, CA, USA: Sage Publications; 2006.
- Austin MP, Belbin L, Meyers JA, Doherty MD, Luoto M: Evaluation of statistical models used for predicting plant species distributions: Role of artificial data and theory. Ecol Model 2006, 199: 197–216.View Article
- Baden BM, Noonan DS, Turaga RMR: Scales of Justice: Is there a Geographic Bias in Environmental Equity Analysis? J Env Plan Mgmt 2007, 50: 163–185.View Article
- Bhaduri B, Bright E, Coleman P, Urban M: LandScan USA: A High Resolution Geospatial and Temporal Modeling Approach for Population Distribution and Dynamics. GeoJournal 2007, 69: 103–117.View Article
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.