Conference on Geospatial Approaches to Cancer Control and Population Sciences

Time: 1:10 - 1:25 p.m.
Co-Author: Xiuling Zhang, PhD, New York State Cancer Registry

Innovations in cancer detection and treatment are introduced in an uneven fashion, typically originating in research and teaching hospitals and eventually diffusing to smaller community hospitals. One such major innovation was the replacement of film mammography with digital mammography in the early 2000s. We quantify the diffusion of digital mammography within the state of New York from 2004, when approximately 10% of all mammogram images were digital, to 2012, when the figure was over 90%, using Medicare claims data from a sample of over 100,000 cancer-free women for whom the zip code of residence was known. The percentage of mammograms that were digital was calculated for each zip code for each 12-month trailing period in the study. The data were smoothed using a flexible spatial filter to capture a minimum of at least 100 mammograms centered on each zip code. Viewing the resulting smoothed percentages as a series of sequential maps reveals the times, locations, and rates at which digital mammography was adopted. Early adopters tended to be in areas with younger, upscale populations and communities with teaching hospitals, but there were individual exceptions; for example, the small Binghamton market was years ahead of other upstate areas. This type of analysis can help us better understand the persistence of disparities in cancer, as differential access to technology can translate into differences in incidence and survival. The process is now being repeated with the introduction of 3-D digital tomography. Health care providers are likely interested in seeing where they fall on the innovation spectrum; maps like these could help accelerate future innovation.

Time: 1:25 - 1:40 p.m.
Co-Authors: Tzy-Mey Kuo, PhD, MPH, University of North Carolina at Chapel Hill; Angela Smith, MD, MS, University of North Carolina at Chapel Hill; Matthew Nielsen, MD, MS, University of North Carolina at Chapel Hill

The epidemiology of prostate cancer (CaP) in the US is influenced by patient demographics, community, and provider characteristics which vary geographically. We explore this variation using incidence rates (IR) to identify areas of highest risk and deconstruct the relative influence of community and provider characteristics in the state of North Carolina (NC). Methods: We examine age-adjusted IR data from the NC cancer registry and county-level data (number of primary care & urology physicians per capita, race, education/insurance/employment) from the Area Resource File to describe disease risk in NC. Spatial clustering analysis methods were applied to evaluate spatial autocorrelation and identify clusters with significant high (HH) and low (LL) incidence relative to the surrounding area. Analyses were used to compare defining characteristics between clusters. Results: The geographic distribution of CaP Incidence in NC is not random (p < .001). Local Indicator of Spatial Association (LISA) analysis identified 12 HH clusters and 11 LL clusters. The HH clusters had a significantly larger proportion of black residents, fewer urologists and primary care physicians per capita, and lower SES than LL clusters. Conclusion and Impact: This study suggests that county-level socioeconomic and racial variables, as well as access to primary and specialty care, are associated with geographic disparities in prostate cancer. These observations support additional work with spatial regression to further characterize factors affecting disparities in incidence and outcomes to better identify targets for intervention.

Time: 1:40 - 1:55 p.m.
Co-Authors: James Hebert, ScD, University of South Carolina; Jan Eberth, PhD, University of South Carolina

Regional disparities exist in rates of lung and bronchus cancer within the United States. Underlying reasons may include access to care, screening utilization, quality of treatment, socioeconomic, or cultural (including lifestyle) characteristics, or some combination of these. The mortality-to-incidence ratio (MIR) provides a relative measure of cancer survival and a means for identifying regional disparities that is easily derived from traditional cancer incidence and mortality data. Mortality and incidence rates for lung and bronchus cancer for 49 states (i.e., excluding Nevada) and the District of Columbia (D.C.) were obtained from the National Cancer Institute State Cancer Profiles. Rates were given as 5-year averages for the years 2008-2012. MIRs were calculated by dividing the age-adjusted mortality rate by the age-adjusted incidence rate per 100,000 for a given state. States were ranked by ascending MIR and divided into deciles. These categories were mapped using ArcGIS software in order to highlight the geographic variation in lung and bronchus cancer survival. A mean MIR of 0.75 was observed for the 49 states and D.C. MIRs ranged from 0.65 to 0.83. The five states with the lowest relative survival from lung and bronchus cancer, as measured by the highest MIRs, were all in the South: Arkansas, Oklahoma, Tennessee, Alabama, and Louisiana. The states with the lowest MIR, indicating the highest relative survival, were mostly in the Northeast: Connecticut, New York, Massachusetts, New Jersey, and Hawaii. The calculation and mapping of state level MIRs for lung and bronchus cancer revealed that many southern states have the lowest relative lung and bronchus cancer survival. Future research should focus on identifying the underlying risk factors for poor lung and bronchus cancer outcomes. This may entail exploring variation at smaller scales (e.g., counties).

Time: 1:55 - 2:10 p.m.
Co-Authors: Su Yeon Han, MS, San Diego State University; Atsushi Nara, PhD, San Diego State University; Joseph Gibbons, PhD, San Diego State University; Caroline A. Thompson, PhD, San Diego State University

We developed an interactive web mapping tool for visualizing San Diego County cancer rates with socioeconomic data at the sub-regional area (SRA) level. This tool will be extended to provide dynamic choropleth mapping and spatial analysis functions for areas beyond San Diego using California Cancer Registry data. This dynamic web GIS/mapping tool was created with an open-source JavaScript library, Leaflet, and free web authoring tools (bootstrap, jquery, and Google Chart) to provide user-friendly maps and data mining functions, including multiple classification methods, correlation analysis, data export, and side-by-side map displays. Data obtained from San Diego County Health & Human Services Agency includes age-adjusted rates of death and hospitalization for 19 different cancer sites. Contextual data were compiled from the 2009-2013 wave of the American Community Survey, including variables on race/ethnicity and other demographic factors. The web tool is currently available in beta testing and undergoing a usability review by public health staff and cancer researchers. Conclusion: Different from existing web mapping tools for cancer data, our tool can provide a side-by-side visual comparison between cancer data and socioeconomic data to facilitate hypothesis testing for cancer population health research. Version-2 will add advanced GIS analysis functionality including multivariate analysis capabilities using Geographically Weighted Regression (GWR) models, Principle Component Analysis for covariate control, and correction for spatial autocorrelation. This free web map tool can be re-created by other researchers easily to help cancer researchers and policymakers visualize and analyze the spatial correlation between cancer outcomes and socioeconomic variables.

Time: 1:00 - 1:30 p.m.

This presentation will (1) explore the use of wearable environmental sensors to reduce cancers in vulnerable populations such as infants and children; and (2) describe a research collaboration with NAACCR that seeks to develop novel methods (geospatial cryptography) that will enhance sharing and analysis of geospatial cancer data among disease registries and researchers.

Time: 1:30 - 1:45 p.m.
Co-Authors: Shalini Pahrek, MPH, Maryland Department of Health and Mental Hygiene, Center for Cancer Prevention and Control; Jennifer Hayes, MEd, MPH, Maryland Department of Health and Mental Hygiene, Center for Cancer Prevention and Control; Kimberly Stern, MHA, CTR, Maryland Department of Health and Mental Hygiene, Center for Cancer Prevention and Control; Frank Curriero, PhD, Johns Hopkins University; Amber D'Souza, PhD, Johns Hopkins University

Invasive cervical cancer (ICC) is a highly preventable cancer; yet some women in the state of Maryland, like in other parts of the United Sates, continue to develop the disease. We hypothesized that there is space-time variation in ICC incidence. Methods: Data on all 2172 ICC cases reported by the Maryland Cancer Registry between 2003 and 2012 was utilized. Our analysis involved a retrospective space-time cluster detection analysis (SaTScan), searching for clusters of both high and low rates. All analyses were conducted at the county level, using 2-year aggregates of time and adjustments were made for age distribution and race using census data as well as percentage of females per county receiving Pap test using Behavioral Risk Factor Surveillance Data. Results: Median age of cases was 50 years (IQR: 40-63.5) and at diagnosis, cases had mainly grade 2 or 3 cancers (83%) with no metastasis (88%). After adjustments, four significant clusters were identified. A cluster of higher than expected rates comprising 3 counties (Baltimore city, Anne Arundel and Howard); RR 1.52, p=0.011 was observed for the period 2011-2012. In addition three significant clusters of lower than expected rates were observed for the period 2003-2004 in Cecil county (RR 0.21, p=0.011), 2005-2008 in Somerset county (RR 0.063, p <0.001) and 2011-2012 in Fredrick county (RR 0.25, p<0.001). Conclusion and Impact: Overtime some counties have experienced significantly lower than expected rates of disease. However there is a more recent cluster of disease that is not explained by age distribution, racial composition or the percentage of females screened for cervical cancer and this cluster may require targeted efforts to decrease disease rates.

Time: 1:45 - 2:00 p.m.
Co-Authors: Emre Eftelioglu, MCS, University of Minnesota; Xun Tang, PhD, University of Minnesota

Finding spatial clusters, i.e., places with unusually high disease rate, is an important task for the geo-surveillance of cancer. Many current tools (e.g., NCI GeoViewer) use spatial auto-correlation map-visualizations for manual searching of spatial clusters. Since manual searches are expensive for large datasets, other tools (e.g., SaTScan) use algorithms to automatically identify statistically significant spatial clusters with simple footprints (e.g., circles) to reduce manual labor. However, SaTScan fails to detect many interesting spatial clusters. For example, towers (or chimneys) emitting carcinogenic material may lead to a ring-shaped cluster surrounding the source. In addition, related lung-cancer clusters related to vehicular emissions may lead to linear footprints along high-traffic transportation corridors (e.g., roads, rivers, etc.). We recently proposed new algorithms for detecting (statistically significant) clusters with ring-shaped footprints. The proposed algorithms outperformed SaTScan in a case study with dataset describing the legionnaire's disease outbreak in New York in 2015, where a cooling tower that was determined to the source the outbreak by public health officials. Our algorithm identified a statistically significant ring containing the location of the source. However, the source was not contained in any cluster identified by SaTScan. We have also proposed new algorithms for detecting (statistically significant) clusters with linear footprints. In case studies with pedestrian fatality datasets, our algorithms identified several statistically significant spatial clusters with linear footprints. However, SaTScan missed all but one statistically significant cluster with linear footprints. We acknowledge support from the USDOD, the NSF and the University of Minnesota.

Time: 2:00 - 2:15 p.m.
Co-Authors: Dustin T. Duncan, ScD, New York University; Jessica K. Athens, PhD, New York University; Brian Elbel, PhD, MPH, New York University

Colorectal cancer is one of the most prevalent cancers in the United States. Cluster detection methods have been applied to colorectal cancer previously but rarely done in New York City (NYC). These analyses may yield a better understanding of how cancer incidence concentrates in certain areas and insight into key environmental determinants. The objective of this study was to detect geospatial clusters of colorectal cancer incidence in NYC. We used data on gender-specific colorectal cancer incidence of five NYC boroughs at the zip-code level from the NY State Cancer Registry in 2005-2009. A spatial scan statistic based on the unadjusted discrete Poisson model was used to identify statistically significant (p <0.05) in NYC, individuals had 19-21% greater risks compared to outside the cluster. Of two low-risk clusters (p<0.05), individuals had 20-25% lower-risks, compared to outside the cluster. For women, of three high-risk clusters, individuals had 13-92% greater risks than individuals outside the cluster. In one low-risk cluster, individuals had a 17% lower-risk, relative to outside cluster. For both gender, high risk clusters appeared in Bronx and Brooklyn, whereas low risk clusters emerged in Manhattan and a few areas in Queens. Colorectal cancer incidence spatially clustered at the zip-code level in NYC. Future research should examine the geographic stability of identified colorectal cancer clusters over time.

Time: 2:15 - 3:30 p.m.
Co-Authors: Zheng Li, PhD, Penn State; Stephen A. Matthews, PhD, Penn State; Vern Chinchilli, PhD, Penn State

Prostate cancer is one of the most common cancers diagnosed among males, and is an important public health issue in Pennsylvania. The incidence rate and mortality vary substantially across geographical regions (counties) and over time (years). The widely-used Cox Proportional Hazard (PH) model does not apply due to the violation of PH assumption. In this work, we propose to use Bayesian accelerated failure time (AFT) models to analyze prostate cancer survivorship by incorporating random effects with multivariate conditional autoregressive (MCAR) priors for taking spatial temporal variation into account. The models are fitted based on Monte Carlo Markov Chain (MCMC) technique under the Bayesian framework. Extensive simulations are performed to examine and compare the performances of various Bayesian AFT models with MCAR priors. The criteria for model selection via the deviance information criterion (DIC) is also evaluated in the simulation study. Finally, we implement our method to the prostate cancer data obtained from the Pennsylvania Cancer Registry which includes all reported prostate cancer diagnosed and death cases by county from years 2000-2011.

Time: 1:00 - 1:30 p.m.

Cancer prevention and control is founded on inferential tools that should highlight important etiological and behavioral patterns in the variation of the disease. To this end it can be fundamentally important to consider a model-based approach to such disease analysis. A model-based approach is the most flexible tool in the inferential box. It allows the inclusion of multiple observed effects within the inference framework, something that testing-based paradigms do not provide. In addition, it can adjust for unobserved effects (confounding). Within a modeling framework, geospatial information can play a significant role in highlighting risk effects not observable in time based studies (cohort or longitudinal studies).

Bayesian disease mapping addresses these issues by focusing on hierarchical modeling of disease risk effects (e.g. clusters, hot spots, ecological relations) from a spatial view point while allowing sensible inference to be made. Confounding can easily be included within these models via predictor adjustment and random effects. These latter effects can be spatially correlated or uncorrelated and can be seen as latent effects underlying the risk. An example of the analysis of multiple cancer outcomes where between-disease shared effects are modelled and an optimal model selection strategy is pursued.

Time: 1:30 - 1:45 p.m.
Co-Authors: Jaime E. Hart, ScD, Brigham and Women's Hospital/Harvard University; Kimberly A. Bertrand, ScD, Boston University; Zhibin Sun, PhD, Colorado State University and U.S. Department of Agriculture UV-B Monitoring and Research Program; Rulla M. Tamimi, ScD, Harvard University; Francine Laden, ScD, Harvard University

Ultraviolet B (UV-B) radiation, the primary source of vitamin D for most humans, is associated with cancers. Current U.S. UV exposure models are limited in spectral, spatial, and/or temporal resolution. Area-to-point (ATP) residual kriging is a geostatistical method that can be used to create a spatiotemporal exposure model. A stratified ATP residual kriging approach was used to predict average July noon-time erythemal UV (UVEry) (mW/m2) from 1998-2012 by downscaling National Aeronautics and Space Administration (NASA) Total Ozone Mapping Spectrometer and Ozone Monitoring Instrument remote sensing images to a 1 km spatial resolution and incorporating ancillary data (aerosol optical depth [AOD], surface albedo, cloud cover, dew point, elevation, latitude, ozone, surface incoming shortwave flux, year, and/or interactions [e.g., year and AOD]). Modeling was performed separately within nine U.S. regions. Cross-validation was used to compare ATP models and NASA grids to UV-B Monitoring and Research Program (UVMRP) measurements (gold standard) by calculating mean absolute errors (MAEs) and root mean square errors (RMSEs). From 1998-2012, ATP models more accurately estimated UVEry at UVMRP stations on average compared to NASA grids in the Northeast (average MAE for ATP, NASA: 13.9, 17.4; average RMSE 16.5, 19.2), Mid-Atlantic (MAE 14.8, 14.9; RMSE 15.8, 16.4), and Southeast (MAE 25.3, 36.9; RMSE 30.6, 43.3). Preliminary results indicate ATP residual kriging can provide accurate fine-scale UVEry estimates. This spatially and temporally varying exposure model will be applied to epidemiologic studies to elucidate the role UVEry may play in cancer risk.

Time: 1:45 - 2:00 p.m.
Co-Author: Lisa McKenzie, PhD, MPH, Colorado School of Public Health

Preliminary studies indicate that populations living near oil and gas (O&G) wells have a higher risk of childhood leukemia and other adverse health outcomes compared to populations living without O&G wells. While these studies used GIS methods to determine inverse distance weighted well counts around homes, they have not accounted for changes in intensity. Our objective is to use GIS methods to build an industrial-activity model to estimate the intensity of O&G activities surrounding homes over relevant exposure periods. We obtained dates for O&G activities such as drilling, fracturing, and production for each O&G well in Colorado. Additionally, we acquired monthly production data and other variables relevant to air emissions. This information was used to build a GIS model that estimates the wind-adjusted intensity of O&G activities for any given month. By categorizing wells into activity phases, we can produce the high degree of spatial and temporal heterogeneity in intensities expected around these sites. Our model allows us to evolve from a spatial model to a spatiotemporal one where we can assign time-varying exposure scores due to O&G development. Changes in activities, pollutants, and wind direction can have a large impact on the environmental exposure one receives from a particular source. This is particularly important when investigating associations with health outcomes such as childhood cancers where the exposure window for developing cancer may be fairly specific and relatively short. Our space/time GIS model should allow us to advance research investigating the association between decentralized sources of air toxics, such as O&G development, and childhood cancers.

Time: 2:00 - 2:15 p.m.
Co-Authors: Van Tran, PhD, National Cancer Institute; Mark P. Little, DPhil, National Cancer Institute; Philip S. Rosenberg, PhD, National Cancer Institute

Age-period-cohort (APC) models are frequently used to study population trends in vital rates. However, tools for a comprehensive evaluation of geographic heterogeneity of model parameters are not well developed. We outline a general and scalable method for modeling rates in geographically-organized regions. We apply our approach to state-level variation in all-cause mortality trends for white non-Hispanic women under age 50; recent studies have noted an overall rise in mortality for this group in the United States. Thus, state-level estimates may provide new insights into this emerging public health issue. We allow region-specific parameters to be correlated spatially (e.g., among neighboring states), and to one another (e.g., allowing for correlations between state-specific mean rate and birth cohort trends (net drift)), via a random-effects formulation using a generalized multivariate conditionally auto-regressive model. This allows parameters to vary spatially, while retaining interpretability. Estimation is carried out using Bayesian methods in JAGS. Longitudinal age trend (LAT) (i.e., trend with age) exhibited significant geographic clustering with 95% posterior CI for spatial auto-correlation parameter of (0.19, 0.85). LAT varied between 5% per year in northern mountain states, and 9% per year in East North Central states. Birth cohort trend (net drift) was not significantly clustered and it varied between -0.7% per year (-0.9%, -0.5%) in New York, and 3.0% per year (2.6%, 3.5%) in West Virginia. National APC analyses that do not account for spatial heterogeneity in model parameters can over-generalize findings. Our model provides more nuanced estimates, aiding in performing reliable inference.

Time: 2:15 - 2:30 p.m.
Co-Authors: Erica Sercy, MSPH, University of South Carolina; Abdoulaye Diedhiou, MD, PhD, University of South Carolina; Daniel Kilpatrick, PhD, University of South Carolina; Jan Eberth, PhD, University of South Carolina

Local health behavior and outcome data are important for program planning and evaluation, resource allocation, and policy-making. Although interest in local-level data is growing, surveys powered to make national or state-level inferences are rarely adequate to support direct estimation for small areas such as counties. Model-based small area estimation techniques have shown to be a viable solution. Using data from the 2014-2015 SC Adult Tobacco Survey, we examined several tobacco-related outcomes at the county level using a spatial multilevel, post-stratification approach. Specifically, we used a 2-level, random-intercept multilevel model and a spatially intrinsic ICAR model. Stratum-specific (race, age and sex) estimates for each county were created and then averaged based on stratum-specific population estimates obtained from the U.S. Census. The estimated prevalence of current smoking among SC adults ranged from 9-36%, and was lowest in Lee and Edgefield counties, both rural counties with a high proportion of African-Americans. The percentage of adults reporting ever trying an e-cigarette ranged from 4-29%, and was highest in Colleton County, a rural and majority-white county. Among smokers, 22-88% reported being advised to quit, and 36-92% reported a desire to quit. Data from the SC Adult Tobacco Survey was used to develop county-level estimates of >20 tobacco-related outcomes using a spatial multilevel, post-stratification approach. The results showed heterogeneity in smoking behaviors across the state along with marked spatial correlation. Our approach could be adapted to other states to produce accurate small area estimates for tobacco use and other health behaviors.

Time: 2:45 - 3:15 p.m.

The Environmental Health Institute at the Geisinger Health System has been engaged in epidemiologic research using electronic health records since 2007. Geisinger has > 6 hospitals and > 40 community outpatient clinics in over 40 counties in central and northeastern Pennsylvania; its EHR began in 2001; its >450,000 primary care patients represent the general population in the region; and the 40 counties have a range of communities from urban to rural with a wide range of built and social environments. Each of these features facilitates epidemiologic research. This presentation will show how EHR data can be used to identify health outcomes and give examples of how environmental and community conditions can be measured at the individual or contextual level, assigned to individuals, and evaluated in relation to health. EHR data will be contrasted with what can be learned from claims data, which are available from the Geisinger Health Plan for approximately one-third of primary care patients. Environmental and community conditions that will be discussed include unconventional natural gas development (aka "fracking"); food environments; land use environments relevant to physical activity; community socioeconomic deprivation; animal feeding operations; building radon levels; and abandoned coal mine lands. Health conditions that will be discussed include childhood obesity, methicillin-resistant Staphylococcus aureus, type 2 diabetes, and asthma exacerbations. These were selected to show the range of health conditions that can be measured with EHR data, including the use of diagnoses, laboratory tests, medications, and procedures. EHR-based epidemiologic research can be enriched with primary data collection, including to measure health-related behaviors or to obtain samples for genetic or epigenetic studies, and examples will be presented.

Time: 3:15 - 3:30 p.m.

The major difficulty in the analysis of health outcomes is that the patterns observed reflect the influence of a complex constellation of demographic, social, economic, cultural and environmental factors that likely change through time and space, and interact with the different types and scales of places where people live. A suite of techniques is introduced for the visualization and spatial analysis of time series of health data, including 3D display in a combined time and geography space, binomial kriging for noise filtering, joinpoint regression to model temporal trends, and cluster analysis to group geographical units with similar temporal trends. These techniques are used to explore spatio-temporal disparities in the incidence of breast cancer late-stage diagnosis for counties of Michigan Lower Peninsula over the period 1985-2007. Overall the proportion of late-stage diagnosis declined significantly at a rate of 3 to 5.3% per year until 1999 when it started rising again at a significant rate of 2.38% per year. Temporal trends greatly vary among counties and such discrepancies peaked in the early 1990s when the number of mammography clinics in operation started increasing steadily. The first significant decline in proportion of late-stage diagnosis started much later on the Eastern border of the State along Lake Huron, in particular in the Thumb area where late-stage diagnosis has been more prevalent over the years and both access to screening and socio-economic status are less favourable.

Time: 3:30 - 3:45 p.m.
Co-Authors: Sandi Pruitt, PhD, University of Texas Southwestern Medical Center; Tammy Leonard, PhD, University of Texas at Dallas

Social observatories comprise longitudinal, linked multi-sector datasets of individuals within their local geospatial context. These datasets uniquely facilitate assessment of residential mobility and historic neighborhood exposures, thus presenting scalable, flexible platforms for longitudinal geospatial cancer prevention and control research. We link three administrative and clinical datasets tracking low-income families across the charitable food, housing, and medical sectors to evaluate potential for differential measurement of historic neighborhood exposures. Household members recorded in a food bank administrative database were matched to a safety-net healthcare system electronic medical record (EMR). Two years of food bank address histories and all associated EMR addresses were linked to housing parcel data using a hierarchical geocoding scheme. We assessed address congruence in space-time using descriptive statistics, and individual and housing correlates of congruence using logistic regression models. In all, 3,696 households and 23,192 addresses (6,289 food bank; 16,903 EMR) were included. Of 6,289 food bank addresses, 4,103 (65.2%) overlapped temporally with EMR addresses. Congruence in space-time was low (2,293 addresses; 36.5%). Congruence was more likely for older, married, Medicare/Medicaid- or un-insured and residentially-zoned households. Congruence was less likely for males and if the household moved. Multi-sector datasets measure markedly different historic neighborhood exposures. Congruence across component datasets in our social observatory is associated with household characteristics (e.g., uninsured) and behaviors (e.g., mobility), and independent of social safety-net system engagement. Social observatory datasets facilitate assessment of geospatial influences on cancer prevention and control behaviors. Future research is needed to characterize geospatial measurement error.

Time: 3:45 - 4:00 p.m.

Colorectal cancer (CRC) mortality is higher among persons diagnosed with late-stage CRC; therefore it is likely that areas with high CRC mortality rates will also have high rates of late-stage CRC. We tested whether the proportion of late-stage CRC cases among all CRC cases was a better predictor of CRC mortality than was the rate of late-state cases among the population at risk for CRC. We matched individual-level cancer incidence data from the Iowa Cancer Registry to mortality data from the Iowa Department of Health, for Iowans aged 50 years and older. In this population from 1999 – 2010 there were 21,756 persons diagnosed with CRC, of whom 14,435 were late-stage, and 7,950 deaths from CRC. Incidence, mortality and population data for ZIP Code tabulation areas were aggregated to 3,423 small areas, which each contained at least 30 expected deaths. Indirectly age-sex adjusted measures, reported as standardized ratios, were calculated for each area. Adjusted R-squares from ordinary least squares regression quantified the correlation between these measures. In Iowa, between 1999 and 2010, the late-stage incidence rates per 100,000 person-years ranged from 114.9 to 243.6, the late-stage proportions per 100 diagnosed cases ranged from 52.9 to 77.1, and the mortality rates per 100,000 person-years ranged from 41.6 to 96.5. The correlation between late-stage incidence rates and mortality rates (r2 = 0.411, p < 0.001) was nearly 400 times higher than the correlation between the proportion of cases that were late and mortality rates (r2 = 0.003, p < 0.001). The late-stage incidence rate is a better predictor of CRC mortality in Iowa than the proportion of late-stage cases. Impact: State CRC prevention programs should consider using the late-stage incidence rate as a surveillance measure and for program planning. Disclaimer: The findings and conclusions in the report are those of the author and do not necessarily represent the official position of the Centers for Disease Control and Prevention.

Time: 4:00 - 4:15 p.m.
Co-Author: Eugene Lengerich, VMD, MS, Penn State University

While many cancer studies have examined how distance to the nearest facility may affect cancer treatment decisions and outcomes, little research has been undertaken on what factors affect the actual distance patients travel and what this means in terms of the types of patients different cancer facilities see. We use 2010-2014 Pennsylvania Cancer Registry data and a Stata user-provided open-source routing machine program to first calculate patients' driving time to the hospital where they receive treatment as well as the driving time to the nearest hospital. We then examine patient (age, sex, race, and insurance type, disease (primary site, stage), and facility characteristics that are associated with distance traveled to the treating hospital. For all analyses we estimate linear regression models with standard errors clustered at the county level. Preliminary results suggest that younger age, being insured, VA/Tricare insurance, a higher US News hospital quality score, hospital volume, and distance to nearest hospital were significantly associated with greater travel time. In addition, we found significantly greater travel times for prostate, cervical/uterine, and ovarian cancers but significantly shorter times for colorectal, breast, and lung cancers as well as cancers at a distant stage. Patient, hospital, and disease characteristics are all associated with significant differences in cancer care travel times. Understanding what factors are associated with patient cancer care location choices is critical as policymakers and payers attempt to encourage patients to seek care at higher quality facilities and will be particularly critical to understanding the potential future impact of alternative care model, particularly in oncology.

Time: 2:45 - 3:15 p.m.

It is well known that the accuracy of input data used for scientific investigations can affect the results generated from experiments, as well as the validity and power of the conclusions that can be drawn from them. These facts are particularly relevant when considering the spatial accuracy of data used in health research. For many years, qualitative metrics have been used to define the accuracy of geocoded data, which have in turn been used to judge the appropriateness of input data for epidemiological and disease surveillance and control purposes in public health. The existing measures used to indicate the spatial accuracy of geocoded data in health research are inadequate, sometimes misleading to research end-users, and are difficult to integrate into accuracy assessments in meaningful ways. Given prior results from large national-scale health investigations, it is possible to predict in advance what level of spatial accuracy can be obtained for different localities. Further, these same approaches can be used to re-assess the spatial accuracy of the data which underlie prior investigations to evaluate if inaccuracies impacted the validity of findings. This talk will examine the sources, magnitudes, and implications of error sources in the geocoding process, as well propose alternative spatial accuracy assessments that could and should be used in place of qualitative measures and hierarchical coding schemes that may imply false precision. Geocoding is an important first step in the health research process. It usually works well, but when it does not – it matters – and researchers must be provided with understandable and actionable measures to evaluate just how poorly it went.

Time: 3:15 - 3:30 p.m.
Co-Authors: Anne Moudon, PhD, University of Washington; Jason Y. Scully, MUP, University of Washington; Eric J. Howard, MUP, University of Washington; Brian E. Saelens, PhD, University of Washington

Spatial energetics research relies on processing of large amounts of sensor and GIS data. Correct data handling and processing is critical. The Transportation and Community Study (TRAC), focusing on longitudinal effects of the new light rail system in Seattle, WA, collected global positioning systems (GPS; Qstarz BTQ1000XT), accelerometry (Actigraph GT3X), and travel diary data from hundreds of adults for one week over three measurement waves. In the course of data processing and analysis we identified several technical problems induced by the device vendors' proprietary software (QTravel, Actilife), as well as commonly used analytic software (MS Excel and Access). We also identified issues surrounding the use of geographic information systems (GIS) software (ArcGIS) for geocoding and network routing. We present errors in time stamp handling related to time zones and daylight savings time in common software used for handling device data spatial energetics research. We also present some problems encountered integrating instrument data with GIS data. Furthermore, we present solutions to these problems. Results: Incorrectly accounting for standard/daylight savings time can induce time shifts of plus or minus one hour in GPS and accelerometry data. This can affect time synchronization among time stamped data tables, which can have downstream effects in estimating when and where physical activity takes place. Such locational problems may affect measurement of the built environment characteristics of physical activity bout locations. Likewise, problems with GIS data, software, and processing during geocoding and network routing can induce similar errors affecting estimation of built environment variable values at geocoded locations, and of the distances required to travel between known points. The large amount of data required for spatial energetics research precludes detailed review of all data, thus unforeseen errors may arise without analysts' knowledge. Data and software used in spatial energetics research may contain errors or omissions that compromise analysis and findings. More attention needs to be paid to these issues, with researchers sharing information and website and refereed journals reporting on the limitations of both data and software used. Solutions to identified problems may help other researchers in improving data processing and analysis.

Time: 3:30 - 3:45 p.m.
Co-Authors: Hong Gao, MPP, New York University; Andrew Anasetti-Rothermel, PhD, Truth Initiative

There is a great deal of interest in the way exposure to risk and protective factors within neighborhoods affects the health of people who traverse their streets. Real-time exposure estimates derived from actual mobility patterns are inherently more specific than residential approximations, but empirical methods for contrasting the predictive power of residential versus real-time exposure estimates are not well established. A nationwide density surface of convenience and related retail outlet locations was generated using kernel density estimation (KDE). This surface was linked to participants' (N=363) residential location, as well as their real-time geographic locations, recorded every 10-minutes over 180-days. Hourly mobility (N=1,567,156 total hours) was characterized by radius of gyration and associated contact with the retail density surface. Overall, 61.3% of real-time, hourly exposures were of relatively low intensity, and after controlling for temporal and seasonal variation, 72.8% of these low-level exposures was accounted for by residence in one of the two lowest residential density quintiles. Residence in the two highest residential density quintiles accounted for approximately 50% of extreme exposure levels, but extreme levels of exposure were rare, constituting about 1% of the data. Altogether 55.2% of real-time exposures were not explained by participants' residential density level, and of the unexplained exposures, 38.5% fell in the central, moderate range of the distribution. Results document the degree that residential exposure estimates may lead to misclassification, suggesting that a substantial proportion of real-time exposures will be misleadingly ignored, especially among people residing in areas characterized by moderate levels of residential density. This paper describes the development of an analytic framework for direct comparison of estimates based on residential versus real-time exposure to the landscape of retail outlets across the US, an industry vector for distribution of products known to heavily influence a range of health-related behaviors and clinical outcomes.

Time: 3:45 - 4:00 p.m.
Co-Authors: Juan Yang, PhD, MPH, Cancer Prevention Institute of California; Chiuchen Tseng, MS, University of Southern California; Shannon Conroy, PhD, MPH, Cancer Prevention Institute of California; Lianfa Li, PhD, University of Southern California; Jun Wu, PhD, University of California, Irvine; Daniel O. Stram, PhD, University of Southern California

Vehicle exhaust contains a mixture of gases and particulate matter with known mutagenic and carcinogenic effects. Using three complementary exposure methods, we examined the association between NOx, a marker of traffic-related air pollutants, and breast cancer risk among female Multiethnic Cohort participants, residing predominately in Los Angeles County. We estimated average NOx exposure for 57,530 females, using 93,148 residential addresses over a 17 year period of follow-up by: 1) temporally adjusted land use regression (LUR; high spatial and high temporal resolution); 2) California line source dispersion model, version 4 (CALINE4; high spatial and moderate temporal resolution); 3) Bayesian kriging interpolation of routine monitoring station data (low spatial and high temporal resolution). Cox proportional hazard regression was used to estimate hazard ratios (HR) and 95% confidence intervals (CI). Stratified analyses were conducted by race/ethnicity, moving status, and distance to monitoring stations. Breast cancer risk was significantly associated with an increase in the interquartile range in NOx exposure in African Americans (LUR: HR=1.24; 95% CI: 1.01-1.53) and Japanese Americans (LUR: HR=1.39; 95% CI: 1.05-1.83 and CALINE4: HR=1.13; 95% CI: 1.00-1.27). For the LUR model, stronger significant associations were observed in African Americans and Japanese Americans among non-movers (single residential address) and those living closer to monitoring stations ( < median distance=5 miles). No significant associations were identified by kriging interpolation. Exposure assessment with high spatial and high/moderate temporal resolutions identified long-term effects of NOx on breast cancer risk. Future analyses will examine associations by neighborhood- and individual-level factors.

Time: 4:00 - 4:15 p.m.
Co-Authors: James E. Chapman, MS C.E., Urban Design 4 Health, Inc.; Eric H. Fox, MScP, Urban Design 4 Health, Inc.; Jessica Schoner, PhD, Urban Design 4 Health, Inc.

Planners and public health officials seek to prioritize built environment investments which contribute to health behaviors important for chronic disease and cancer prevention. Despite this relevance, there is no widely available, nationally applicable decision-support planning tool to quantify health impacts at fine-grained geographies. This study creates resources to measure and optimize health impacts of transport and land use investments: 1) a national, consistent built, natural and social environment database at the Census block group level, and 2) a web-based health module to enhance the functionality of multiple existing scenario planning software tools. The health module contains statistical models of health behaviors and outcomes developed using participant address level data from the California's Health Interview and Household Travel Surveys. The models show sensitivity of health behaviors and outcomes to several useful environment variables, such as residential density and tree canopy. Using the health module, communities can estimate health behaviors and outcomes using the national database (existing conditions) and custom scenarios that reflect their land use and transport planning investment decisions. Successful health interventions may be implemented through strategies using the health module to enhance scenario planning tools' functionality to estimate health outcomes, including cancer risk factors. These data, models and tools allow more jurisdictions, even those which are challenged by insufficient data, staff and resources, to more easily include health impacts in planning decisions.

Time: 3:00 - 3:15 p.m.

To protect patients' confidentiality, the data provider often employs the geomasking technique to de-identify individuals. Geomasking moves a point from its original location to a random location that is within a certain spatial extent. The main concern in geomasking is the balance between de-identifying individuals and preserving the spatial pattern determined by the individual locations. So far there is a lack of method for the data provider to evaluate this balance, and often geomasking in practice is performed in a subjective or even arbitrary way. Based on the understanding that as long as the de-identification can be achieved, the moving distance should always be minimized, an adaptive geomaksing process is proposed, which adjusts the moving distance according to the local population density. To achieve this adaption, instead of specifying a geographic distance as the maximum moving distance, a safe size of population is specified, which eventually has more explicit statistical and epidemiological meanings than a subjectively chosen distance. Furthermore, when choosing a random location, a location that has a population density similar to that of the original location has a greater chance to be selected. An experiment with a simulated dataset demonstrates that the adaptive geomasking is more advantageous in both preserving the original spatial pattern and evaluating (and thus controlling) the balance between confidentiality and fidelity. The adaptive geomasking improves the conventional geomasking that is based on specified geographic distances. The process should be able to facilitate dissemination of health data for research purposes.

Time: 3:15 - 3:30 p.m.
Co-Authors: Stephen Mooney, PhD, Columbia University; Michael D.M. Bader, PhD, American University

Physical activity prevents several cancers, yet less that 50% of U.S. adults meet physical activity recommendations. There is a growing interest in using urban design to foster physical activity. These analyses investigate the associations between physical activity among residents of NYC and residential neighborhood walkability and access to green space. Sociodemographic information and data on physical activity measured by accelerometers were collected from 803 residents of NYC between November 2010 and November 2011. Physical activity was measured as minutes of moderate intensity equivalent activity per week ((minutes of moderate activity + (2 x minutes of vigorous activity)). The residential neighborhood was defined as a 1km radial buffer around the home. Neighborhood walkability was measured using a scale that combines data on residential density, land use mix, intersection density, retail space configuration and subway stop count. Access to green space was measured as the proportion of residential neighborhood covered with large parks spaces (park areas>6 acres) and small park spaces and the number of street trees per km of neighborhood area. After adjustment for individual and neighborhood sociodemographic characteristics, total weekly minutes of moderate intensity equivalent physical activity were significantly associated with neighborhood walkability (+26% across the interquartile range (IQR) of walkability, p<0.01), access to small and large park spaces (+12%, P<0.05, +11%, p=0.09 respectively, across the IQR of park access) and street tree density (+15% across the IQR of tree density, p<0.05). Physical activity among NYC residents is associated with neighborhood walkability and some measures of green space access.

Time: 3:30 - 3:45 p.m.
Co-Authors: Stephen A. Matthews, PhD, Penn State; Daniel Kifer, PhD, Penn State

Due to concerns about confidentiality, cancer patients' information is censored and any residential information is typically replaced by reference to coarse areal units (e.g., ZIP code). Censoring does not necessarily preserve the relative spatial associations in the data, leading to an empirical question of whether researchers can obtain meaningful statistical inference with censored geospatial data. We develop an approach based on Kullback-Leibler (KL) divergence (NSF, 1228669) to address this problem. We compute a distribution of areas where respondents may reside and sample from this distribution to generate an imputed dataset. To obtain this distribution, we combine hierarchical Census data with censored geospatial data to estimate a probability distribution over areas with multiple imputation. We offer the KL primer by applying the method to a large dataset (over 700,000 geocoded observations) to the study of housing in Arkansas, and discuss extensions in cancer research. Our approach facilitates the use of spatial analytic tools to censored geospatial data. Geographically weighted regression was used on the imputed data and we found that coefficient estimates fall within the 95% confidence intervals of the true parameters (drawn from the geocoded data). The same patterns hold for other spatial clustering measures (e.g., Moran's I). The results indicate that our approach can help researchers make meaningful statistical inference with censored geospatial data. Our approach serves as a primer on handling restricted cancer patient data as it allows researchers to obtain preliminary findings about the spatial associations with public data, facilitating the decision to apply for restricted data.

Time: 3:45 - 4:00 p.m.
Co-Authors: Shannon N. Zenk, PhD, MPH, RN, University of Illinois at Chicago; Zhenhui Li, PhD, Penn State; Aster Xiang, MA, University of Illinois at Chicago; Amber Kraft, University of Illinois at Chicago; Kirby McKinnon, MPH, University of Illinois at Chicago

Activity Space (AS) research offers a promising approach to understanding environmental influences on cancer-risk behaviors and outcomes. An objective of the ASCMEE measurement grant (5R21CA195543) is to develop tools to examine the joint spatial and temporal stability of AS. ASCMEE included the collection of 28-day GPS data (97 adults, Chicago); generating a geospatial database of 3+ million data points. We focus on the application of a new, flexible data aggregation tool to describe and synthesize the spatiotemporal patterns generated from GPS data. We introduce general and conditioned AS profile graphs to both represent accumulated exposure to places across time and to identify places where participants spend most time. We compare data for different sequences and durations of days across the 28-day period; and demonstrate the robustness of results through sensitivity analysis and the examination of specific cases. Results demonstrate heterogeneity in the size and complexity of individual AS over time, reveal different subgroups (rapid vs. slow accumulators) and the role of specific days or journeys in accumulated exposure. This method development scratches the surface of what is possible in AS research. The new ASCMEE tools are fast and efficient (do not require GIS), parameters options control spatial and temporal metrics and output can easily generate AS profile graphs. These methods/results have conceptual and substantive implications for the design of AS studies vis-à-vis the duration of data collection. This new tool can provide more accurate measures of spatiotemporal environmental exposures and ultimately a better understanding of environmental contributions to cancer risk behaviors.