Heart disease is the leading cause of death in the United States, with a large proportion of heart disease deaths attributable to acute myocardial infarction (AMI).1 Although age-adjusted rates are decreasing steadily in New York State (NYS), heart disease has remained the leading cause of death for more than a decade.2 Rates of AMI hospitalization in NYS for patients 35 years and older declined from 48.1 per 10 000 in 2000 to 32.2 per 10 000 in 2007, based on the 2000 US Standard Population, and then remained steady through 2008.3 AMI is a chronic condition of particular interest not only because of its sudden onset but also because quick medical response is critical to survival.4 Thus, patients experiencing AMI are likely to access hospitals in the nearest city irrespective of the state in which they are located.5 Spatial analysis can lend itself as a valuable tool to explore these relationships.
Previous studies have shown that residents seeking out-of-state care are “missed” in surveillance counts and morbidity counts. Rates of emergency department (ED) visits at health care institutions close to out-of-state hospitals tend to be lower, suggesting that residents in those areas sought out-of-state care.6 About 25% of patients with more than 1 hospital visit tend to seek emergency care outside their primary health care system.6 Several reasons for patients utilizing out-of-state health care have been posited, including health care cost and insurance,7 availability of providers,7–9 location of the nearest hospital,6 , 10 and perceptions of area hospitals.8 , 9 One study found that NYS residents receiving care at an Ohio clinic were not included in the NYS registry, which is expected since Ohio does not systematically share health data with NYS.5 A NYS hospital observed that its patients living on the state border sought care in Vermont; in response, they joined a Vermont health care network,10 which, in turn, joined a larger NYS health care network.11 The NYS hospitals in the network report to the NYS Department of Health (NYSDOH), but the Vermont hospitals in the network do not, although they treat NYS residents.
Experts have stressed the need for a statewide electronic repository for patient information to provide EDs with essential medical information on patients outside their network.6 Similarly, an electronic repository that includes care for in-state residents regardless of where they are treated could assist physicians to stay updated on all care received, as well as provide a full picture of acute health outcomes for the state.
We posit that patients who live near the state border will travel out of state for acute health care issues if the closest hospital is out of state. As a preliminary test, we mapped AMI rates by census tracts throughout NYS to show whether census tracts on the state border had noticeably lower AMI rates than census tracts farther from the state border. We hypothesized that there would be a difference in AMI rates between border and nonborder census tracts that may be due to NYS residents seeking care out of state, since out-of-state care is not captured in our data.
This study is a retrospective analysis of data from the Statewide Planning and Research Cooperative System (SPARCS) for NYS managed by the NYSDOH. Data were collected prospectively. Institutional review board approval was not necessary for this study. No confidential information was collected during this study. The SPARCS data set is described elsewhere.12 We used the NYS Environmental Public Health Tracking (EPHT) tier 1 system database to locate hospitals. EPHT data are useful for studying AMI because AMI has been linked to both air pollution and temperature changes.13
We defined “cases” as incidents of AMI among NYS residents, who were at least 35 years old, and discharged from a hospital that was not a pediatric or psychiatric specialty facility between 2005 and 2014. We defined incidents of AMI as having a primary diagnosis of 410 (AMI) in the International Classification of Diseases, Ninth Revision.14
Both ED data and inpatient data were downloaded for the years of interest and then merged in SAS15 to create one data set. We created a subset of the data using the following criteria: age at least 35 years with a principal diagnosis of AMI and a hospital discharge date between 2005 and 2014. The variables retained included residence details, basic demographics, and billing details; however, we did not use the billing details in this analysis. A total of 410 712 potential cases were identified (see Figure 1, Supplemental Digital Content, available at http://links.lww.com/JPHMP/A327).
We developed 2 geocoding keys to geocode cases' home addresses using MapMarker16 and ArcMap.17 The address details for the data were concatenated and merged with these keys to assign point coordinates using R18 for the MapMarker geocoding key and SAS for the ArcMap geocoding key.
After geocoding, 377 459 cases remained in the sample. These cases were mapped in ArcMap, where they were merged with the shapefile for the US Census 2010 census tracts,19 which resulted in 377 418 cases being successfully matched to census tracts. We exported the resulting table, with census tract data, to a database file and then imported into R using the foreign package20 for R.
In R, the data were compiled by census tract by year for a total of 4842 census tracts in the data set. The demographics for each of the 2010 census tracts were exported from a census file using MapInfo and imported into R. We used these demographics as denominators to calculate the rates for each subpopulation and overall population by census tract. All census tracts with infinite values for overall AMI rates (13 tracts, 65 cases) or with overall or age-adjusted AMI rates over 1000 (12 tracts, 47 cases) were removed from the sample. Infinite values occurred when no one lived in those tracts according to the 2010 US Census, which suggests that we may not have had the correct addresses for those cases. The final sample size was 377 306 cases.
We exported the table of census tracts to ArcMap and joined it to the 2010 census tracts shapefile19 to generate thematic maps of crude (unadjusted) and age-adjusted AMI rates. We overlaid the 2 layers of hospital data onto these maps. The NYS hospital locations were obtained from the EPHT database.21 The locations of hospitals both in neighboring states and in Canada were identified using those states' Department of Health Web sites if available (Pennsylvania22 and Vermont23) or from Google Maps.24 The locations of the hospitals for neighboring states were plotted at zip code level using MapMarker. We retained only the hospitals within 20 miles of NYS in the map.
AMI rates include both fatal cases and nonfatal cases, so individuals who experienced multiple AMI events are included for each event. We calculated both unadjusted and age-adjusted AMI rates for each year and census tract in R. We calculated unadjusted AMI rates for each year and census tract that contained at least one case discharged that year using census tract–level population data from the 2010 US census. We calculated age-adjusted AMI rates using the 2010 US Standard Population.25 When census tracts had zero cases in a given year, we treated the data for that year as missing under the assumption that no cases meant no health care utilization for AMI. We calculated the unadjusted and age-adjusted overall rates for each year as the means of the census tracts' rates for that year, excluding missing rates.
We joined the mean age-adjusted and unadjusted AMI rates to a 2010 census tract shapefile in ArcGIS. We calculated a mean of the rates over 10 years for each census tract, which controls for some of the instability that may arise from very few cases in a census tract by averaging the rates over time. We coded the rates into 4 categories based on distance from the state mean. We created thematic maps and inspected them visually for patterns of higher and lower AMI rates. We compared these patterns with locations of hospitals both in and outside NYS and with t tests of border versus nonborder census tracts.
Finally, we ran Cramer's V measures of effect test to compare the demographics for the 377 306 cases that were included in the map with the demographics for the 33 406 cases that were excluded from the map.
The study population was mostly male (59%) and mostly white (72%) (see Table 1, Supplemental Digital Content, available at http://links.lww.com/JPHMP/A328). Only 11% of the geocoded cases self-reported their race as black. Half of the study population was 70 years and older, and 90% was 50 years and older.
The statewide annual AMI rate was highest in 2005 (Table). Adjusting for age reduced the AMI rates slightly for all years included in the study. Results from t tests were significant (P < .05) when comparing the unadjusted AMI rates in border (mean = 29.64) and nonborder (mean = 39.41) census tracts (95% confidence interval [CI], 3.46-13.73 cases per 10 000 people) and the age-adjusted AMI rates in border (mean = 25.19) and nonborder (mean = 33.78) census tracts (95% CI, 6.94-12.60 cases per 10 000 people).
Since the proportion of blacks in the study population was low despite blacks being considered a high-risk population for heart disease,26 we used Cramer's V measures of effect tests to check whether this discrepancy could be explained by amenability of cases to geocoding. The results show a very weak association (ϕ = 0.034 with 95% CI, 0.031-0.037) between race and whether or not cases were included in the map. We also calculated measures of effect for border versus nonborder counties, age groups, and sex. All tests showed very weak measures of effect (ϕ < 0.05 with no 95% CI > 0.05).
The population density (data not shown) varies by census tract throughout NYS. When compared with the mean annual unadjusted AMI rate by census tract (see Figure 2, Supplemental Digital Content, available at http://links.lww.com/JPHMP/A329), higher unadjusted AMI rates roughly corresponded with denser populations in the western part of the state but not in the eastern and southern parts of the state. Along the New England border with Vermont, Massachusetts, and Connecticut, the mean AMI rates in all census tracts were lower than the statewide mean of annual AMI rates before adjusting for age. Along the Pennsylvania border, there was a greater variation in rates, but generally higher rates were closer to NYS hospitals and lower rates were closer to Pennsylvania hospitals. Along the Canada border, rates tended to be higher than those along the borders with other states.
The means of the annual age-adjusted AMI rates at census tract level (Figure) were lower near Connecticut and Massachusetts than the statewide mean of age-adjusted AMI rates. Along the Vermont border, there was a greater variation in rates. Along the Pennsylvania border, higher rates were closer to NYS hospitals and lower rates were closer to Pennsylvania hospitals. Along the New Jersey border, rates were generally higher than the mean of rates over the 10 years. Along the Canada border, rates were higher than those along the borders with other states. Areas in western NYS tended to have higher age-adjusted AMI rates irrespective of population density. In eastern NYS, areas with high population densities had higher rates and areas with low population densities had lower rates.
As with NYS overall, in New York City (NYC), areas with lower population densities had lower age-adjusted AMI rates, with the exception of Staten Island (see insets in Figure 2, Supplemental Digital Content, available at http://links.lww.com/JPHMP/A329, and the Figure).
Rapid medical response is critical for patients who experience AMI,4 so it is reasonable to assume that patients experiencing AMI would be brought to the nearest hospital irrespective of the state in which the patient or hospital is located. As other studies have found,6 and as we expected, AMI rates are lower in border census tracts and higher in nonborder census tracts when assessed both visually and with t tests.
Along the border with New England, the unadjusted AMI rates in all census tracts were below the mean of unadjusted AMI rates over the 10 years. After adjusting for age, AMI rates in the census tracts near Massachusetts and Connecticut were below the statewide mean age-adjusted AMI rate. This supported our supposition that lower AMI rates along this border are due to out-of-state care because most of the closest hospitals are in New England and not in NYS. Near the Vermont border, there was a greater variation in rates, perhaps due, in part, to Lake Champlain separating NYS from Vermont. In one section along the Vermont border, some AMI rates were more than 125% of the mean; this area is close to a NYS hospital with a cardiac specialization. In a section along the Connecticut border, AMI rates were also more than 125% of the mean; this section is close to multiple hospitals on both sides of the border, as well as multiple hospitals with cardiac specializations in NYC. Nassau County and western Suffolk County had lower AMI rates despite higher population densities and being located on an island, which could be due to factors that were not considered in this study.
Along the southern border with Pennsylvania, the unadjusted and age-adjusted AMI rates showed more variation than along the eastern border, but generally, higher rates were closer to NYS hospitals and lower rates were closer to Pennsylvania hospitals. This pattern did not hold along the New Jersey border, where AMI rates were generally higher irrespective of hospital location. This may be due to road networks or topography, which are beyond the scope of this preliminary study.
We chose to exclude nearby Canadian hospitals from our analysis for several reasons. First, the logistics of traveling across an international border might be too time-consuming in a medical emergency. Second, the border between NYS and Canada is separated by water with a limited number of bridges, so the closest hospital on a map may not be the closest hospital by road. Third, the NYSDOH does not have access to Canadian insurance/billing information. Along the Canada border, near Niagara County and the St Lawrence River, unadjusted and age-adjusted AMI rates tended to be higher than those along the borders with other states. This supports the supposition that while people may be seeking care in other states, they are not seeking care in Canada. This is reasonable because interstate travel is faster and easier than international travel. In the northeastern corner of NYS, AMI rates are low, but that could be attributed to seeking care in Vermont and not in Canada.
When compared with the population density map (not shown), we expected higher unadjusted AMI rates to correspond with denser populations. This is somewhat true for the western part of NYS, but it is not true for the eastern and southern parts of NYS. When comparing the age-adjusted AMI rates and population densities, areas in the western half of NYS tended to have higher age-adjusted AMI rates irrespective of population density. In the eastern half of NYS, areas with higher population densities also tended to have higher age-adjusted AMI rates and areas with lower population densities tended to have lower age-adjusted AMI rates.
As with NYS overall, areas in NYC with lower population densities (data not shown) also appeared to have lower age-adjusted AMI rates. This makes sense, since none of the NYC census tracts share borders with other states, and given the presence of numerous hospitals in NYC, it is unlikely residents seek care outside NYC for acute conditions. The notable exception is on Staten Island, despite the fact that it is unlikely someone would leave the island to seek care for AMI when there are 2 hospitals on the island with cardiac specializations. This discrepancy may be due to additional determinants that were not examined in this study.
In nonborder census tracts with lower AMI rates, we may need to explore additional variables. In our data set, 72% of cases were white and 11% were black. In border census tracts with high AMI rates, it is possible that there is no direct road to nearby hospitals across the border.27 We did not address topography or transit routes that could affect where people seek health care in this study.
This study has several limitations. Transfer from one health care facility to another health care facility within the state during treatment of the same AMI event could have led to an overestimation of AMI rates. We also did not consider transit systems or topography in our analysis, so those residents living in a census tract close to a hospital but separated from it by a land or water barrier may have utilized a hospital farther away and possibly out of state. In addition, people along the border may be more likely to die before they get to a hospital for reasons beyond the scope of this study. It is possible that these cases would not be captured in SPARCS. Finally, the EPHT database records billing address, not residential address, and in some cases, the 2 addresses may not be the same.
Our preliminary analysis found that both unadjusted and age-adjusted AMI rates differ between border and nonborder census tracts. These visual patterns and results seem to suggest that people along state borders seek care out of state. As a result, states cannot get a complete picture of the health of their residents unless they share data with other states. Our findings seem to support the view that states should share health data with neighboring states.
Because our findings are preliminary, additional research is required. We plan to explore whether patients utilized the hospitals closest to their addresses and whether specific hospitals see the most AMI patients. Furthermore, we can explore whether higher AMI rates tend to cluster around hospitals, indicating that people farther from hospitals may die before they can reach them. We can also consider whether patients in NYS who live in neighboring states tend to have addresses close to the NYS border, which would support our supposition that people seek care in states other than their state of residence for acute conditions such as AMI.
Implications for Policy & Practice
- To address needs of NYS residents, the NYSDOH should be able to identify those needs. Residents who receive care out of state are not captured currently by the state EPHT database. As a result, the NYSDOH currently has no information on the care they receive, so its understanding of the health of NYS residents is incomplete. If the NYSDOH could collect this information routinely, it could be used to assess residents' health care needs and to identify areas in the state where more resources are needed.
- Developing a data-sharing relationship with neighboring states would assist the other states as well as NYS and could provide insight into regional trends. It would also assist patients treated out of state by helping their primary physicians to identify treatments they have received elsewhere.
In our study, the application of spatial analysis allowed us to visualize census tract rates in NYS in relation to their neighbors and in relation to the location of hospitals both in NYS and near the NYS border. Our findings can inform steps to establish memoranda of understanding with neighboring states to share patient data and help determine whether lower AMI rates along state borders are due to NYS residents seeking treatment at hospitals out of state.
1. Mozaffarian D, Benjamin EJ, Go AS, et al AHA statistical update: heart disease and stroke statistics—2015 update. Circulation. 2015;131(4):e29–e322.
2. Public Health Information Group. Leading causes of death, New York State, 2005-2014. New York State Department of Health Web site. http://www.health.ny.gov
/statistics/leadingcauses_death/nys_by_year.htm. Updated June 2016. Accessed September 15, 2016.
3. Talbott EO, Rager JR, Brink LL, et al Trends in acute myocardial infarction hospitalization rates for US states in the CDC Tracking Network. PLoS ONE. 2013;8(5):e64457.
4. Warning signs of heart attack, stroke & cardiac arrest. American Heart Association Web site. http://http://www.heart.org
/HEARTORG/Conditions/911-Warnings-Signs-of-a-Heart-Attack_UCM_305346_SubHomePage.jsp. Published June 2016. Accessed October 5, 2016.
5. Yamashita T, Kunkel SR. The association between heart disease mortality and geographic access to hospitals: county level comparisons in Ohio, USA. Soc Sci Med. 2010;70(8):1211–1218.
6. Finnell JT, Overhage JM, Grannis S. All health care is not local: an evaluation of the distribution of emergency department care delivered in Indiana. AMIA Annu Symp Proc. http://www.researchgate.net
/publication/51921209_All_health_care_is_not_local_an_evaluation_of_the_distribution_of_Emergency_Department_care_delivered_in_Indiana. Published January 2011. Accessed October 6, 2016.
7. Omoigui NA, Miller DP, Brown KJ, et al Outmigration for coronary bypass surgery in an era of public dissemination of clinical outcomes. Circulation. 1996;93(1):27–33.
8. Hare TS, Barcus HR. Geographical accessibility and Kentucky's heart-related hospital services. Appl Geogr. 2007;27(3):181–205.
9. Caliendo M. Hospitals turn focus on keeping flock in NJ. NJBIZ. February 18, 2013:1–6.
10. Velasquez J. With borders irrelevant, New York hospitals seek out-of-state partners. Politico. July 25, 2016. http://http://www.politico.com
/states/new-york/albany/story/2016/07/new-york-hospitals-pair-with-out-of-state-health-systems-104138. Accessed September 16, 2016.
11. Member Directory. Healthcare Association of New York State. http://www.hanys.org
/member_directory. Accessed October 7, 2016.
12. Bureau of Health Informatics, Office of Quality and Patient Safety. 2014 SPARCS Operations Guide. Albany, NY: NYS Department of Health; 2014. http://http://www.health.ny.gov
/statistics/sparcs/training/docs/sparcs_operations_guide.pdf. Accessed October 7, 2016.
13. Claeys MJ, Rajagopalan S, Nawrot TS, Brook RD. Climate and environmental triggers of acute myocardial infarction. Eur Heart J. 2017;38(13):955–960.
14. International Classification of Diseases, 9th Revision; Clinical Modification, 6th edition. Hospital Edition. Los Angeles, CA: Practice Management Information Corporation; 2011.
15. SAS [computer program]. Version 9.4. Cary, NC: SAS Institute Inc, 2012.
16. MapMarker USA [computer program]. Version 28. Troy, NY: Pitney Bowes Software Inc; 2015.
17. ArcGIS for Desktop [computer program]. Version 10.3. Redlands, CA: Environmental Systems Research Institute; 2014.
18. R: A language and environment for statistical computing [computer program]. Version 3.1.0. Vienna, Austria: R Foundation for Statistical Computing; 2014.
19. TIGER/Line® Shapefiles and TIGER/Line® Files. United States Census Bureau Web site. http://www.census.gov
/geo/maps-data/data/tiger-line.html. Updated March 26, 2012. Accessed August 8, 2016.
20. foreign: Read Data Stored by Minitab, S, SAS, SPSS, Stata, Systat, Weka, dBase, ... [computer program]. Version 0.8-66. Vienna, Austria: R Foundation for Statistical Computing; 2015.
21. Talbot TO, Kumar S, Babcock GD, Haley VB, Forand SP, Hwang S. Development of an interactive environmental public health tracking system for data analysis, visualization, and reporting. J Public Health Manag Pract. 2008;14(6):526–532.
22. Directory of Pennsylvania Hospitals and Ambulatory Surgery Centers. Bureau of Health Statistics, Pennsylvania Department of Health Web site. http://http://www.health.state.pa.us
/pdf/hpa/stats/hospdir98/hospdir98.pdf. Updated November 1999. Accessed September 15, 2016.
23. Division of Licensing and Protection, Vermont Department of Disabilities, Aging, & Independent Living. Hospitals. http://http://www.dlp.vermont.gov
/other/hospitals. Accessed September 15, 2016.
25. Standard populations (millions) for age-adjustment. Surveillance, Epidemiology, and End Results Program, National Cancer Institute Web site. http://seer.cancer.gov/stdpopulations. Accessed August 8, 2016.
26. Understand your risks to prevent a heart attack. American Heart Association Web site. http://http://www.heart.org
/HEARTORG/Conditions/HeartAttack/UnderstandYourRiskstoPreventaHeartAttack/Understand-Your-Risks-to-Prevent-a-Heart-Attack_UCM_002040_Article.jsp. Published June 2016. Accessed October 26, 2016.
27. Boscoe FP, Henry KA, Zdeb MS. A nationwide comparison of driving distance versus straight-line distance to hospitals. Prof Geogr. 2012;64(2):188–196.