Disparities in geographic access to medical oncologists

Abstract

Objective

The objective of this study is to identify disparities in geographic access to medical oncologists at the time of diagnosis.

Data Sources/Study Setting

2014–2016 Pennsylvania Cancer Registry (PCR), 2019 CMS Base Provider Enrollment File (BPEF), 2018 CMS Physician Compare, 2010 Rural‐Urban Commuting Area Codes (RUCA), and 2015 Area Deprivation Index (ADI).

Study Design

Spatial regressions were used to estimate associations between geographic access to medical oncologists, measured with an enhanced two‐step floating catchment area measure, and demographic characteristics.

Data Collection/Extraction Methods

Medical oncologists were identified in the 2019 CMS BPEF and merged with the 2018 CMS Physician Compare. Provider addresses were converted to longitude‐latitude using OpenCage Geocoder. Newly diagnosed cancer patients in each census tract were identified in the 2014–2016 PCR. Census tracts were classified based on rurality and socioeconomic status using the 2010 RUCA Codes and the 2015 ADI.

Principal Findings

Large towns and rural areas were associated with spatial access ratios (SPARs) that were 6.29 lower (95% CI −16.14 to 3.57) and 14.76 lower (95% CI −25.14 to −4.37) respectively relative to urban areas. Being in the fourth ADI quartile (highest disadvantage) was associated with a 12.41 lower SPAR (95% CI −19.50 to −5.33) relative to the first quartile. The observed difference in a census tract's non‐White population from the 25th (1.3%) to the 75th percentile (13.7%) was associated with a 13.64 higher SPAR (Coefficient = 1.10, 95% CI 11.89 to 15.29; p < 0.01), roughly equivalent to the disadvantage associated with living in the fourth ADI quartile, where non‐White populations are concentrated.

Conclusions

Rurality and low socioeconomic status were associated with lower geographic access to oncologists. The negative association between area deprivation and geographic access is of similar magnitude to the positive association between larger non‐White populations and access. Policies aimed at increasing geographic access to care should be cognizant of both rurality and socioeconomic status.

What is known on this topic

• Geographic access to oncology care varies considerably across the United States.

• Access to timely diagnosis and treatment is associated with increased survival, improved quality of life, and lower treatment costs.

What this study adds

• Our measure of geographic access to oncologists, which considers patient demand, provider capacity, and travel time, is negatively associated with rurality and low socioeconomic status.

• Larger non‐White populations are associated with increased geographic access; however, minority populations tend to concentrate in deprived areas, which are associated with lower geographic access.

• These results highlight the importance of focusing not only on rurality but also on socioeconomic status when designing policies targeted toward improving geographic access to medical oncologists.

1. INTRODUCTION

Geographic access to oncology care varies considerably across the United States. ¹ Previous research has found geographic access to oncology is higher in census tracts that are densely populated ² and that oncologists choose to practice in urban areas more often than in rural areas. ³ , ⁴ Geographic access can improve the timeliness of diagnosis and treatment, which are associated with increased survival rates, improved quality of life, and lower cost of treatment for many types of cancer. ⁵ , ⁶ , ⁷ Limited geographic access to oncology providers is associated with late‐stage diagnoses, longer intervals between diagnosis and treatment, lower treatment completion, and lower quality of life. ³ , ⁸ , ⁹

Prior research examining demographic and geographic disparities in access to cancer care has yielded mixed results, indicating a complex relationship between geography, access to care, and outcomes that are also associated with factors such as race and ethnicity. One study concluded that patients, particularly non‐Hispanic Black and Hispanic patients, may opt to forgo treatment due to geographic barriers such as distance to providers. ¹⁰ Rates of late‐stage diagnosis are particularly high among non‐Hispanic Black patients and populations with low socioeconomic status. ¹¹ , ¹² Even after accounting for geographic access, there are racial, ethnic, and socioeconomic disparities in colorectal cancer late‐stage diagnosis. ¹² However, other studies have found evidence of a “reversed racial disadvantage” in which Non‐Hispanic Whites have the lowest geographic access to NCI‐designated cancer treatment centers. ¹ This “reversed racial disadvantage” may still be consistent with the aforementioned findings since NCI‐designated cancer centers do not represent all cancer treatment facilities, and other studies show worse outcomes for minority racial and ethnic groups. ¹¹ , ¹² Further, research has indicated that hospital systems may focus expansion of specialized services, such as oncology care, in more affluent neighborhoods, which may lead to disparities in geographic access to care by area socioeconomic status. ¹³ , ¹⁴

Access is a broad and important concept in health services research that has been conceptualized as involving multiple dimensions including availability, accessibility, accommodation, affordability, and acceptability. ¹⁵ The measure of geographic access that we use in this analysis captures two geographic dimensions of access, availability, and accessibility (i.e., the supply of available health care services and their relationship to patient needs and geographic proximity). No prior study has used a robust measure of geographic access to care that accounts for provider supply, patient demand, and travel time to simultaneously examine disparities in geographic access to medical oncologists by rurality, race/ethnicity, and socioeconomic status.

In this study, we use an Enhanced Two‐Step Floating Catchment Area (E2SFCA) ¹⁶ approach to examine geographic access to medical oncologists in Pennsylvania, the state with the fifth highest number of new cancer cases annually. ¹⁷ Pennsylvania has a mix of urban and rural areas and generally reflects national demographics with the exception that Pennsylvania has a smaller non‐Black minority population. ¹⁸ , ¹⁹ The E2SFCA simultaneously considers potential patient demand, provider capacity, and travel time to measure geographic access to care. It has been consistently used in the geography literature for over a decade but has only been used to a limited extent in the clinical oncology literature. ¹ , ³ , ¹¹ , ²⁰ Our objective was to identify geographic disparities in access to medical oncologists by rurality, socioeconomic status, and demographic characteristics, particularly race and ethnicity, using the E2SFCA approach.

Characterizing disparities in geographic access and identifying areas of low geographic access can support policy makers and providers seeking to allocate resources efficiently and equitably. For instance, decisions regarding facility openings, closings, or changes in service offerings; incentives to clinicians related to practice location; or programs supporting telemedicine or visiting providers can use the information on geographic access to improve efforts to target patient populations that lack access to care.

2. METHODS

2.1. Data

We used information on the locations of medical oncologists and all incident cancer cases in the state of Pennsylvania. We used incident cancer cases in an area because it gives us a consistent measure of the burden of new cancer cases and is the clearest available measure of the relative demand for care.

Pennsylvania's demographics generally reflect national demographics in terms of age, race (except for smaller non‐Black minority population), and ethnicity, persons per household, education, income, and labor force participation. ¹⁸ , ¹⁹ We used the 2019 Centers for Medicare & Medicaid Services (CMS) Base Provider Enrollment File (BPEF), ²¹ a list of all Medicare providers, to identify medical oncologists in Pennsylvania. We identified medical oncologists as providers with a specialty in hematology, hematology/oncology, or medical oncology. Because the BPEF does not contain providers' addresses, we merged it with the 2018 CMS Physician Compare data, ²² which contains the addresses of Merit‐based Incentive Payment System‐eligible Medicare providers. In cases where there was not a match between the two data sources—this was the case for 24.3% of the identified providers in the BPEF—we used a Google search to find street addresses. Specifically, we queried the provider's first and last name along with “MD,” and we used the first returned address. We converted provider addresses to longitude‐latitude coordinates using OpenCage Geocoder. ²³

We identified the number of newly diagnosed cancer patients in each census tract by averaging annual patient counts in the 2014–2016 Pennsylvania Cancer Registry data, ²⁴ a statewide data system responsible for collecting information on all new cases of cancer diagnosed or treated in Pennsylvania.

We used two other data sources to classify census tracts based on rurality and socioeconomic status of the local population: (1) the 2010 Rural‐Urban Commuting Area Codes (RUCA), ²⁵ which classify U.S. census tracts based on population density, urbanization, and area commuting patterns, from the U.S. Department of Agriculture Economic Research Service; and (2) the 2015 Area Deprivation Index (ADI) ²⁶ from the University of Wisconsin, which ranks neighborhoods by socioeconomic status and includes factors for the theoretical domains of income, education, employment, and housing quality. We used the 2018 5‐Year American Community Survey ²⁷ by the US Census Bureau to obtain other census tract demographics.

2.2. Measures

The unit of analysis was the census tract, the most granular level at which we observed patient locations. We assumed patients were located at the geographic center of their census tract. We used MapBox, ²⁸ a provider of customizable online maps, to create car travel‐based provider location catchment areas (e.g., the area within a 60‐min drive of a physician's practice location). We used driving time, not driving distance, because it explicitly considers road structure, congestion, and so forth, and represents a uniform estimate of travel time regardless of the built environment, something that is likely to differ between urban and rural areas. We considered providers within 60 min of a census tract geographic center to be “accessible,” or within the patient census tract catchment area. Three drive time sub‐zones were defined —20, 40, and 60 min—to represent low, medium, and high travel times. These threshold drive times are broadly consistent with CMS Medicare Advantage network adequacy standards. ²⁹

Next, we used an E2SFCA method ¹⁶ , ³⁰ to measure geographic access to medical oncologists. The E2SFCA simultaneously accounts for provider capacity, patient demand, and travel time facing patients. Intuitively, it is a ratio of medical oncologists to cancer patients weighted by the capacity of medical oncologists, demand from cancer patients, and travel times facing cancer patients.

We created the E2SFCA measure using a two‐step process. First, we measured provider availability by calculating oncologist‐to‐cancer patient ratios for each medical oncologist location. We calculated these ratios by dividing the number of providers at a location by the number of patient‐weighted census tracts whose geographic centers fall within the provider location's catchment area. We assigned lower weights to census tracts located in sub‐zones further away from provider locations (i.e., census tracts that are closer to providers are weighted more heavily because patients living in them are more likely to demand nearby providers' services). Second, for each census tract, we summed the oncologist‐to‐cancer patient ratios from all the provider locations inside the patient census tract's catchment area (i.e., ratios in further sub‐zones received smaller weights to account for higher travel times). This sum, which is the E2SFCA measure, represents the availability and accessibility of medical oncologists within a feasible driving time of a given census tract's patient population.

We determined the magnitude of the weights using an algorithm developed by Kwan ³¹ —the standard in the E2SFCA literature—that uses a Gaussian function to assign distance decay weights to different sub‐zones. Intuitively, this approach allows travel time costs to increase similarly to a bell curve, increasing gradually at first and more abruptly as travel time increases. The weights for our 0–20, 20–40, and 40–60 min sub‐zones were 1, 0.48, and 0.05. This implies, for example, that oncologists in the 20–40 min sub‐zone were counted slightly less than half as much as those in the 0–20 min sub‐zone, and that patients that were 20–40 min away from an oncologist placed 48% as much demand on them as patients 0–20 min from an oncologist. We tested the sensitivity of our results to these weights below.

We report E2SFCA values as spatial access ratios (SPARs). A census tract's SPAR is simply its E2SFCA value over the mean of E2SFCA value for the sample. ³² We multiply SPAR by 100 for scaling purposes. For example, a census tract with a SPAR of 130 would have 30% more geographic access than the mean census tract, which has a SPAR value of 100. We provide a more detailed description of the E2SFCA approach and weighting in the technical appendix (Data S2). For further discussion of the E2SFCA measure, see Luo and Qi ¹⁶ and Drake et al. ³⁰ For further discussion of the weight approach, see Kwan, ³¹ Wang, ³³ and Drake et al. ³⁰

We grouped RUCA codes into three categories of rurality: urban (RUCA codes 1–3), large town (RUCA codes 4–6), and rural (RUCA codes 7–10). We aggregated national ADI rankings of census block groups to the census tract level based on the mean rankings for census block groups within a census tract. We then divided census tracts into four ADI quartiles so that each quartile had an equal number of patients. We measured other census tract demographics as percentages of the population, including race/ethnicity (percentage non‐Hispanic White, non‐White, and unknown), insurance type (Medicare, Medicaid, private, uninsured, other/unknown), age (<50, 50–64, 65–79, 80+), and gender.

2.3. Statistical analysis

We calculated mean geographic access (i.e., SPAR) values across RUCA codes (urban, large town, rural), ADI quartiles, and census tract demographic characteristics, as well as by rurality and ADI quartile simultaneously.

We estimated multivariate linear regressions and spatial error models at the census tract level to examine the associations between geographic access and rurality, ADI quartiles, and census tract demographics. Spatial error models account for spatial dependence in error terms, which could exist due to the granular geographic nature of our analysis. We estimated our spatial error models using a generalized spatial two‐stage least squares estimator. We tested for spatial dependence using a Moran test, which tests whether the residuals of linear regression are correlated with the residuals of nearby geographic units (i.e., census tracts). We defined census tract adjacency using queen contiguity, which considers census tracts to be adjacent if any two points of their borders touch. For a detailed discussion of spatial error models, see Anselin and Bera. ³⁴

This study was deemed exempt by the institutional review board at the University of Pittsburgh. We conducted our analyses in R 3.6.3, Stata‐SE 16, and SAS 9.4.

3. RESULTS

3.1. Descriptive results

The study included 3190 census tracts, 255,198 patients in the Pennsylvania Cancer Registry, and 744 Pennsylvania‐based medical oncologists.

Figure 1 maps the number of cancer patients by census tract along with medical oncologist locations. Figure 2 displays the values of SPARs of medical oncologists across Pennsylvania census tracts. Geographic access, as measured by SPARs, was higher near Pennsylvania's largest cities, including Pittsburgh, Philadelphia, and Harrisburg. There was a large variation in geographic access across rural Pennsylvania. For instance, geographic access was relatively high along the northern border with New York, but relatively low along the southwestern border with West Virginia.

FIGURE 1.

Locations of cancer patients and medical oncologists. The 2019 CMS Base Provider Enrollment File was used to identify medical oncologists in Pennsylvania. The 2018 CMS Physician Compare data and Google searches were used to find medical oncologist addresses. These addresses were geocoded using OpenCage Geocoder. Data on cancer patient census tracts were obtained from the 2014 to 2016 Pennsylvania Cancer Registries. [Color figure can be viewed at wileyonlinelibrary.com]

FIGURE 2.

Geographic access to medical oncologists from Pennsylvania census tract. Geographic access is reported as spatial access ratios (SPARs). For example, a census tract with a SPAR of 150 implies that a census tract has 50% greater geographic access than the mean Pennsylvania census tract. A census tract's SPAR is its enhanced two‐step floating catchment area (E2SFCA) value over the mean E2SFCA for all census tracts. The E2SFCA, as described in the text and the technical appendix (Data S2), simultaneously incorporates medical oncologist supply, patient demand, and travel time to measure geographic access to oncologists. [Color figure can be viewed at wileyonlinelibrary.com]

Table 1 reports population‐weighted demographic characteristics of sample census tracts and cancer patients in Pennsylvania. It shows demographics overall and stratified by RUCA and ADI. The majority (86.8%) of the census tracts in the sample were urban. The mean census tract was 11.8% non‐White. The non‐White cancer population percentage was 1.3% in rural tracts and 13.2% in urban census tracts. In the lowest three ADI quartiles, the non‐White cancer population percentage varied from 6.3% in the second quartile to 7.8% in the third quartile. However, in the fourth quartile—the most disadvantaged—the non‐White cancer population was 26.2%.

TABLE 1.

Demographic characteristics of Pennsylvania cancer patients

Demographic		Rural‐urban commuting area group a			Area deprivation index quartile b
	Overall	Urban	Large Town	Rural	First	Second	Third quartile	Fourth
Census tract characteristics
Census tracts	3190	2769 (86.80%)	273 (8.56%)	148 (4.64%)	706 (22.84%)	662 (20.75%)	769 (24.11%)	1053 (33.00%)
Rural–urban commuting area group a
Urban	2769 (86.80%)	—	—	—	701 (99.29%)	627 (94.71%)	621 (80.75%)	820 (77.87%)
Large town	273 (8.56%)	—	—	—	4 (0.57%)	24 (3.63%)	97 (12.61%)	148 (14.06%)
Rural	148 (4.64%)	—	—	—	1 (0.14%)	11 (1.66%)	51 (6.63%)	85 (8.07%)
Area deprivation index quartile b
First	706 (21.98%)	701 (25.32%)	4 (1.47%)	1 (0.68%)	—	—	—	—
Second	662 (20.61%)	627 (22.64%)	24 (8.79%)	11 (7.43%)	—	—	—	—
Third	769 (23.94%)	621 (22.43%)	97 (35.53%)	51 (34.46%)	—	—	—	—
Fourth	1053 (32.78%)	820 (29.61%)	148 (54.21%)	85 (57.43%)	—	—	—	—
Characteristics of patients in census tracts c
Cancer patients	85,066	73,875 (44.5%)	7246 (8.5%)	3946 (4.6%)	21,284 (25.0%)	21,233 (25.0%)	21,267 (25.0%)	21,282 (25.0%)
Gender
Male	48.0%	47.7%	50.3%	49.5%	48.0%	48.5%	48.0%	47.4%
Female	52.0%	52.3%	49.7%	50.5%	52.0%	51.5%	52.0%	52.6%
Age
0–49	11.9%	12.1%	11.0%	10.0%	12.4%	11.2%	11.4%	12.5%
50–64	32.0%	32.2%	31.1%	30.2%	31.9%	30.8%	31.3%	34.1%
65–79	38.7%	38.3%	40.6%	42.1%	38.1%	39.5%	39.5%	37.6%
80+	17.4%	17.4%	17.4%	17.6%	17.6%	18.5%	17.8%	15.8%
Race/ethnicity d
NH‐White	85.5%	83.8%	95.5%	97.7%	88.0%	91.2%	90.4%	72.2%
Nonwhite	11.8%	13.2%	3.4%	1.3%	6.9%	6.3%	7.8%	26.2%
Unknown	2.8%	3.0%	1.2%	0.9%	5.2%	2.5%	1.8%	1.6%
Insurance
Medicare	46.9%	45.9%	52.9%	54.4%	43.7%	47.6%	48.1%	48.2%
Medicaid	5.7%	5.6%	6.1%	5.5%	2.3%	3.5%	5.4%	11.5%
Private	33.4%	34.1%	30.0%	27.9%	39.7%	32.6%	32.3%	29.0%
Uninsured	0.8%	0.8%	0.9%	0.8%	0.6%	0.9%	0.8%	0.9%
Other/unknown	13.2%	13.6%	10.1%	11.5%	13.8%	15.4%	13.4%	10.3%

Abbreviation: NH, non‐Hispanic.

^a 2013 Rural‐Urban Continuum Codes from the Department of Agriculture Economic Research Service. RUCA codes are grouped into three categories: urban (RUCA codes 1–3), large town (RUCA codes 4–6), and rural (RUCA codes 7–10).

^b 2015 Area Deprivation Index from the University of Wisconsin's Department of Medicine.

^c Patient data averaged over 2014–2016 Pennsylvania Cancer Registries.

^d Pennsylvania populations are small (particularly in rural areas); therefore, the Asian and Hispanic population sizes are negligible and not included here.

Table 2 reports mean SPAR values by ADI and RUCA groups. Mean SPARs (SD) did not exhibit a consistent pattern by ADI, ranging from 89.1 (63.6) in the second‐highest quartile to 106.6 (87.7) in the highest (most disadvantaged) quartile. This may reflect a concentration of poverty and medical oncologists in urban areas. Turning to rurality, mean SPARs were highest in urban census tracts, 105.2 (78.7), and lowest in rural ones 61.0 (94.5). This pattern of lower geographic access in rural census tracts was consistent within ADI quartiles one and two with slight deviations from the pattern in quartiles three and four.

TABLE 2.

Mean spatial access ratios of geographic access to medical oncologists by rurality and area deprivation index

Spatial access ratio (Mean [SD])		Rural‐urban commuting area group
	Overall	Urban	Large Town	Rural
Overall	100 (92.6)	105.2 (78.7)	68.7 (174.7)	61.0 (94.5)
Area deprivation index quartiles (1 = Highest; 4 = Lowest)
First quartile	104.1 (72.0)	104.4 (72.1)	64.2 (4.5)	37.0 (−)
Second quartile	89.1 (63.6)	92.0 (64.1)	42.9 (11.6)	29.5 (21.0)
Third quartile	96.6 (128.9)	96.8 (82.2)	109.0 (286.7)	70.8 (112.4)
Fourth quartile	106.6 (87.7)	122.3 (88.1)	46.6 (33.0)	59.4 (88.7)

Note: The technical appendix (Data S2) details the calculation of the enhanced two‐step floating catchment area and spatial access ratios. Spatial access ratios are relative measures of geographic access (e.g., a census tract with a SPAR of 150 implies that a census tract has 50% greater geographic access than the mean Pennsylvania census tract).

3.2. Regression results

Table 3 reports the coefficients of the spatial error model. We present the spatial model here, and the linear regression results in Table A1 because the Moran test indicates that the spatial error model is appropriate here (i.e., the test indicates spatial correlation is present among the error terms that are not accounted for by the linear model). Coefficients should be interpreted as unit changes in the SPAR values, where the mean census tract has a SPAR of 100.

TABLE 3.

Associations between spatial access ratios of medical oncologists and census tract demographics

Covariate	Estimates of associations with spatial access ratio
	Coefficient	SE	Lower CI	Upper CI	p‐Value
Rural‐urban commuting area group
Urban	—	—	—	—	—
Large town	−6.29	5.03	−16.14	3.57	0.21
Rural	−14.76	5.30	−25.14	−4.37	0.01
Area deprivation index
First quartile	—	—	—	—	—
Second quartile	−4.62	2.82	−10.14	0.90	0.10
Third quartile	−5.96	3.16	−12.16	0.24	0.06
Fourth quartile	−12.41	3.61	−19.50	−5.33	<0.01
Gender (%)
Male	—	—	—	—	—
Female	0.06	0.09	−0.12	0.24	0.54
Age (%)
0–49	—	—	—	—	—
50–64	−0.25	0.15	−0.54	0.04	0.09
65–79	−0.08	0.17	−0.41	0.24	0.62
80+	−0.17	0.17	−0.51	0.17	0.33
Race/ethnicity (%)
NH‐White	—	—	—	—	—
Nonwhite	1.10	0.07	0.96	1.23	<0.01
Unknown	−0.61	0.33	−1.25	0.03	0.06
Insurance (%)
Medicare	—	—	—	—	—
Medicaid	0.00	0.20	−0.38	0.39	0.98
Private	0.13	0.13	−0.13	0.40	0.32
Uninsured	−0.49	0.53	−1.53	0.54	0.35
Other/unknown	0.13	0.16	−0.18	0.44	0.41
Spatial error	1.26	0.02	1.22	1.30	<0.01

Note: The outcome variable is spatial access ratio (SPAR), measured at the census tract level, and indexed such that the mean census tract has a SPAR of 100. We estimate SPAR values using a spatial error model, not a linear model because Moran's test indicates that there is spatial dependence among the error terms. For gender, age, race/ethnicity, and insurance variables, coefficients represent the mean differences associated with a 1 percentage point change in their values. We used the algorithm from Kwan ³¹ to calculate distance decay weights.

Abbreviation: NH, non‐Hispanic.

Census tracts with a RUCA designation of rural were associated with SPARs that were 14.76 lower (95% CI = −25.14 to −4.37; p = 0.01) than those in urban census tracts. We did not detect a significant difference in SPARs between urban and large town census tracts. Census tracts in the fourth ADI quartile were associated with 12.41 lower SPARs (95% CI = −19.50 to −5.33; p < 0.01), relative to the first ADI quartile census tracts, and census tracts in the third ADI quartile were marginally associated with 5.96 lower SPARs (95% CI = −12.16 to 0.24; p = 0.06). A change in a census tract's non‐White population from the 25th percentile (1.3%) to the 75th percentile (13.7%) was associated with 13.64 higher SPAR values (Coefficient = 1.10, 95% CI = 11.89 to 15.29; p < 0.01).

3.3. Sensitivity analyses

We conducted several sensitivity analyses to test the robustness of our findings. First, we determined whether our baseline model overlooks key interactive relationships between our key variables of interest: rurality, ADI, and race/ethnicity. In a series of regressions that each interact two out of the three of these sets of variables shown in Tables A2A–A2C, we do not find any significant interactions between rurality, ADI, and race/ethnicity. Second, we incorporated spatial dependence in the outcomes—another common form of spatial dependence ³⁴ —in our model by estimating a spatial autoregressive model (i.e., a spatial error model with spatial lags). As shown in Table A3, the results of this model are not substantively different.

Third, we tested whether our findings were sensitive to weights we used to measure travel time costs. To do so, we used two different sets of weights that allowed travel time costs to (a) increase more quickly (Weights = {1, 0.32, 0.01}); and (b) decrease more slowly (Weights = {1, 0.59, 0.12}), relative to our baseline model (Weights = {1, 0.48, 0.05}). The resulting regression results are shown in Tables A4A and A4B. Our findings are strikingly similar in both cases, suggesting that, within reasonable bounds, are findings our robust to the selection of weights. To further explore the role that travel time plays in access disparities, we also estimated a model where patients are indifferent to traveling 0–20 min, 20–40 min, and 40–60 min (i.e., Weights = {1,1,1}). As shown in Table A4C, we find that ADI‐ and race/ethnicity‐based access disparities persist, but that rurality‐based disparities disappear. These findings suggest that rural disparities in oncology access are driven by the time, monetary, and logistical burdens of traveling to visit oncologists. Race/ethnicity‐ and ADI‐based disparities, however, are driven by the interaction of supply and demand in urban settings, where the supply of oncologists is relatively higher.

4. DISCUSSION

In this study of geographic access to medical oncologists in Pennsylvania, we found that rurality and area deprivation were the two strongest predictors of geographic access to care. Rural and deprived census tracts exhibited disparities in geographic access to oncologists relative to urban, non‐deprived census tracts. We also found that census tracts with larger non‐White populations were associated with higher geographic access, suggesting what has previously been referred to as a “reversed racial disadvantage” in geographic access. ¹

This study is consistent with prior literature in finding a negative association between rurality and geographic access to care. ¹ , ³ , ⁴ , ³⁵ Although a small number of clinical studies have used a Two‐Step Floating Catchment Area (2SFCA) method, ¹ , ³ , ¹¹ , ²⁰ few have used the enhanced 2SFCA (E2SFCA), and none have used it to study geographic access to medical oncologists. Using this approach, we simultaneously examined associations between area deprivation and demographics in addition to rurality. This innovative approach showed that neighborhood socioeconomic status is roughly as important to geographic access as rurality, a point that has not been emphasized in the previous literature.

As in previous literature, ¹ this study found evidence of a “reversed racial disadvantage” in geographic access to care. Although this “reversed racial disadvantage” was statistically significant, it is statistically indistinguishable from the negative association between geographic access and census tracts in the lowest ADI quartile (i.e., the most socioeconomically disadvantaged), where non‐White populations are concentrated. Specifically, a census tract that is in the fourth ADI quartile is associated with a 12.41 lower SPAR value (relative to being in the first ADI quartile). On the other hand, a census tract that is in the 75th percentile of the percentage of the population that is non‐White, relative to the 25th percentile, is associated with 13.64 higher SPAR values. These findings are consistent with Xu et al., who found that geographic access to NCI cancer centers was higher for non‐White populations but also lower for populations under the poverty line. ¹ Thus, while a “reversed racial disadvantage” in geographic access does exist, its benefits are counteracted by the negative association between socioeconomic status and geographic access to oncologists that disproportionately affects neighborhoods with higher shares of non‐White residents. We note that, while improving geographic access to care is a necessary condition for persons with low socioeconomic status to access care, it is likely, not sufficient due to non‐geographic barriers to care (e.g., affordability, accommodation, etc.) ¹⁵ It also is important to note that, while rural areas in Pennsylvania have relatively small non‐White populations, this is not always the case in other areas of the United States, notably the south. ³⁶ Therefore, “reversed racial disadvantage” may not be present in all regions.

The results of this study have important policy implications. Previous county‐based analysis has suggested that a focus on rural areas should be prioritized. ³⁷ However, the census‐tract level analysis of this study revealed that both rurality and low area socioeconomic status were associated with low geographic access to medical oncologists, revealing a complexity that has been overlooked in previous literature. Efforts aimed at improving access to care should account for the complex determinants of geographic access. For example, under the Pennsylvania Rural Health Model, some rural hospitals are being paid under global budgets to cover services in the hopes of increasing rural Pennsylvanians' access to high‐quality care. ³⁸ The results of this study imply that this model and similar efforts targeted towards maintaining or increasing geographic access to care should be cognizant not only of an area's rurality but also of its socioeconomic status.

This study is subject to limitations. First, the study was focused solely on Pennsylvania, which has a small proportion of non‐Black minority populations (7.4%), limiting the racial analysis to the White population relative to the non‐White population. ¹⁸ Second, like many spatial analyses, this study is subject to the “edge effect.” That is, it did not capture the availability of medical oncologists or the presence of cancer patients in bordering states. Given available data sources, we are unable to directly address this limitation, though we expect that most Pennsylvania residents receive some if not all cancer‐focused care within the state and do not anticipate that out‐of‐state provider availability would substantially change our findings. Third, this study did not consider other dimensions of access to care such as racial stigma, language barriers, transportation barriers, and affordability. These non‐geographic factors can affect whether patients seek care alongside geographic access, ¹⁵ , ³⁹ , ⁴⁰ , ⁴¹ but they are beyond the scope of this study. Fourth, it is possible that some medical oncologists were excluded from this study. However, this is likely a minor limitation because the Base Provider Enrollment File records all Medicare‐accepting physicians—93% of non‐pediatric primary care physicians accept Medicare. ⁴² Lastly, incident cancer cases are not a perfect measure of demand for care but a reasonable proxy that captures the relative cancer burden in an area.

5. CONCLUSIONS

This study uses the E2SFCA approach to identify disparities in geographic access to medical oncologists by rurality, socioeconomic status, and demographic characteristics. The results point to two key findings that can inform efforts to improve access to care for underserved cancer patient populations. First, rural areas and those with low socioeconomic status are associated with lower geographic access to medical oncologists. Interventions to improve geographic access to medical oncologists should therefore focus on socioeconomically disadvantaged urban areas as well as rural areas. Second, while the results indicated that “reversed racial disadvantage” exists in geographic access to medical oncologists, it is counteracted by the decreased geographic access experienced by Pennsylvanians living in socioeconomically disadvantaged areas, where many non‐White residents tend to be concentrated. Future research should continue to examine how the distribution of oncology services results in poor geographic access to care for socioeconomically disadvantaged, minority neighborhoods.

CONFLICT OF INTEREST

The authors declare no conflicts of interest.

Supporting information

Data S1. Supporting information.

Data S2. Technical appendix.

Muluk S, Sabik L, Chen Q, Jacobs B, Sun Z, Drake C. Disparities in geographic access to medical oncologists. Health Serv Res. 2022;57(5):1035‐1044. doi: 10.1111/1475-6773.13991

DATA AVAILABILITY STATEMENT

Cancer registry data used in this study were supplied by the Bureau of Health Statistics and Registries, Pennsylvania Department of Health, Harrisburg, Pennsylvania. The Pennsylvania Department of Health specifically disclaims responsibility for any analyses, interpretations, or conclusions. These data are not available for public use.