The impact of patient travel time on disparities in treatment for early stage lung cancer in California

Background Travel time to treatment facilities may impede the receipt of guideline-concordant treatment (GCT) among patients diagnosed with early-stage non-small cell lung cancer (ES-NSCLC). We investigated the relative contribution of travel time in the receipt of GCT among ES-NSCLC patients. Methods We included 22,821 ES-NSCLC patients diagnosed in California from 2006–2015. GCT was defined using the 2016 National Comprehensive Cancer Network guidelines, and delayed treatment was defined as treatment initiation >6 versus ≤6 weeks after diagnosis. Mean-centered driving and public transit times were calculated from patients’ residential block group centroid to the treatment facilities. We used logistic regression to estimate risk ratios and 95% confidence intervals (CIs) for the associations between patients’ travel time and receipt of GCT and timely treatment, overall and by race/ethnicity and neighborhood socioeconomic status (nSES). Results Overall, a 15-minute increase in travel time was associated with a decreased risk of undertreatment and delayed treatment. Compared to Whites, among Blacks, a 15-minute increase in driving time was associated with a 24% (95%CI = 8%-42%) increased risk of undertreatment, and among Filipinos, a 15-minute increase in public transit time was associated with a 27% (95%CI = 13%-42%) increased risk of delayed treatment. Compared to the highest nSES, among the lowest nSES, 15-minute increases in driving and public transit times were associated with 33% (95%CI = 16%-52%) and 27% (95%CI = 16%-39%) increases in the risk of undertreatment and delayed treatment, respectively. Conclusion The benefit of GCT observed with increased travel times may be a ‘Travel Time Paradox,’ and may vary across racial/ethnic and socioeconomic groups.


Study population
We included 23,571 patients diagnosed with first primary, stages I-II, NSCLC, as defined by the American Joint Committee of Cancer 7th edition, between 2006 and 2015, and alive at the time of diagnosis. Of these, we excluded patients due to the following reasons: missing lymph node (N) staging (n = 122) or missing date of diagnosis (n = 127), which were required to determine receipt of GCT; missing race (n = 43) or those who were classified as multiracial (n = 288) or other Hispanics (n = 9) due to race being required to assess differences by race, no validated methods to analyze multiracial categories, and a small sample size of other Hispanics; transsexual or transgender (n = 4) individuals due to small sample sizes; missing residential census block group (n = 20), missing treatment facility (n = 68), or requiring a ferry for transit/driving time incalculable (n = 3), which were required to determine travel times; driving distance >250 miles (n = 66), which were outliers for travel times. After applying these exclusions, the final study population comprised 22,821 patients.

Assessment of GCT
The primary outcome was receipt of GCT according to the 2016 National Comprehensive Cancer Network (NCCN) guidelines defined as the administration of proper initial and adjuvant surgery, chemotherapy, or radiation treatment(s) according to cancer site and stage. If a patient did not receive surgery, they were assumed inoperable and assessed for GCT according to lymph node staging (N0 or N1). Alternatively, undertreatment was less than minimum siteand stage-specific recommended treatment.
The secondary outcome was receipt of timely (versus delayed) GCT. The Research ANd Development Corporation recommends treatment initiation within 6 weeks of diagnosis [54] (i.e., the initiation of surgery, radiation, or chemotherapy within 45 days of diagnosis), and The Commission of Cancer Quality of Care Measures recommends adjuvant treatment of chemotherapy administration within 6 months of surgery, when required [55] (i.e., the initiation of chemotherapy +/-radiation within 6 months of initial surgery for N1 patients); Fig 1. To determine receipt of GCT and timely treatment, full dates for diagnosis, surgery, radiation, and chemotherapy are required. If only month and year were available, the middle of the month day was imputed.

Assessment of travel time
Mean-centered travel time [56,57] to treatment facilities including driving and public transit travel times (minutes) to a patient's chosen treatment facility from their residence was calculated from the centroid of their census block group [58]. ArcGIS Online's Connect Origins to Destinations Analysis was used to compute driving travel time based on historical and live traffic data [59]. Public transportation was calculable for 11,607 patients living in census blocks with transit service available (nearest transit stop within 0.75 miles). The Google Maps Application Programming Interface with the gmapsdistance function in R was used to compute public transit travel time; gmapsdistance requires a future travel time and was specified as an arrival date and time of Monday, October 9 th , 2020 at 5pm; 5pm was chosen to account for less traffic during the COVID-19 pandemic. Driving time was also calculated using gmapsdistance with the same specifications to compare the two methods of calculating driving travel time.

Effect modifiers
Patient race/ethnicity and nSES were investigated as potential effect modifiers of the association between travel time and receipt of GCT. Race/ethnicity was classified as non-Hispanic White (NHW), non-Hispanic Black (NHB), Hispanic (including those who identify as white or Black), Native Hawaiian and Pacific Islander (NHPI), Chinese, Japanese, Filipino, Korean, Vietnamese, Asian Indian, Other Asian, or American Indian. Race/ethnicity data in the CCR is based on hospital records that use self-report, assumptions of hospital personnel, or extrapolation from birthplace, race/ethnicity of parent, maiden name, or surname [60]. nSES in the CCR is determined from the American Community Survey using a composite residential neighborhood-level index that combines census measures of education, income, occupation, and cost of living at the census block group level and categorized into quintiles [61].

PLOS ONE
The impact of patient travel time on disparities in treatment for early stage lung cancer in California marital status (single/never married, partnered (married/unmarried or domestic partner), unpartnered (separated/divorced, widowed)), whether or not the reporting facility with the earliest date of admission had an ACOS-approved cancer program, and rural-urban continuum codes. Rural-urban continuum codes (1-9) distinguishes metropolitan counties by the population of their metro area, and nonmetropolitan counties by degree of urbanization and adjacency to a metro area are assigned to each county [62]. To resolve unavailability of payer (n = 298), marital status (n = 564), and cancer program (n = 46) information, we used multiple imputation, a valid statistical procedure for recovering missing data to create complete datasets that can then be analyzed through standard procedures [63].

Statistical analysis
Exposure, clinical and sociodemographic information were stratified by race/ethnicity. We quantified average disproportionality in receipt of GCT and timely treatment across categories of race/ethnicity, nSES, and driving and public transit travel times (<15, 15-30, 30-60, and �60 minutes) using three disproportionality functions: Between-Groups Variance (BGV), The Theil Index (T), and Mean Log Deviation (MLD). BGV is a useful metric of absolute disparity for unordered groups, such as race/ethnicity, because it weights by population size and is sensitive to larger deviations from the population average. T and MLD are entropy-based measures that quantify the relative disparity, meaning the disproportionate receipt of GCT and timely GCT across effect modifiers and exposures. T and MLD are complementary measures because T can be influenced by groups with high ratios of GCT and timely GCT in a group relative to the average GCT and timely GCT in the population, and MLD can be influenced by groups with larger population shares [64]; formulas provided in Fig 2. We used multivariable generalized logistic regression models (PROC GENMOD) with a Poisson distribution and log link function to explore all combinations of the following associations: outcomes (undertreatment and delayed GCT), exposures (mean-centered driving and public transit travel time), and effect modifiers (race/ethnicity and nSES), to estimate the impact of travel time to treatment facilities on both racial/ethnic and socioeconomic disparities in undertreatment and delayed GCT. The intraclass correlation coefficient (ICC) of treatment hospital was assessed to determine if treatment hospital needed to be included as a random effect. Driving and public transit travel times were mean-centered and scaled to represent a 15-minute increase from the population average. Patient racial/ethnic groups with less than 100 persons (NHPI, Asian Indian, and American Indian) were excluded from models due to small sample sizes. In addition to disaggregating Asian groups with sufficient sample sizes, an aggregated Asian American, Native Hawaiian, and Pacific Islander (AANHPI) models was run separately including NHPIs and Asian Indians. Overall, we had 28 covariate-adjusted models. Models 1,8,15, and 22 regressed the outcomes (undertreatment and delayed GCT) on the effect modifiers (race/ethnicity and nSES). Models 2,5,9,12,16,23, and 26 regressed the outcomes (undertreatment and delayed GCT) on the exposures (driving and public transit time). Models 3, 6, 10, 13, 17, 20, 24, and 27 combined the above models. Models 4,7,11,14,18,21,25, and 28 extended the previous models by adding an interaction term between the effect modifiers (race/ethnicity and nSES), and the exposures (driving and public transit time). The interaction models were the primary models of interest. nSES was not adjusted for when considering race/ethnicity as an effect modifier, but race/ethnicity was adjusted for when considering nSES as an effect modifier. Risk Ratios (RR) and 95% Confidence Intervals (CI) for the effect measure modifier analyses are presented in Table 4, while the betas and 95% CIs for all 28 models are available in (S1 Table, effect modifier: race/ethnicity; S2 Table, effect modifier: nSES). A sensitivity analysis considering driving time calculated using the gmapsdistance function were compared to the above results. All analyses were performed using SAS version 9.4 (SAS Institute Inc., Cary, NC).

Clinical and sociodemographic characteristics
Stage at diagnosis varied by race/ethnicity with NHBs having the highest proportion of Stage IIB diagnosis (14.9%). Females accounted for 54.5% of patients overall, but 64.4% of Japanese and 40.6% Vietnamese patients. The mean age at diagnosis was 70.4 years overall and ranged from 67.1 years for NHPI to 74.2 years for Japanese patients. Less than 1% of patients were uninsured, and half were married or in a domestic partnership. Most patients were treated at hospitals with an ACOS-approved cancer program (60.5%) with lower rates among NHBs (51.3%), NHPIs (51.5%), and Chinese (52.3%). nSES differed by race/ethnicity; overall, 14.2% of patients lived in the lowest nSES, but NHBs (29.7%), Hispanics (26.5%), and NHPIs (19.7%) proportions were much higher, and most patients lived in metro areas.

Payer
Not insured

PLOS ONE
The impact of patient travel time on disparities in treatment for early stage lung cancer in California and public transit times, stratified by nSES, are provided in (S3 Table). Patients with the highest nSES have the shortest travel times.

Absolute and relative disparity measures
The proportions of receipt of GCT ranged from 76.35% among NHBs to 84.70% among Chinese and the proportions of receipt of timely treatment ranged from 49.80% among Filipinos to 72.06% among Other Asians. Patient's living in the highest nSES had the highest proportion of GCT (84.53%) and timely treatment (66.25%), followed by upper-middle, middle, lowermiddle, and lowest SES (GCT = 75.33%; timely GCT = 50.43%) nSES (Table 2). Patients with a �60 minutes driving time had the highest percent GCT (86.90%) and timely treatment (64.95%), followed by 30-60, 15-30, and <15 minutes (GCT = 77.36%; timely treatment = 56.29%). Patients with a �60 minutes public transit time had the highest proportion of GCT (82.33%) and timely GCT (58.65%) ( Table 3). BVG, Theil, and MLD values range from 0 to 1 (higher inequality) and should be used to compare the level of inequality across outcomes and groups. We observed more absolute disparity in rate of timely GCT, compared to GCT, between race/ethnicity (GCT = 3.65; timely GCT = 8.65) and nSES (GCT = 10.10; timely GCT = 28.35), with higher absolute disparity in nSES compared to race/ethnicity. There was more absolute disparity in GCT (driving = 10.73; public transit = 8.60) compared to timely GCT (driving = 5.65; public transit = 2.18), between travel times. There was very little relative disparity in rate of GCT and timely GCT.
To explain how driving and public transit time impacted the risk of undertreatment and delayed GCT, multivariable mean-centered models are described below. Treatment hospital had an intraclass correlation coefficient of < 5% and therefore was not included as a random effect.

Discussion
Racial/ethnic and socioeconomic disparities in receipt of GCT and timely treatment exist among early-stage NSCLC patients in California. NHBs experienced the lowest rate of GCT and Filipinos and NHBs experienced the lowest rates of timely treatment, and patients living in the highest nSES experienced the highest rate of timely GCT with a linear decreasing trend with decreasing nSES. On average, a 15-minute increase in travel time was associated with a decreased risk for undertreatment and delayed treatment. This protective effect observed from increased travel times was unexpected and may be a 'Travel Time Paradox," but this paradox was not uniform across all groups.
NHBs and Hispanics were at higher relative risk as compared to Whites for undertreatment and delayed treatment. NHBs and Hispanics had shorter travel times and the highest proportions of patients in lower nSES. Interestingly, when considering the interaction between travel time and race/ethnicity, a 15-minute increase in driving time for Hispanics attenuated the risk of undertreatment and delayed treatment, compared to NHWs. This could be explained by healthcare facilities near Hispanic neighborhoods being poorer. Opposing, a 15-minute increase in public transit time for Hispanics increased the magnitude of risk of undertreatment and delayed treatment, compared to NHWs. It is unclear why this "Travel Time Paradox" would not hold in Hispanics for public transit, but it may be that patients requiring public transit are less likely to travel farther for better care when travel times are already three times longer than driving. Further, a 15-minute increase in driving and public transit time for NHBs increased risk of undertreatment and delayed treatment, compared to NHWs. This supports a racial/ethnic disparity that is not overcome by a farther, more qualified, healthcare facility.
In aggregate, AANHPIs were not at increased relative risk for undertreatment or delayed treatment, however, by disaggregated Asian groups important heterogeneity was illuminated. Compared to NHWs, Koreans and Vietnamese were at higher risk for undertreatment and Filipinos were at higher risk for delayed treatment. Filipinos and Vietnamese had shorter travel times and relatively average nSES. For Vietnamese, however, a 15-minute increase in driving time for Vietnamese appears to protect against undertreatment compared to NHWs and reveals the benefit for Vietnamese to travel farther for better cancer care. On the other hand, a 15-minute increase in public transit time for Vietnamese increases the risk of undertreatment, compared to NHWs. A 15-minute increase in driving and public transit time for Filipinos increases the risk of undertreatment and exaggerates the risk of delayed treatment, compared to NHWs. Lastly, Other Asians are at higher risk for undertreatment and lower risk for delayed treatment compared to NHWs, but a 15-minute increase in travel time significantly increases risk for undertreatment and delayed treatment, compared to NHWs.
We observed a linear relationship between increased travel time and risk of undertreatment and treatment delay by decreasing quintile of nSES. For patients in the lowest nSES, a 15-minute increase in travel time resulted in 33-39% and 27-44% increased risks of undertreatment and delayed treatment, respectively. This may be explained by lower socioeconomic patients not having as good of choices, even if traveling farther. Interestingly, a 15-minute increase in driving time for non-highest nSES patients increases the risk of delayed treatment and a 15-minute increase in public transit time for the non-highest nSES patients attenuates the risk of delayed treatment, compared to the highest nSES patients. This may be due to patients in lower nSES wanting to drive farther for better care, but it simply taking longer to find the time.
In previous U.S. studies [65][66][67], increased travel distance within urban areas decreased receipt of timely treatment, while within rural areas, the inverse relationship was found. These studies considered distance as opposed to time, which may have influenced results as driving the same distance in an urban setting likely takes longer than in a rural setting. Our public transit time results generally represent urban areas in which this 'Travel Time Paradox' holds, although attenuated compared to driving time, and contradictory to the above studies' findings. Most other U.S. studies considered assessed travel distance as opposed to travel time, and found that increased travel distance decreased likelihood of treatment [28,[35][36][37].
This "Travel Time Paradox" has not been previously reported in U.S. patients. In one Australian study, early-stage NSCLC patients living farther away were less likely to have surgery and more likely to attend a general hospital rather than a specialist hospital. But, for patients that were treated in specialist hospitals, the relationship with distance was inverse showing a protective effect with longer distance [21]. Although our study is not directly comparable due to differences in healthcare systems, our study supports the hypothesis that patients may choose, if resources allow, to travel farther for better cancer care, and the closest hospital may not have the resources to provide proper treatment. Further, two recent U.S. studies showed that early-stage NSCLC who were treated at an academic facility compared to a community facility had significantly higher median overall survival, and Black patients were more likely to undergo surgery at academic facilities [68,69]. Our study controlled for ACOS-approved cancer program to try and account for quality of care and the importance of facility type, but also found no random effect by treatment facility.
We considered a patient's chosen treatment facility as opposed to the nearest facility, as often examined [14-23, 28, 32, 34, 36, 43, 49, 50, 65-67]. Considering the nearest treatment facility may make sense in countries with universal healthcare or clearly defined catchment regions, but this topic is much more complex in the U.S. where patients' healthcare utilization is driven by insurance, choice, and convenience [70]. Thus, our observed 'Travel Time Paradox' may be driven by a patient's choice to travel further for improved cancer care.
The findings from this study should be interpreted in light of the limitations. The CCR does not provide patient refusal or comorbidities preventing treatment which could result in outcome misclassification. Further, a patient's ability to get appropriate care may be attributable to more than just proximity to care. One consideration is that wealthier patients may choose to travel farther for their cancer care than a poorer patient. We tried to unpack this by assessing nSES as an effect modifier, but due to limited sample sizes, we were unable to stratify our results by both race/ethnicity and nSES. A strength of this study includes the presentation of disaggregated Asian groups; aggregating Asians into one group masks heterogeneity between groups. Additionally, we consider a patient's chosen treatment facility, as opposed to nearest treatment facility, and so our exposure is representative of the treatment facility a patient chose to attend.
These findings help elucidate the cancer-related health disparities within California's highly diverse population. Undertreatment and delayed treatment for early-stage NSCLC disproportionately affect minorities and those living in lower socioeconomic status neighborhoods. The protective effect observed from increased travel times may be a 'Travel Time Paradox'. This paradox effect may be partially explained by patients choosing to travel farther for better care or having to travel farther to receive treatment. However, a patient's ability to travel farther for care could be prohibited for many reasons such as lack of time or personal transportation thus additional healthcare facilities may not be the solution. While cancer treatment transportation options may be beneficial to patients who lack a private vehicle [71,72], accessible high-quality healthcare facilities that offer surgery, radiation, and chemotherapy are required.
Supporting information S1