Integrating Electronic Health Record, Cancer Registry, and Geospatial Data to Study Lung Cancer in Asian American, Native Hawaiian, and Pacific Islander Ethnic Groups

This article is featured in Highlights of This Issue, p. 1453

Asian American, Native Hawaiian, and Pacific Islander (AANHPI) females who never smoked may experience a particularly high burden of lung cancer: while an estimated 20% of all females with lung cancer in the United States (U.S.) have never smoked, previous reports have suggested that up to 70% of AANHPI females with lung cancer have never smoked (1, 2). Putative risk factors have been identified for lung cancer among those who have never smoked (i.e., passive smoking, air pollution, radon, family history of lung cancer, and cooking oil fumes) (3–5). However, the degree to which these and other suspected risk factors contribute to lung cancer among those of specific AANHPI ethnic groups who have never smoked remains largely unknown, with much of our knowledge based on studies in Asia (3–5). Moreover, in the United States, AANHPIs are individuals with considerable diversity in country of origin, socioeconomic levels, cultural beliefs and behaviors, degree of English proficiency, immigration experience, generational status, and acculturation (6–9). Proportions of AANHPIs that identify as multiple races are uniquely expanding, with increases of 60% among Asian Americans and 44% among Native Hawaiians and Pacific Islanders from 2000 through 2010 (10, 11). Despite this substantial diversity, research has mostly considered AANHPIs as one aggregate group, which masks health inequities that exist across ethnicities. Disaggregation of this population in recent studies shows substantial health and exposure disparities (7, 8, 12) that reinforce the importance of considering specific AANHPI ethnicities in research (13, 14).

The study of lung cancer among AANHPI females who have never smoked has been hindered by the lack of a sufficiently-large and representative data source to document population-level incidence rates of lung cancer stratified by sex, detailed race/ethnicity, and smoking status (10, 15). Integrative data analysis, which combines data from multiple sources to enrich the number of observations or data on explanatory variables, takes advantage of existing resources (e.g., health care data) to examine rare diseases among “small populations” (16, 17) of interest, like AANHPI ethnic groups (16–22).

We assembled a large-scale, multilevel electronic health record (EHR)-based cohort to facilitate research on the incidence and etiology of lung cancer among specific AANHPI ethnic groups. The cohort comprises males and females of any race/ethnicity but is designed specifically to quantify the burden of lung cancer among detailed single- and multi-racial/ethnic AANHPI groups. The dataset contains EHR data from two large healthcare systems linked to (i) their respective statewide cancer registries and (ii) geospatial data. The healthcare systems—Sutter Health in northern California and Kaiser Permanente Hawai'i—were specifically selected for their robust AANHPI representation and categorization, mature EHRs, and research infrastructure. We describe our process of data integration with an emphasis on methodology and the generalizability of the resulting pooled cohort, and present proportions of persons with lung cancer who never smoked for groups according to sex and detailed race/ethnicity.

All aspects of the study protocol were approved by the Institutional Review Boards (IRB) of the State of California Protection for Human Subjects; University of California, San Francisco; Sutter Health; Kaiser Permanente Hawai'i; and the Hawai'i Medical Association.

The cohort was assembled with data for Sutter Health and Kaiser Permanente Hawai'i patients with an in-person encounter any time from 2000 through 2013. Sutter Health is a not-for-profit multispecialty healthcare delivery system in northern California serving 23 northern California counties through five geographically-based medical foundations; Sutter patients make up 36% of the catchment area population (23). Kaiser Permanente Hawai'i is a not-for-profit, integrated health care delivery system with over 254,000 members and covers 18% of the underlying catchment population. Cohort inclusion and exclusion criteria are detailed in Fig. 1. The final pooled cohort includes N = 2,211,476 (by system: Sutter Health n = 1,871,175, Kaiser Permanente Hawai'i n = 340,301). The date of the first in-person encounter was considered the date of cohort entry (baseline date), with follow-up through December 31, 2013 or date of death. Median follow-up was 4.8 years with a maximum of 14 years.

All Sutter Health patients were previously linked to the California Cancer Registry for all invasive cancers diagnosed from 1988 to 2013 (23). The Kaiser Permanente Hawai'i cohort was linked to the Hawai'i Tumor Registry for lung cancers diagnosed from 1973 to 2013. For all lung cancer cases, cancer registry data included date of diagnosis, tumor stage (localized, regional, remote), and tumor histology (small cell lung carcinoma, adenocarcinoma, squamous cell carcinoma, large cell or other cell carcinoma, and unspecified) (24). Incident lung cancer (n = 7,274) was defined as a diagnosis of invasive lung or bronchus carcinoma (International Classification of Disease for Oncology, third edition, site codes C34.0–34.9; excluding histologic codes: 8580–9999 and 8500) occurring during the study period.

Both Sutter Health and Kaiser Permanente Hawai'i use EHR systems designed by Epic (Epic Systems Corporation, Verona, WI). Sutter Health's EpicCare EHR system was in three of five Sutter Health foundations by 2000 and in all by 2008. Kaiser Permanente Hawai'i EHR, KP HealthConnect (KPHC), was fully implemented in 2004; however, inpatient diagnoses and procedures have been captured electronically since 1985 and pharmacy, laboratory tests, and claims since 1995.

For both Sutter Health and Kaiser Permanente Hawai'i, patients' race and ethnicity are collected through self-report either in person or online and allow patients to indicate more than one race; only the most recently reported data is retained. We developed a race/ethnicity classification algorithm (Fig. 2) that prioritizes small AANHPI populations and distinguishes between single- and multi-racial/ethnic groups. Individuals indicating any Pacific Islander group (alone or in combination with other racial groups) were classified as Native Hawaiians and Pacific Islanders (NHPI) and were further disaggregated into categories of any Native Hawaiian (NH) or any Pacific Islander (PI) not including NH. Individuals who were not NHPI, but indicating any Asian American group were categorized as Asian American, even if they also indicated another race group. Among Asian Americans, individuals who indicated a single Asian group were classified as solely: Asian Indian, Chinese, Filipinx, Japanese, Korean, or Vietnamese, or Other single Asian group. Individuals indicating multiple races that included an Asian group were categorized as either multiple Asian groups only (Asian, multiple group) or as Asian and non-Asian groups (Asian and non-Asian multiple). Individuals not categorized as AANHPI were categorized as non-Hispanic White, Black, Hispanic, multiple (non-AANHPI) races, Other (including American Indian and Alaska Native), and Unknown/missing. With this algorithm, the AANHPI and Black categories included Hispanics and non-Hispanics; the Hispanic category included individuals who indicated Hispanic ethnicity and White race or ‘Other’ race (including American Indian and Alaska Native) as well as those with unknown/missing race.

Patient smoking status is routinely collected and recorded in the social history module of the EHR, with mutually exclusive categories of current, former, passive, never, unknown/missing, and not-asked. The EHR retains all answers and corresponding collection dates. For each individual, we extracted up to two non-missing smoking status values (current, former, passive, never): (i) the first available value recorded on the day of or after the baseline visit and (ii) the last available recorded before date of lung cancer diagnosis, death, or study end (December 31, 2013). For individuals with discordant entries at these two time points, we applied a simple algorithm to define a single smoking status value (ever, never) (Supplementary Table S1A–S1C). Those with “current” or “former” for either smoking status value were categorized as ever smoking. Those with “never” or “passive” for both status values, or with “never” or “passive” for one value and a non-informative value for the other were categorized as never smoking. Because “passive” smoking was only available as a mutually exclusive category of smoking and only 1.0% of the cohort had a status of “passive” smoking, we did not utilize this categorization further. N = 2,118,945 (95.5%) individuals in the pooled cohort were assigned one of these smoking status categories, the remainder were coded as “unknown” (Supplementary Table S1A).

At both sites, self-reported language preference was collected via free text and patients reported whether a translator was requested (no/yes; ref. 25); only the most recently reported version of this information was retained. We derived a 3-category ordinal variable to serve as a proxy for individual acculturation (English preference, non-English preference/no translator request, and non-English preference/translator request; ref. 26).

To study lung cancer risk factors specifically among AANHPI females who never smoked, additional EHR data were extracted (Fig. 1). Data on medical history, diagnostic codes, and medication orders recorded from the baseline date to the end of follow-up (December 31, 2013) were extracted to classify exposure to previous lung diseases (chronic obstructive pulmonary disease, emphysema, chronic bronchitis, asthma, and idiopathic pulmonary fibrosis), infections [pneumonia, tuberculosis, chlamydia, human papillomavirus (HPV), and human immunodeficiency virus (HIV)], reproductive factors, and body mass index. Details of these disease definitions are presented in Supplementary Table S2.

For the Sutter Health cohort, individuals with a California residential address available at the baseline visit were geocoded with baseline latitude/longitude coordinates. For N = 1,873,650 total in-state addresses (representing addresses for Sutter patients before cohort exclusions were applied), batch geocoding was performed with ArcGIS (27) to a point address (n = 1,269,578) or street locators (n = 436,487). Manual geocoding resulted in another n = 20,216 addresses matched to point address or street locators. A total of 107,248 (5.7%) were determined to be not geocodable (e.g., post-office boxes). This resulted in a total of n = 1,726,281 successfully geocoded addresses (97.7% of geocodable addresses). Baseline addresses from 2000 to 2005 were assigned to Census 2000 tracts and those from 2006 to 2013 were assigned to Census 2010 tracts. After cohort exclusions were also applied (Fig. 1), there were n = 1,721,000 individuals from Sutter Health assigned to a baseline census tract of residence.

Kaiser Permanente Hawai'i patient addresses are geocoded centrally using ArcGIS; census tract identifiers for baseline addresses were provided for the pooled cohort dataset. After excluding n = 3,228 with addresses outside of Hawai'i, n = 35,344 with missing addresses, and cohort exclusions; there were n = 301,729 individuals in the Kaiser Permanente Hawai'i cohort with a 2010 Census tract assigned in Hawai'i. The final pooled cohort comprised n = 2,022,729 individuals assigned to a census tract.

We created census tract-level maps of the geographic distribution of the cohort with individuals' baseline addresses. For each state/site, ArcGIS software (27) was used to map frequency quintiles for the total cohort population and AANHPI female cohort population. To display a single set of maps, 2010 Census tract boundaries were used for all individuals.

Individuals' baseline geocoded addresses (addresses available at the time of their first in-person encounter with the healthcare institution during the study period) were linked via census tracts to area-based measures [neighborhood socioeconomic status (nSES), ethnic enclave, ethnic composition, and traffic density] (28–30). nSES, a composite index developed previously from principal components analysis, incorporates area-level information on education, occupation, employment, household income, poverty, housing value, and rental value from the Census 2000 Summary Files, and American Community Survey (ACS) data from 2007 to 2011 data (28, 29). Each census tract is assigned the appropriate composite score for nSES and then categorized into quintiles based on the statewide distribution of scores across all census tracts separately for each decennial year (2000,2010). Ethnic enclave is a composite measure of proportions of Asian, recent immigrants and households with limited English or that are linguistically isolated and speak an API language. Statewide quintiles are used to identify ethnically distinct enclaves (quintiles 4 and 5). Traffic density is based on the volume of traffic on major roadways within each tract (30, 31).

Additionally, Sutter Health patients were assigned estimates of locally distributed air pollutant PM_2.5 according to baseline geocoded addresses. Monthly ambient data on PM_2.5 collected from air monitoring stations from 2000 to 2013 were spatially interpolated between stations using an empirical Bayesian kriging model implemented in ArcGIS to account for the uncertainty of semivariogram estimation (32). For everyone in the Sutter Health cohort, monthly exposure estimates relevant to the latitude/longitude coordinates of their geocoded baseline address were averaged across one year (12 months) according to the baseline date to derive the one-year average estimate of exposure at baseline. Because PM_2.5 does not capture the extent of air pollution in the Hawaiian Islands due to volcanic smog and fog (“vog”), we limited our PM data collection to the California sample.

We used descriptive statistics (frequencies and percentages) to describe the cohort and lung cancer case population overall, and by age, sex, detailed race/ethnicity, and nSES. For assessing the generalizability of the cohort to the source population, we calculated and plotted standardized mean differences using comparison estimates of background prevalence extracted from the California Health Interview Survey [CHIS (2003–2009 and 2011–2013; ref 33], the Behavioral Risk Factor Surveillance System in California and Hawai'i (2000–2013; ref 34), and the decennial Census (2010; ref 35). Generalizability of lung cancer cases was assessed with covariate balance plots comparing with statewide and national (SEER-18) cancer registry data (36, 37).

Our schema for the categorization of race/ethnicity (Fig. 2) resulted in the disaggregation of the AANHPI group into ten single- and multi-racial/ethnic AANHPI ethnic groups. The full pooled cohort comprised 2,211,476 individuals (1,275,838 females and 935,638 males) including 49,983 Native Hawaiians, 31,506 Pacific Islanders, and 352,076 Asian Americans, representing 2.26%, 1.42%, and 15.92% of the overall cohort, respectively. There were 7,274 incident lung cancer diagnoses (3,867 females and 3,407 males) in the study period (2000–2013). These included 1,228 lung cancers among AANHPI, including 328 Native Hawaiians, 81 Pacific Islanders, and 819 Asian Americans (Tables 1 and 2).

The cohort included 889,870 females who have never smoked (69.7% of females) (Table 1). Among AANHPI, 198,208 females (79.3%) never smoked, with Asian Indian females having the highest proportion (94.7%) and Native Hawaiian females the lowest proportion (54.8%) across all racial/ethnic groups. Among females with lung cancer, the prevalence of never-smoking ranged from a low of 11.2% among non-AANHPI multiple races/ethnicities and 14.4% among Native Hawaiians to a high of 85.7% among Asian Indian and 78.7% among Chinese females (Table 1).

Among males, 545,543 (58.3%) never smoked (Table 2). Among AANHPIs, 115,573 males (63.0%) never smoked. Among males with lung cancer, the prevalence of never-smoking ranged from a low of 6.0% among Native Hawaiian to a high of 41.1% among Chinese males (Table 2).

Supplementary Tables 3A and 3B provide cohort characteristics by joint language use/translator request. Supplementary Table S4 provides the distribution of age and detailed race/ethnicity according to nSES among the cohort.

Our generalizability analysis (Fig. 3; Supplementary Fig. S1, and Supplementary Tables S5–S7) compares the pooled cohort with the demographics of the U.S. (Fig. 3), and Sutter Health and Kaiser Permanente Hawai'i cohorts with the demographics of California and Hawai'i, respectively (Supplementary Fig. S1). Compared with the U.S. population, the pooled cohort overrepresents the NHPI population and underrepresents non-Hispanic White, Black, and Hispanic populations. The cohort overrepresents females in California, and cohort members are slightly younger on average at baseline compared with the selected representative surveys from the states of California and Hawai'i, especially for Native Hawaiian females. In the Sutter Health portion of the cohort, AANHPI proportions are generally comparable with the underlying state population, with the Sutter female population slightly overrepresenting Asian Indian (+13%), and underrepresenting Filipinx (-10%), Korean (-9%), and Vietnamese (-13%) females. The Kaiser Permanente Hawai'i female cohort slightly underrepresents NHPI (by -12%) compared to the state of Hawai'i. While not markedly different, the cohort also includes a slightly higher proportion of persons who have never smoked than the underlying population for most groups, especially among males, older adults, and the Kaiser Permanente Hawai'i cohort. Case demographics reveal a marked underrepresentation of non-Hispanic White lung cancer cases in California (-69% compared to the California Cancer Registry) but also includes enrichment for AANHPI cases (+38% compared to SEER-18). An important caveat for these findings is that our deliberate prioritization of smaller AANHPI groups in classifying cohort members' race/ethnicity differs from the Census classifications, which reflect counts of individuals identifying with “any” racial or ethnic group and thus may result in apparent over- or under-representativeness due to noncomparable classification schemes.

The geographic distribution of the cohort in California and Hawai'i are shown in Fig. 4. The Sutter Health cohort, especially AANHPI females, is more concentrated in the San Francisco Bay Area and Sacramento regions of California, but has baseline residences that extend across California. Even though Kaiser Permanente Hawai'i is based on Oahu, this cohort is distributed evenly across the major islands of the state. The case population in the pooled cohort is distributed toward higher nSESs, compared with the distribution of lung cancer cases in the nation (Supplementary Tables S7A and S7B).

We have described a large-scale, EHR-based cohort inclusive of males and females of any race/ethnicity, but notably sufficient in scale to study lung cancer incidence and etiology among detailed AANHPI ethnic groups (1). By utilizing prospectively collected EHR data and integrative data analysis, this multilevel dataset includes routine clinical data from two EHRs, data on lung cancer diagnoses from state cancer registries, geospatial data on built and social neighborhood environments, and air pollution data. The study population is enriched for AANHPI groups in the U.S. while being generally representative of the target populations of California and Hawai'i. We have shown that the proportion of lung cancer diagnoses among persons who never smoked varies widely across specific AANHPI groups, with more than three-quarters of Chinese, Asian Indian, and Vietnamese female lung cancer diagnoses occurring among those who have never smoked. With ongoing work, we will estimate smoking-specific incidence of lung cancer, including incidence of histologic cell-types, according to sex and detailed race/ethnicity. With multilevel data on important known and putative risk factors, we will also leverage this linked dataset to examine multilevel etiology of lung cancer among AANHPI females who have never smoked. More broadly, this cohort and dataset illustrate a methodological approach to effectively study cancer outcomes among small populations. Our approach highlights many benefits of an integrative data analysis approach including (i) data pooling to enrich small group representation, (ii) population-based registry linkages for outcome ascertainment, and (iii) integration of multilevel data, each described in more detail below (38, 39).

Previous cancer studies in the U.S. largely considered AANHPI populations in aggregate, although smoking prevalence and proportions of lung cancer among AANHPI who have never smoked has been shown to vary widely (16). Our inclusion of multiple healthcare systems with high representation of AANHPI groups and with detailed race categories, Sutter Health and Kaiser Permanente Hawai'i, provided robust representation of AANHPI single- and multi-racial/ethnic groups. This allows us to study cancer among AANHPI groups in a way that has not been possible with prior U.S. studies and, more broadly, highlights the potential for use of EHRs to study cancer outcomes among small populations.

However, an EHR-based cohort may not be generalizable to the target population and, as a result, studies using this type of resource may be limited in external validity (23, 40, 41). We aimed to address this concern with our analysis of the representativeness of the cohort, which indicated that the cohort was fairly generalizable. Representativeness, especially important when studying disease occurrence and risk factors across heterogeneous populations, is often logistically infeasible in longitudinal cohort studies (38, 42). Generalizability analyses such as ours can help to establish “target validity” for EHR-based research and improve confidence in study results and subsequent interventions or policy recommendations (43, 44).

Relying solely on EHR data for identification of incident cancer cases is problematic due to the low sensitivity of EHR-based algorithms for identifying cancer patients and the frequent migration of patients in and out of healthcare systems (23, 45, 46). Therefore we leveraged a previous linkage of Sutter Health patients to the California Cancer Registry (46) and linked the Kaiser Permanente Hawai'i patient population to the Hawai'i Tumor Registry, which allowed for higher quality, adjudicated cancer outcome ascertainment. The state cancer registries have near complete capture of cancer diagnoses in their respective states, so only individuals who moved out of the state would be lost to follow-up. To assure even greater follow-up for outcome, we required at least one record of an in-state address (California for Sutter Health and Hawai'i for Kaiser Permanente Hawai'i) during the study period. Through these registry linkages, the cohort dataset can be expanded with additional registry data to study other cancer outcomes (e.g., treatment or survival).

In recognition of the multilevel factors that shape cancer risk, this cohort has been geocoded to enable inclusion of data on geospatial risk factors. Current data includes baseline residential neighborhood defined as census tracts, which can be used to study neighborhood built and social environments (47–52). Linkage of the Sutter Health dataset to regional monthly air pollution estimates facilitates studies of air pollution exposure and cancer outcomes. Our use of a single address at baseline does limit the understanding of cumulative effect of geospatial exposures over time since we do not know how long individuals' have resided at this address. Geocoding of longitudinal addresses over follow-up is possible with EHR datasets (assuming the individual remains associated with the healthcare system), though not feasible within the scope of the present study. Our use of data linkages will allow us to investigate the relative importance of many clinical and geospatial lung cancer risk factors, but these data can be limited for addressing the contribution of exposures not captured as part of routine healthcare delivery like nutrition, occupational exposures, individual socioeconomic status, and cultural factors. However, contemporary efforts to support routine capture of information on social determinants of health in the EHR will lead to increased availability of such exposures for future studies (53, 54).

Use of EHR data for epidemiologic studies has considerable application and promise. This approach to cohort development is time- and cost-efficient in comparison to traditional prospective cancer cohort creation, which involves recruitment, survey development, questionnaire administration, and follow-up and retention (21, 22, 55). Moreover, EHR-based cohorts are particularly advantageous for studying small populations and groups that are traditionally left out of epidemiologic studies, particularly as they are not subject to other common selection biases (including due to language barriers), non-response, attrition, and survival biases. Of note, too, the efficient approach we outline here can provide rationale for more costly prospective cohort studies able to generate consistent exposure data not available through EHR. However, as others have discussed, there are important considerations as EHR data are not collected for research purposes (21, 22, 55). EHR data are more limited in scope and substantial effort is required to compile and carefully curate the data while attending to the inherent biases in the use of routine healthcare data for observational research (40, 56–59). We discuss below these considerations related to (i) operationalizing routinely collected EHR data for research and (ii) patient privacy and institutional risk.

Careful planning and harmonized strategies for defining and classifying clinical data elements assured consistent data extraction across Sutter Health and Kaiser Permanente EHRs and allowed us to take advantage of the full extent of the available data. This was especially important here as a unique aspect of this project was the differing nature of the partner healthcare systems. While Kaiser Permanente is a managed care provider and has an enrolled population with very little “out of plan” use not captured in the EHR, Sutter Health is a mixed payer environment with many of its patients being able to see other providers for their care. While presenting challenges in the harmonization of pooled clinical data elements, it afforded the opportunity to directly examine heterogeneity of data and data collection methods across the healthcare systems. We developed common variable definitions and, in many cases, several alternative definitions, in consideration of the availability and structure of EHR data at the two sites. Alternative definitions allowed us to specify upfront the sensitivity analyses to assess the robustness of results and assure necessary EHR data were collected without excessive repeat extractions. Coding dictionaries specifying data elements, source, timing of extraction, and category values were provided to collaborators at both healthcare systems. Moreover, because the availability of EHR data will differ across individuals with respect to healthcare utilization, we included several metrics of healthcare utilization in our data collection scheme with which to assess these potential biases in future studies of lung cancer etiology. With EHR-based research becoming more commonplace, this work may thus serve as an example for researchers pooling across system types.

Our approach to pooling data from multiple healthcare systems carries distinct advantages for creating a sufficiently large dataset for our research purposes. It also required extensive attention to protect against data breach and adherence to use of minimum necessary protected health information (PHI) to accomplish our goals. We used data stewards for data extraction from the EHR and, except for the geocoding team, our researchers only had access to Health Insurance Portability and Accountability Act (HIPAA)-defined “limited” PHI. When direct identifiers were necessary (e.g., for geocoding) multiple strategies were employed to ensure safe handling of data and reduce institutional risk. For example, clinical data (from the EHR) was processed separately from patient geocodes; patient addresses and associated latitude/longitude coordinates were destroyed before combining geocode-derived exposure data (e.g., air pollution measures) with any clinical data. The development of these data safety protocols along with obtaining the appropriate approvals needed for data sharing is complicated and time-consuming. Researchers should be aware of the time and effort required and build this into their project timelines. However, once the dataset is developed, the data can be leveraged to study other cancer outcomes, especially those available in cancer registry data, for any of the racial/ethnic populations captured in the EHR.

This diverse, multilevel dataset will allow for much-needed research on lung cancer risk, especially among AANHPI female never-smokers, and serve as a critical evidence base to inform screening, research, and public health priorities in this growing population. More broadly, the innovative approach and methods used to develop this multilevel, integrated dataset can serve as an example of an effective way to conduct valuable cancer research assessing multilevel factors on cancer outcomes among small populations (60–62).

M.C. DeRouen reports grants from NCI during the conduct of the study. A.J. Canchola reports grants from NCI during the conduct of the study. S. Nie reports grants from NCI; and grants from National Center for Advancing Translational Sciences during the conduct of the study. D.Y. Lichtensztajn reports grants from NIH during the conduct of the study. Y.G. Daida reports grants from NIH during the conduct of the study. S. Shariff-Marco reports grants from NCI during the conduct of the study. S. Gomez reports grants from NCI during the conduct of the study. No disclosures were reported by the other authors.

M.C. DeRouen: Formal analysis, supervision, investigation, visualization, methodology, writing–original draft, project administration. C.A. Thompson: Data curation, formal analysis, supervision, investigation, visualization, methodology, writing–original draft. A.J. Canchola: Data curation, formal analysis, methodology, writing–original draft. A. Jin: Data curation, formal analysis, writing–review and editing. S. Nie: Data curation, formal analysis, writing–review and editing. C. Wong: Data curation, formal analysis, writing–review and editing. J. Jain: Data curation, methodology, writing–review and editing. D.Y. Lichtensztajn: Data curation, formal analysis, methodology, writing–review and editing. Y. Li: Data curation, methodology, writing–review and editing. L. Allen: Project administration, writing–review and editing. M.I. Patel: Conceptualization, investigation, methodology, writing–review and editing. Y.G. Daida: Data curation, supervision, investigation, writing–review and editing. H.S. Luft: Conceptualization, resources, writing–review and editing. S. Shariff-Marco: Supervision, investigation, methodology, writing–review and editing. P. Reynolds: Investigation, methodology, writing–review and editing. H.A. Wakelee: Conceptualization, investigation, methodology, writing–review and editing. S.Y. Liang: Resources, supervision, investigation, methodology, writing–review and editing. B.E. Waitzfelder: Resources, supervision, investigation, methodology, writing–review and editing. I. Cheng: Conceptualization, supervision, funding acquisition, investigation, methodology, project administration, writing–review and editing. S.L. Gomez: Conceptualization, supervision, funding acquisition, investigation, methodology, project administration, writing–review and editing.

We acknowledge Dr. Aleymayehu Midekisa, Ms. Katherine Lin, and Ms. Nam Phuong Le for analytic support. We thank Dr. Brenda Hernandez with the Hawai'i Tumor Registry for providing the distribution of neighborhood socioeconomic status among Hawai'i lung cancer cases. This work was supported by NCI (R01 CA204070; multiple PIs: S.L. Gomez and I. Cheng) and National Center for Advancing Translational Sciences (KL2TR001444; to C.A. Thompson).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.