Linking Data From the Multiethnic Cohort Study to Medicare Data: Linkage Results and Application to Chronic Disease Research

The Multiethnic Cohort Study of Diet and Cancer (MEC), a prospective study, was established between 1993 and 1996 to study risk factors for cancer by taking advantage of the ethnic and cultural diversity of Hawaii and California (1). The cohort included more than 215,000 men and women and was comprised of 5 racial/ethnic populations. At recruitment, the participants were aged 45–75 years. By 2013, more than 90% of surviving participants had become eligible for Medicare (aged ≥65 years). Given its ethnic diversity, the value of MEC for etiological research on cancer is widely recognized (2). To expand and identify health outcomes, we recently completed a study of linkage of MEC data to Centers for Medicare & Medicaid Services claim data. Medicare claims have been widely used as a population-based source of data for research (3–6). Here, we describe the success of linking MEC with Medicare data and use congestive heart failure (CHF) to illustrate its value for etiological research. The characteristics of MEC have been described previously (1). The baseline questionnaire and 4 follow-up questionnaires assessed diet, lifestyle, anthropometric factors, and familial and personal medical history. Between 1995 and 2006, blood and urine specimens were collected from approximately 70,000 participants for genetic and biomarker studies. The participants who reached age 65 years, including those who were deceased, were linked to Medicare records using Social Security number, sex, and birthdate (7). We obtained in- and outpatient hospitalization and carrier (physicians and other suppliers) claim files from 1999 through 2011. In the CHF analysis, we restricted the study to fee-for-service (FFS) participants (n = 107,031) and excluded participants (n = 18,324) who were not from the 5 MEC ethnic groups or who had missing risk factor data. CHF cases were identified from the presence of ≥1 diagnosis codes for CHF (International Classification of Diseases, Ninth Revision, code 428.x) in the Medicare Provider Analysis and Review File or ≥2 diagnosis codes in the carrier or outpatient files. The date of the first CHF claim was used as the index date. For each case, we constructed a set of at-risk individuals (alive and without a CHF diagnosis at the time of the index case's diagnosis) matched on ethnicity, sex, disability/end-stage renal disease status, birth year, year of cohort entry, study area, and length of Medicare coverage (8). We identified 16,938 risk sets (1 per case) with an average number of 24 persons per risk set. Self-reported data on body mass index, hypertension, diabetes, smoking status, physical activity, and alcohol intake were obtained from the baseline questionnaire. The associations between risk factors and CHF were estimated using proportional hazards models stratified by risk set and mutually adjusted for the above risk factors. Analyses were performed with SAS 9.3 (SAS Institute, Inc., Cary, North Carolina). Of the 215,000 MEC participants, the records of 85% were sent to the Centers for Medicare & Medicaid Services for linkage; records for the remaining 15% were not sent because those participants were not eligible for Medicare by the cutoff date of January 1, 2011 (74%), died before reaching age 65 years (20%), or had no Social Security number (6%). Of 184,299 unique identifiers, 170,766 (93%) were successfully linked to Medicare data. The characteristics of MEC participants linked to Medicare data are shown in Table 1 by Medicare enrollment status (FFS or managed care). The majority of participants (63%) had at least some coverage in the FFS portion of the Medicare program, Medicare Parts A and B (n = 107,031), while 37% had been enrolled in managed-care plans the entire time (n = 63,735). The proportion of managed-care participants was higher in California than in Hawaii and higher among African Americans and Latinos than among whites, Japanese, and Native Hawaiians. Participants with a lower educational level were more likely to enroll in managed care. No differences by sex or marital status were observed. From the Chronic Condition Summary File, we identified a large number of health outcomes among the FFS participants (see Supplementary Data, available at http://aje.oxfordjournals.org/). Participants with blood samples are also shown for these selected conditions. Characteristics of Multiethnic Cohort Study Participants Linked to Medicare Data, by Medicare Enrollment Status, 1999–2011 a Enrolled in Medicare Parts A and B (fee-for-service) for at least 1 month. b Enrolled in a managed-care plan throughout the entire time period. Characteristics of Multiethnic Cohort Study Participants Linked to Medicare Data, by Medicare Enrollment Status, 1999–2011 a Enrolled in Medicare Parts A and B (fee-for-service) for at least 1 month. b Enrolled in a managed-care plan throughout the entire time period. In the CHF analysis, the mean duration from cohort entry to the first occurrence of CHF was 10 years, and more than 72% of cases had been enrolled in Medicare for at least 5 years (Supplementary Data). Increasing body mass index, diabetes, hypertension, and smoking were associated with higher CHF risk. Alcohol intake of ≤48 g/day and increasing vigorous physical activity were associated with reduced risk (Supplementary Data). These associations were consistent across ethnic groups (Supplementary Data). Using high-quality identifiers, we successfully linked 93% of MEC participants who survived to age 65 years to the Medicare claims. Our linkage rate was comparable to the rates of other cohort studies (6, 7, 9, 10). The majority (63%) of MEC participants had traditional FFS Medicare coverage, with claims data to document their health-care use and provide information on various health outcomes. We used CHF to examine the value of Medicare data in the study of novel outcomes for which data are otherwise unobtainable in the MEC cohort. We showed that the associations of known risk factors with CHF among MEC participants were consistent with the literature (11–15). The algorithms for identifying CHF using claims data have a high positive predictive value and high specificity (>90%) (16). While algorithms are available for many chronic conditions (17, 18) and a similar approach can be applied to study them, the success of this approach for analyses will be disease-dependent. Among MEC participants, 23% in California and 15% in Hawaii were enrolled in managed care the entire time and had to be excluded from analyses because of incomplete claim histories (19). Despite the modest differences between managed-care and FFS participants, our results may not be generalizable to managed-care enrollees or to the entire MEC cohort. The FFS subcohort, however, was large and still had significant ethnic and socioeconomic diversity. Another issue regarding the analysis is left-truncation (there is a gap between baseline and Medicare data) among participants who reached age 65 years prior to 1999, the first year for which Medicare data are presently available. To minimize this issue, we can compare the results for the overall MEC cohort and only those participants who turned 65 after 1999; and for California, we could use available hospital discharge data to identify cases during the gap period. Because a large number of MEC participants have genome-wide association data, we can readily perform genetic analyses. We can also identify pre-event exposures and prediagnostic blood samples for various conditions, because event dates are available. Other potential use includes examining treatment patterns and outcomes following diagnosis (e.g., cancer survivorship studies). These possibilities extend the value of MEC baseline and follow-up data. In summary, linking MEC data to Medicare data allows identification of information on many health outcomes that was otherwise unobtainable. This new resource, together with the cohort's ethnic diversity, wealth of accumulated data on genetic, dietary, and lifestyle factors, and large biorepository, will enable MEC to make research contributions to the etiology and prevention of various chronic diseases in diverse populations. This work was supported by grant CA164973 from the National Cancer Institute. We thank Dr. Malcolm Pike for his invaluable advice on the statistical analysis. Conflict of interest: none declared.