[Skip to Content]
Sign In
Individual Sign In
Create an Account
Institutional Sign In
OpenAthens Shibboleth
[Skip to Content Landing]
Figure 1.  Survival Curves From the Emulated Target Trial of Adjuvant Fluorouracil-Based Chemotherapy in Stage II Colorectal Cancer Using Surveillance, Epidemiology, and End Results–Medicare Data From 2008 to 2013
Survival Curves From the Emulated Target Trial of Adjuvant Fluorouracil-Based Chemotherapy in Stage II Colorectal Cancer Using Surveillance, Epidemiology, and End Results–Medicare Data From 2008 to 2013
Figure 2.  Survival Curves From the Emulated Target Trial of the Addition of Erlotinib to a Regimen of Gemcitabine in Locally Advanced or Metastatic Pancreatic Cancer Using Surveillance, Epidemiology, and End Results–Medicare Data From 2007 to 2013
Survival Curves From the Emulated Target Trial of the Addition of Erlotinib to a Regimen of Gemcitabine in Locally Advanced or Metastatic Pancreatic Cancer Using Surveillance, Epidemiology, and End Results–Medicare Data From 2007 to 2013
Table 1.  Protocol of the Target Trial to Study Adjuvant Fluorouracil-Based Chemotherapy in Stage II Colorectal Cancer and Emulation Procedure Using the SEER-Medicare Database
Protocol of the Target Trial to Study Adjuvant Fluorouracil-Based Chemotherapy in Stage II Colorectal Cancer and Emulation Procedure Using the SEER-Medicare Database
Table 2.  Protocol of the Target Trial to Study the Addition of Erlotinib to a Regimen of Gemcitabine in Locally Advanced or Metastatic Pancreatic Cancer and Emulation Procedure Using the SEER-Medicare Database
Protocol of the Target Trial to Study the Addition of Erlotinib to a Regimen of Gemcitabine in Locally Advanced or Metastatic Pancreatic Cancer and Emulation Procedure Using the SEER-Medicare Database
Supplement.

eTable 1. Protocol of the Target Trial to Study Adjuvant Fluorouracil-Based Chemotherapy in Stage II Colorectal Cancer and Protocol of the Existing QUASAR Trial (2007)

eTable 2. Protocol of the Target Trial to Study the Addition of Erlotinib to a Regimen of Gemcitabine in Locally Advanced or Metastatic Pancreatic Cancer and Protocol of the Existing Trial (Moore et al. 2007)

eTable 3. Characteristics of Eligible Individuals With Stage II Colorectal Cancer Who Were Included in the Emulation of the Fluorouracil Target Trial at Baseline and the End of the Grace Period (3 Months Post-Baseline), SEER-Medicare 2008-2013

eTable 4. Comparison of Individuals in the Existing QUASAR Trial (2007) in the Emulation of the Fluorouracil Target Trial Using SEER-Medicare 2008-2013

eTable 5. Characteristics of Eligible Individuals With Locally Advanced or Metastatic Pancreatic Cancer Who Were Included in the Emulation of the Erlotinib Target Trial at Baseline and the End of the Grace Period (12 Weeks Post-Baseline), SEER-Medicare 2007-2013

eTable 6. Comparison of Individuals in the Existing Trial (Moore et al. 2007) and in the Emulation of the Erlotinib Target Trial Using SEER-Medicare 2007-2013

eFigure 1. Flowchart of Eligibility for a Target Trial of Adjuvant Fluorouracil-Based Chemotherapy in Individuals With Stage II Colorectal Cancer, SEER-Medicare 2008-2013

eFigure 2. Flowchart of Eligibility for a Target Trial of Addition of Erlotinib to Gemcitabine in Individuals With Locally Advanced or Metastatic Pancreatic Cancer, SEER-Medicare 2007-2013

eFigure 3. Illustration of the Cloning and Censoring Process for the Fluorouracil Target Trial Emulation

eAppendix 1. Codes Used to Identify Variables Used in the Analyses

eAppendix 2. Details of Statistical Analysis

eAppendix 3. Models Used in the Emulation of the Fluorouracil Target Trial

eAppendix 4. Models Used in the Emulation of the Erlotinib Target Trial

eAppendix 5. Sensitivity Analyses for the Fluorouracil Target Trial Emulation

eAppendix 6. Sensitivity Analyses for the Erlotinib Target Trial Emulation

1.
Visvanathan  K, Levit  LA, Raghavan  D,  et al.  Untapped potential of observational research to inform clinical decision making: American Society of Clinical Oncology Research Statement.  J Clin Oncol. 2017;35(16):1845-1854. doi:10.1200/JCO.2017.72.6414PubMedGoogle ScholarCrossref
2.
Hutchins  LF, Unger  JM, Crowley  JJ, Coltman  CA  Jr, Albain  KS.  Underrepresentation of patients 65 years of age or older in cancer-treatment trials.  N Engl J Med. 1999;341(27):2061-2067. doi:10.1056/NEJM199912303412706PubMedGoogle ScholarCrossref
3.
Murthy  VH, Krumholz  HM, Gross  CP.  Participation in cancer clinical trials: race-, sex-, and age-based disparities.  JAMA. 2004;291(22):2720-2726. doi:10.1001/jama.291.22.2720PubMedGoogle ScholarCrossref
4.
Suissa  S.  Immortal time bias in pharmaco-epidemiology.  Am J Epidemiol. 2008;167(4):492-499. doi:10.1093/aje/kwm324PubMedGoogle ScholarCrossref
5.
Hernán  MA, Robins  JM.  Using big data to emulate a target trial when a randomized trial is not available.  Am J Epidemiol. 2016;183(8):758-764. doi:10.1093/aje/kwv254PubMedGoogle ScholarCrossref
6.
Hernán  MA, Sauer  BC, Hernández-Díaz  S, Platt  R, Shrier  I.  Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses.  J Clin Epidemiol. 2016;79:70-75. doi:10.1016/j.jclinepi.2016.04.014PubMedGoogle ScholarCrossref
7.
Hernán  MA, Alonso  A, Logan  R,  et al.  Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease.  Epidemiology. 2008;19(6):766-779. doi:10.1097/EDE.0b013e3181875e61PubMedGoogle ScholarCrossref
8.
Dickerman  BA, García-Albéniz  X, Logan  RW, Denaxas  S, Hernán  MA.  Avoidable flaws in observational analyses: an application to statins and cancer.  Nat Med. 2019;25(10):1601-1606. doi:10.1038/s41591-019-0597-xPubMedGoogle ScholarCrossref
9.
Emilsson  L, García-Albéniz  X, Logan  RW, Caniglia  EC, Kalager  M, Hernán  MA.  Examining bias in studies of statin treatment and survival in patients with cancer.  JAMA Oncol. 2018;4(1):63-70. doi:10.1001/jamaoncol.2017.2752PubMedGoogle ScholarCrossref
10.
Warren  JLKC, Klabunde  CN, Schrag  D, Bach  PB, Riley  GF.  Overview of the SEER-Medicare data: content, research applications, and generalizability to the United States elderly population.  Med Care. 2002;40(8)(suppl):IV-3-IV-18. doi:10.1097/00005650-200208001-00002PubMedGoogle Scholar
11.
Yeh  JM, Tramontano  AC, Hur  C, Schrag  D.  Comparative effectiveness of adjuvant chemoradiotherapy after gastrectomy among older patients with gastric adenocarcinoma: a SEER-Medicare study.  Gastric Cancer. 2017;20(5):811-824. doi:10.1007/s10120-017-0693-xPubMedGoogle ScholarCrossref
12.
Fu  S, Wang  M, Zhao  H,  et al.  Comparative effectiveness of chemotherapy, rituximab, and bendamustine in Medicare beneficiaries with mantle-cell lymphoma.  Clin Lymphoma Myeloma Leuk. 2019;19(11):e616-e623. doi:10.1016/j.clml.2019.08.014PubMedGoogle ScholarCrossref
13.
Zeidan  AM, Davidoff  AJ, Long  JB,  et al.  Comparative clinical effectiveness of azacitidine versus decitabine in older patients with myelodysplastic syndromes.  Br J Haematol. 2016;175(5):829-840. doi:10.1111/bjh.14305PubMedGoogle ScholarCrossref
14.
Gray  R, Barnwell  J, McConkey  C, Hills  RK, Williams  NS, Kerr  DJ; Quasar Collaborative Group.  Adjuvant chemotherapy versus observation in patients with colorectal cancer: a randomised study.  Lancet. 2007;370(9604):2020-2029. doi:10.1016/S0140-6736(07)61866-2PubMedGoogle ScholarCrossref
15.
Moore  MJ, Goldstein  D, Hamm  J,  et al; National Cancer Institute of Canada Clinical Trials Group.  Erlotinib plus gemcitabine compared with gemcitabine alone in patients with advanced pancreatic cancer: a phase III trial of the National Cancer Institute of Canada Clinical Trials Group.  J Clin Oncol. 2007;25(15):1960-1966. doi:10.1200/JCO.2006.07.9525PubMedGoogle ScholarCrossref
16.
National Cancer Institute: Division of Cancer Control and Population Sciences. SEER-Medicare linked database. https://healthcaredelivery.cancer.gov/seermedicare/. Published 2018. Accessed January 20, 2020.
17.
Berger  ML, Mamdani  M, Atkins  D, Johnson  ML.  Good research practices for comparative effectiveness research: defining, reporting and interpreting nonrandomized studies of treatment effects using secondary data sources: the ISPOR Good Research Practices for Retrospective Database Analysis Task Force Report—part I.  Value Health. 2009;12(8):1044-1052. doi:10.1111/j.1524-4733.2009.00600.xPubMedGoogle ScholarCrossref
18.
von Elm  E, Altman  DG, Egger  M, Pocock  SJ, Gøtzsche  PC, Vandenbroucke  JP; STROBE Initiative.  The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies.  Ann Intern Med. 2007;147(8):573-577. doi:10.7326/0003-4819-147-8-200710160-00010PubMedGoogle ScholarCrossref
19.
Gagne  JJGR, Glynn  RJ, Avorn  J, Levin  R, Schneeweiss  S.  A combined comorbidity score predicted mortality in elderly patients better than existing scores.  J Clin Epidemiol. 2011;64(7):749-759. doi:10.1016/j.jclinepi.2010.10.004PubMedGoogle ScholarCrossref
20.
Davidoff  AJ, Gardner  LD, Zuckerman  IH, Hendrick  F, Ke  X, Edelman  MJ.  Validation of disability status, a claims-based measure of functional status for cancer treatment and outcomes studies.  Med Care. 2014;52(6):500-510. doi:10.1097/MLR.0000000000000122PubMedGoogle ScholarCrossref
21.
Davidoff  AJ, Zuckerman  IH, Pandya  N,  et al.  A novel approach to improve health status measurement in observational claims-based studies of cancer treatment and outcomes.  J Geriatr Oncol. 2013;4(2):157-165. doi:10.1016/j.jgo.2012.12.005PubMedGoogle ScholarCrossref
22.
Dobie  SA, Baldwin  LM, Dominitz  JA, Matthews  B, Billingsley  K, Barlow  W.  Completion of therapy by Medicare patients with stage III colon cancer.  J Natl Cancer Inst. 2006;98(9):610-619. doi:10.1093/jnci/djj159PubMedGoogle ScholarCrossref
23.
García-Albéniz  X, Hsu  J, Hernán  MA.  The value of explicitly emulating a target trial when using real world evidence: an application to colorectal cancer screening.  Eur J Epidemiol. 2017;32(6):495-500. doi:10.1007/s10654-017-0287-2PubMedGoogle ScholarCrossref
24.
O’Connor  ES, Greenblatt  DY, LoConte  NK,  et al.  Adjuvant chemotherapy for stage II colon cancer with poor prognostic features.  J Clin Oncol. 2011;29(25):3381-3388. doi:10.1200/JCO.2010.34.3426PubMedGoogle ScholarCrossref
25.
Weiss  JM, Schumacher  J, Allen  GO,  et al.  Adjuvant chemotherapy for stage II right-sided and left-sided colon cancer: analysis of SEER-Medicare data.  Ann Surg Oncol. 2014;21(6):1781-1791. doi:10.1245/s10434-014-3631-8PubMedGoogle ScholarCrossref
26.
Bao  Y, Maciejewski  RC, Garrido  MM, Shah  MA, Maciejewski  PK, Prigerson  HG.  Chemotherapy use, end-of-life care, and costs of care among patients diagnosed with stage IV pancreatic cancer.  J Pain Symptom Manage. 2018;55(4):1113-1121.e3. doi:10.1016/j.jpainsymman.2017.12.335PubMedGoogle ScholarCrossref
27.
SEER-Medicare Technical Support, Information Management Services. Documentation for the Patient Entitlement and Diagnosis Summary File (PEDSF). https://healthcaredelivery.cancer.gov/seermedicare/aboutdata/PEDSF.pdf. Published December 2018. Accessed January 20, 2020.
28.
National Cancer Institute: Division of Cancer Control and Population Sciences. Measures that are limited or not available in the data. https://healthcaredelivery.cancer.gov/seermedicare/considerations/measures.html. Accessed January 20, 2020.
29.
Hernán  MA, Robins  JM.  Per-protocol analyses of pragmatic trials.  N Engl J Med. 2017;377(14):1391-1398. doi:10.1056/NEJMsm1605385PubMedGoogle ScholarCrossref
30.
Hernán  MA.  How to estimate the effect of treatment duration on survival outcomes using observational data.  BMJ. 2018;360:k182. doi:10.1136/bmj.k182PubMedGoogle ScholarCrossref
31.
Hernán  MA, Hernández-Díaz  S, Robins  JM.  A structural approach to selection bias.  Epidemiology. 2004;15(5):615-625. doi:10.1097/01.ede.0000135174.63482.43PubMedGoogle ScholarCrossref
32.
Cole  SR, Hernán  MA.  Adjusted survival curves with inverse probability weights.  Comput Methods Programs Biomed. 2004;75(1):45-49. doi:10.1016/j.cmpb.2003.10.004PubMedGoogle ScholarCrossref
33.
Robins  JM, Hernán  MA, Brumback  B.  Marginal structural models and causal inference in epidemiology.  Epidemiology. 2000;11(5):550-560. doi:10.1097/00001648-200009000-00011PubMedGoogle ScholarCrossref
34.
Thompson  WA  Jr.  On the treatment of grouped observations in life studies.  Biometrics. 1977;33(3):463-470. doi:10.2307/2529360PubMedGoogle ScholarCrossref
35.
D’Agostino  RB, Lee  M-L, Belanger  AJ, Cupples  LA, Anderson  K, Kannel  WB.  Relation of pooled logistic regression to time dependent Cox regression analysis: the Framingham Heart Study.  Stat Med. 1990;9(12):1501-1515. doi:10.1002/sim.4780091214PubMedGoogle ScholarCrossref
36.
Toh  S, Hernández-Díaz  S, Logan  R, Robins  JM, Hernán  MA.  Estimating absolute risks in the presence of nonadherence: an application to a follow-up study with baseline randomization.  Epidemiology. 2010;21(4):528-539. doi:10.1097/EDE.0b013e3181df1b69PubMedGoogle ScholarCrossref
37.
Henson  KE, Jagsi  R, Cutter  D, McGale  P, Taylor  C, Darby  SC.  Inferring the effects of cancer treatment: divergent results from early breast cancer trialists’ collaborative group meta-analyses of randomized trials and observational data from SEER registries  [retracted in: J Clin Oncol. 2016;34(27):3358-3359].  J Clin Oncol. 2016;34(8):803-809. doi:10.1200/JCO.2015.62.0294PubMedGoogle ScholarCrossref
38.
 Retraction.  J Clin Oncol. 2016;34(27):3358-3359. doi:10.1200/JCO.2016.69.0875PubMedGoogle ScholarCrossref
39.
Yu  JB, Gross  CP, Wilson  LD, Smith  BD.  NCI SEER public-use data: applications and limitations in oncology research.  Oncology (Williston Park). 2009;23(3):288-295.PubMedGoogle Scholar
40.
Giordano  SH, Kuo  YF, Duan  Z, Hortobagyi  GN, Freeman  J, Goodwin  JS.  Limits of observational data in determining outcomes from cancer therapy.  Cancer. 2008;112(11):2456-2466. doi:10.1002/cncr.23452PubMedGoogle ScholarCrossref
41.
Soni  PD, Hartman  HE, Dess  RT,  et al.  Comparison of population-based observational studies with randomized trials in oncology.  J Clin Oncol. 2019;37(14):1209-1216. doi:10.1200/JCO.18.01074PubMedGoogle ScholarCrossref
42.
Ewald  H, Ioannidis  JPA, Ladanie  A, Mc Cord  K, Bucher  HC, Hemkens  LG.  Nonrandomized studies using causal-modeling may give different answers than RCTs: a meta-epidemiological study.  J Clin Epidemiol. 2019;118:29-41. doi:10.1016/j.jclinepi.2019.10.012PubMedGoogle ScholarCrossref
Limit 200 characters
Limit 25 characters
Conflicts of Interest Disclosure

Identify all potential conflicts of interest that might be relevant to your comment.

Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.

Err on the side of full disclosure.

If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.

Not all submitted comments are published. Please see our commenting policy for details.

Limit 140 characters
Limit 3600 characters or approximately 600 words
    Original Investigation
    Oncology
    March 5, 2020

    Estimates of Overall Survival in Patients With Cancer Receiving Different Treatment Regimens: Emulating Hypothetical Target Trials in the Surveillance, Epidemiology, and End Results (SEER)–Medicare Linked Database

    Author Affiliations
    • 1Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, Massachusetts
    • 2Now with the Feinberg School of Medicine, Division of Biostatistics, Department of Preventive Medicine, Northwestern University, Chicago, Illinois
    • 3Mongan Institute, Massachusetts General Hospital, Boston
    • 4RTI Health Solutions, Barcelona, Spain
    • 5Division of Cancer Control and Population Sciences, National Cancer Institute, Bethesda, Maryland
    • 6Center for Evidence Synthesis in Health, Brown University, Providence, Rhode Island
    • 7Department of Health Services, Policy and Practice, Brown University, Providence, Rhode Island
    • 8Department of Epidemiology, Brown University, Providence, Rhode Island
    • 9Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Massachusetts
    • 10Harvard-MIT Health Sciences and Technology, Harvard Medical School, Boston, Massachusetts
    JAMA Netw Open. 2020;3(3):e200452. doi:10.1001/jamanetworkopen.2020.0452
    Key Points español 中文 (chinese)

    Question  Does the Surveillance, Epidemiology, and End Results–Medicare linked database contain sufficient information to estimate the effectiveness of adding a drug to an existing treatment regimen for elderly individuals with cancer who are systematically underrepresented in randomized trials?

    Findings  After explicitly emulating target trials, findings from this comparative effectiveness study using Surveillance, Epidemiology, and End Results–Medicare data were not meaningfully different from those in elderly subgroup analyses reported from randomized trials. Naive observational estimates, however, were not compatible with those from previous trials.

    Meaning  Analyses using Surveillance, Epidemiology, and End Results–Medicare data may be informative for some research questions examining the comparative effectiveness of oncological therapies for elderly individuals.

    Abstract

    Importance  The Surveillance, Epidemiology, and End Results (SEER)–Medicare linked database may provide insights into the comparative effectiveness of oncological treatments for elderly individuals who are underrepresented in clinical trials.

    Objective  To evaluate the suitability of SEER-Medicare data for assessing the effectiveness of adding a drug to an existing treatment regimen on the overall survival of elderly patients with cancer.

    Design, Setting, and Participants  This comparative effectiveness study analyzed SEER-Medicare data from 9549 individuals who received a new diagnosis of stage II colorectal cancer (2008-2012) and 940 patients who received a new diagnosis of advanced pancreatic adenocarcinoma (2007-2012), with follow-up to December 31, 2013 (SEER-Medicare data released in 2015). Two (hypothetical) target trials were designed and emulated based on 2 existing randomized clinical trials: (1) adjuvant fluorouracil after curative surgery for individuals with stage II colorectal cancer and (2) erlotinib added to gemcitabine for individuals with advanced pancreatic adenocarcinoma. Data were analyzed January 2018 to March 2019.

    Exposures  The following treatment strategies were compared: (1) fluorouracil initiation vs no initiation within 3 months of tumor resection and (2) erlotinib initiation vs no initiation within 12 weeks of gemcitabine initiation.

    Main Outcomes and Measures  All-cause mortality within 60 months of baseline for the fluorouracil trial and within 72 weeks for the erlotinib trial.

    Results  Compared with 3293 individuals in the existing fluorouracil trial, 9549 eligible individuals included in the present analyses were more likely to have colon cancer (8565 [90%] vs 2291 [71%]) and were older (median [interquartile range], 79 [73-84] vs 63 [56-68] years). The 5-year risk difference for initiation vs noninitiation of fluorouracil after surgery was −3.8% (95% CI, −14.8% to 12.6%), and the mortality hazard ratio (HR) was 0.95 (95% CI, 0.85-1.04). Compared with 569 individuals in the existing erlotinib trial, 940 eligible patients included in the present analysis were older (median [range], 74 [66-93] vs 64 [36-92] years) and more likely to be male (547 [58%] vs 298 [52%]). The 1-year risk difference for initiation vs noninitiation of erlotinib was 4.7% (95% CI, −9.4% to 18.0%), and the corresponding mortality HR was 1.04 (95% CI, 0.86-1.42). In naive analyses, the mortality HR estimate was 1.14 (95% CI, 0.95-1.36) for the fluorouracil emulation and 0.68 (95% CI, 0.54-0.87) for the erlotinib emulation.

    Conclusions and Relevance  The present estimates were similar to those from randomized clinical trials that studied adding the same cancer drugs to existing regimens. The published HR was 1.02 (95% CI, 0.70-1.48) in the fluorouracil trial for individuals aged 70 or older and 0.96 (95% CI, 0.74-1.24) in the erlotinib trial for individuals aged 65 years or older. The SEER-Medicare database may be adequate for studying the real-world effectiveness of adding a drug to treatment regimens used for elderly individuals with cancer.

    Introduction

    When randomized clinical trials are not available at the time a clinical decision needs to be made, clinicians often must choose between making decisions based on either observational data or no human data at all. A cautious use of observational data has been recently endorsed by the American Society of Clinical Oncology, which states that observational studies can “inform questions that either have not been or cannot be answered by randomized trials”1(p1845) and “complement the evidence collected in randomized trials.”1(p1845) Observational studies may be particularly important for patients who are underrepresented in trials, such as elderly patients.2,3

    Although health care databases make it possible to study cancer treatments as they are used in clinical practice, observational analyses are subject to several sources of bias.4 Some of these biases, however, can be eliminated by using observational data to emulate a (hypothetical) pragmatic target trial.5,6 Explicitly emulating a target trial is important because some well-known failures of observational research7-9 were the result of deviating from basic principles of study design rather than with the shortcomings inherent to observational data.

    Here, we demonstrate how to use the Surveillance, Epidemiology, and End Results (SEER)–Medicare database10—a linkage between the SEER cancer registries and Medicare—to emulate target trials that compare cancer treatments for elderly individuals. Our goal is to describe procedures to increase the validity of comparative effectiveness analyses using the SEER-Medicare database.11-13 To demonstrate these procedures, we emulated 2 target trials for estimating the effect of the addition of a cancer drug to an existing treatment regimen on overall survival: (1) adjuvant fluorouracil after curative surgery in individuals with stage II colorectal cancer and (2) the addition of erlotinib to a regimen of gemcitabine for individuals with advanced pancreatic adenocarcinoma. As a benchmark, we used the age-specific estimates from 2 published randomized clinical trials,14,15 both of which estimated a near-null and imprecise effect for elderly individuals: hazard ratios (HRs) of 1.02 (95% CI, 0.70-1.48) for fluorouracil and 0.96 (95% CI, 0.74-1.24) for erlotinib. We also compared the emulated randomized clinical trial estimates with those from naive observational analyses that did not attempt to emulate a target trial.

    Methods
    Study Data: SEER-Medicare Linked Database

    The SEER-Medicare database is a linkage of patient demographic and cancer-related variables collected by 17 SEER cancer registries across 12 states with Medicare claim files from the Centers for Medicare & Medicaid Services. The SEER data are summarized in the Patient Entitlement and Diagnosis Summary File, which is linked to Medicare claims. More information about the SEER-Medicare linked database can be found elsewhere.16 This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline and the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) reporting guideline for good practices for comparative effectiveness research.17,18 The institutional review board at the Harvard T.H. Chan School of Public Health deemed this study exempt from review because the research involves the study of publicly available deidentified data.

    To identify eligible individuals, construct a combined comorbidity score19 and a performance status proxy score,20,21 and also identify all other variables in this study, we used Medicare claims from the Inpatient, Outpatient, Home Health Agency, Durable Medical Equipment, Medpar, National Claims History, and Patient Entitlement and Diagnosis Summary Files. To identify treatment, we used the Healthcare Common Procedure Coding System and Common Procedural Terminology codes. Because erlotinib is an oral medication, we determined its use by records of filled prescriptions in Medicare Part D, which started in 2007. See eAppendix 1 in the Supplement for details on data use.

    Emulating a Target Trial in the SEER-Medicare Linked Database

    Our approach had 2 steps. The first step was to fully articulate the clinical question by specifying the protocol of the (hypothetical) target trial.5 The second step was to emulate the target trial using the observational data. Table 1 and Table 2 outline the specifications of 2 target trials: (1) the addition of fluorouracil in stage II colorectal cancer and (2) the use of erlotinib in advanced pancreatic adenocarcinoma, based on existing randomized trials of fluorouracil14 and erlotinib15 (eTable 1 and eTable 2 in the Supplement). Table 1 and Table 2 also summarize the emulation procedure for both target trials using SEER-Medicare data. Key steps of the procedure are described herein.

    Eligibility Criteria

    The patient eligibility process is given in eFigure 1 and eFigure 2 in the Supplement. In addition to the eligibility criteria of the existing trials, our target trials also required having 1 year of continuous enrollment in Medicare Parts A and B, without being enrolled in a health maintenance organization, and having less than 3 months between diagnosis and initiation of first-line treatment.

    For the fluorouracil trial, there were no eligibility restrictions related to the use of neoadjuvant chemotherapy or radiation. In addition, because some colorectal cancers are diagnosed during surgery, we allowed surgery to occur up to 4 weeks before the recorded date of cancer diagnosis to account for time spent in pathologic confirmation of the diagnosis.

    For the erlotinib target trial, 3 months of enrollment in Medicare Part D was also required. In addition, individuals with an early cancer stage (I, II, or III) who had undergone surgery to remove the tumor began a second eligibility period if they initiated gemcitabine at least 12 weeks after surgery (otherwise gemcitabine was considered adjuvant care rather than treatment for progression) and before initiating other adjuvant chemotherapy or radiation; this modification emulated the randomized clinical trial conducted by Moore et al15 (eTable 2 in the Supplement).

    Treatment Strategies and Assignment

    The treatment strategies compared in our hypothetical target trials mimicked those in 2 trials: (1) initiation of fluorouracil vs no initiation of fluorouracil and (2) initiation of erlotinib vs no initiation of erlotinib in individuals receiving gemcitabine. All components of the treatment strategy needed to be in place within a prespecified period (the grace period) after baseline. Because the existing trials did not explicitly specify the length of the grace period, we chose a 3-month period based on our understanding of the clinical management of stage II colorectal cancer and of advanced pancreatic adenocarcinoma.

    Observational analyses attempt to emulate randomized assignment via adjustment for confounders (ie, potential prognostic factors that are imbalanced between treatment groups at baseline). For example, performance status was expected to be a confounder for the effect of erlotinib, whereas perceived risk of relapse (summarized here by size of the tumor, grade of differentiation, number of resected nodes, and T category) was expected to be a confounder for that of fluorouracil. We adjusted for those factors and demographic characteristics, geographic characteristics, and comorbidities (as identified by previously validated procedure and diagnosis codes20-26) listed in Table 1 and Table 2. Target trials must be pragmatic trials in which treatment assignment is not blinded because observational data typically cannot be used to emulate double-blind trials.5

    Start and End of Follow-up

    The start of follow-up (baseline or time zero) for each individual was the time of first eligibility: the date of the first gemcitabine claim while otherwise eligible in the emulation of the erlotinib trial and the date of hospital discharge after surgery while otherwise eligible in the emulation of the fluorouracil trial. The end of follow-up was death, loss of insurance eligibility follow-up (loss of enrollment in Parts A, B, or D; enrollment in a health maintenance organization), or administrative end of follow-up (December 31, 2013; 60 months for the fluorouracil trial and 72 weeks for the erlotinib trial), whichever was earliest.

    Outcome

    The primary outcome of interest in both target trials was all-cause mortality. The SEER-Medicare database contains information about the date of death in both the SEER file and the Medicare claims. We used the Medicare date of death, which is more up to date and is validated by the National Death Index.27 The SEER-Medicare data do not include enough information to reliably study progression-free survival.28

    Causal Contrast

    Our causal contrast of interest was the per-protocol effect, that is, the effect that would have been observed if all trial participants had adhered to the treatment strategies specified in the protocol.29 Because our treatment strategies simply required the initiation of either fluorouracil or erlotinib within the grace period (regardless of future continuation), our causal contrast was similar to the intention-to-treat effect in a trial in which all individuals initiated their assigned treatment during the grace period, even if some of them later discontinued treatment.

    Statistical Analysis

    We refer to analyses whose goal was to estimate the per-protocol effect as “per-protocol analyses”; these analyses should not to be confused with naive per-protocol analyses conducted in some trials.5 There are 3 steps of the per-protocol analysis that we implemented in our emulation of the target trials, which could also be implemented in target trials if they were actually conducted. (Technical details are provided in eAppendix 2 in the Supplement.)

    For the first step, we censored individuals if or when they deviated from their assigned protocol: individuals who (1) initiated fluorouracil in the observation strategy or (2) initiated erlotinib in the gemcitabine alone strategy. In addition, because eligible individuals could not be uniquely assigned to a strategy at baseline when emulating target trials with a grace period, we duplicated (“cloned”) eligible individuals and assigned 1 clone to each treatment strategy (eFigure 3 in the Supplement).30 We then censored clones if they had no data compatible with their assigned treatment strategy by the end of the grace period.6,30 Details on all possible censoring schemes are given in eAppendix 2 in the Supplement.

    Second, we used inverse probability weighting to adjust for the potential selection bias introduced by censoring.31 To do so, we estimated stabilized inverse probability weights as a function of the variables listed in Table 1 and Table 2.32,33 Third, we fit an inverse probability weighted discrete-time hazard model34,35 by pooled logistic regression, with death as the response and the following regressors: the indicator for the assigned treatment strategy, a function of time of follow-up (restricted cubic spline with knots at 3, 16, 30, 44, and 57 months for the fluorouracil emulation; quadratic for the erlotinib emulation), product terms for treatment strategy and time, and baseline covariates. To obtain standardized, treatment-specific risks for time points between baseline and the end of follow-up, we standardized the model-derived estimated values to the joint distribution of the baseline covariates. To estimate the mean HR commonly reported in randomized clinical trials, we fit a Cox proportional hazards model with treatment strategy as the only covariate, as described elsewhere.36 The final models are described in eAppendix 3 and eAppendix 4 in the Supplement.

    We conducted several sensitivity analyses to test the robustness of our estimates to different choices of functional forms and grace periods (results were similar; see eAppendix 5 and eAppendix 6 in the Supplement). We computed 95% CIs via a nonparametric bootstrap based on 500 resamplings. We used SAS, version 9.4 (SAS Institute Inc), for data processing, and R, version 3.4.4 (The R Foundation), for data analyses. The SEER-Medicare data were from those released in 2015 (containing content through December 31, 2013), and our analyses were conducted from January 2018 to March 2019.

    Results

    We identified 9549 eligible patients for the fluorouracil trial and 940 for the erlotinib trial, as described in eFigure 1 and eFigure 2 in the Supplement, respectively.

    Fluorouracil for Stage II Colorectal Cancer

    The baseline characteristics of the eligible individuals are given eTable 3 in the Supplement. Compared with 3293 individuals in the existing randomized clinical trial,14 9549 eligible individuals included in the present analyses were more likely to have colon cancer (8565 [90%] vs 2291 [71%]) and were older (median [interquartile range] age, 79 [73-84] years vs 63 [56-68] years) (eTable 4 in the Supplement).

    Of the 9549 eligible individuals included in the present analysis (23 447 person-years of follow-up), 204 initiated fluorouracil-based chemotherapy within 3 months of their hospital discharge. Fluorouracil initiation was more likely for people who were younger, married, and had a T4 tumor category at diagnosis and rectum involvement (eAppendix 3 in the Supplement). Individuals were less likely to initiate fluorouracil-based chemotherapy if they had a prolonged hospitalization after surgery, preoperative radiotherapy, and, in the year before diagnosis, anemia and asthenia. By the end of the grace period, 185 individuals remained in the fluorouracil group and 6150 in the observation group (eTable 3 in the Supplement).

    There were 2148 deaths during follow-up. The adjusted estimated 5-year survival was 66.6% for fluorouracil and 62.7% for no fluorouracil; the 5-year risk difference was −3.8% (95% CI, −14.8% to 12.6%) (Figure 1). The mortality HR for fluorouracil vs no fluorouracil was 1.00 (95% CI, 0.89-1.12) without any adjustment, 1.01 (95% CI, 0.97-1.05) after adjustment for baseline covariates, and 0.95 (95% CI, 0.85-1.04) after adjustment for baseline and postbaseline covariates.

    Erlotinib for Advanced Pancreatic Cancer

    The baseline characteristics of the eligible individuals are given in eTable 5 in the Supplement. Compared with 569 individuals in the existing randomized clinical trial,15 940 individuals included in the present analysis were older (median [range] age, 74 [66-93] years vs 64 [36-92] years), more likely to be male (547 [58%] vs 298 [52%]), and less likely to have undergone prior chemotherapy (53 [6%] vs 45 [8%]). In addition, all participants in the existing trial had a good performance status at enrollment, whereas 82 of the eligible individuals in our study (9%) had a poor performance status20,21 (eTable 6 in the Supplement).

    Of the 940 eligible individuals (412 person-years of follow-up), 62 initiated erlotinib within 12 weeks of their initial gemcitabine dose. Erlotinib initiation was more likely for individuals with stage IV cancer and for those with comorbidities, and less likely for those with a poor performance status and older age (eAppendix 4 in the Supplement). By the end of the grace period, 44 individuals remained with the gemcitabine plus erlotinib strategy and 480 remained with the gemcitabine alone strategy (eTable 5 in the Supplement).

    There were 659 deaths during follow-up. The adjusted estimated 1-year survival was 15.6% for gemcitabine plus erlotinib and 20.4% in gemcitabine alone; the risk difference was 4.7% (95% CI, −9.4% to 18.0%) (Figure 2). The all-cause mortality HR for gemcitabine plus erlotinib vs gemcitabine alone was 1.08 (95% CI, 0.93-1.25) without any adjustment, 1.03 (95% CI, 0.96-1.09) after adjustment for baseline covariates, and 1.04 (95% CI, 0.86-1.42) after adjustment for baseline and postbaseline covariates.

    Naive Analyses

    For comparison purposes, we conducted an analysis without cloning, censoring by deviation from protocol, or inverse probability weighting. Eligible individuals for the fluorouracil comparison were assigned to the fluorouracil group if they started fluorouracil therapy at any time during the follow-up and to the no fluorouracil group otherwise. We then fit a Cox proportional hazards regression model with treatment group and the baseline covariates. The mortality HR estimate was 1.14 (95% CI, 0.95-1.36). In a similar analysis for the erlotinib comparison, the mortality HR estimate was 0.68 (95% CI, 0.54-0.87).

    Discussion

    We described 2 target trials for the addition of a cancer drug to an existing treatment regimen. These target trials were based on 2 existing randomized clinical trials: 1 pragmatic trial (fluorouracil) and 1 placebo-controlled, double-blind trial (erlotinib).14,15 We emulated these target trials using observational data from individuals 65 years of age or older whose data were in the SEER-Medicare database. Our observational estimates were consistent with the null estimates reported by the existing trials when we explicitly emulated the target trials but not when we conducted naive analyses that did not attempt to emulate a target trial.

    The first emulation yielded a mortality HR of 0.95 (95% CI, 0.85-1.04) for fluorouracil vs no fluorouracil initiation, which is compatible with that reported in the actual fluorouracil trial14 for individuals 70 years of age or older (HR, 1.02; 95% CI, 0.70-1.48). Five-year survival could not be directly compared because the actual trial did not report these numbers for the elderly individuals. The second emulation yielded a mortality HR of 1.04 (95% CI, 0.86-1.42) for erlotinib vs no erlotinib initiation after gemcitabine, which is compatible with that reported in the existing erlotinib trial15 for individuals 65 years of age or older (HR, 0.96; 95% CI, 0.74-1.24). The estimated 1-year survival in the erlotinib strategy was approximately 16% in the emulated trial and 23% in the existing trial; the estimated 1-year survival in the gemcitabine alone strategy was approximately 20% in the emulated trial and 17% in the existing trial.

    The target trial framework requires data sources with adequate longitudinal information on treatments, outcomes, and confounders. We were able to obtain that information from the linkage of detailed cancer information from the SEER registries with administrative data from Medicare claims. Detailed timing information for comorbidity onset and treatment administration exists solely in Medicare; thus, using SEER only would preclude us from emulating target trials. The use of SEER registry data alone may have contributed to erroneous conclusions in recent reports (later retracted)37,38 that compared cancer treatments in the elderly population. The use of Medicare data enabled us to reduce some errors present in SEER alone (eg, in the radiation variable) but does not guarantee perfect coding reliability of all variables (eg, rare histologic results).39

    Similar to all observational analyses, ours assume that the lack of randomized treatment assignment can be approximately replaced by adjustment for the measured confounders. In our analyses, after the target trial was specified and the emulation procedures followed, we found no indication of strong confounding by measured variables. In fact, adjustment for multiple prognostic factors only modestly changed the estimates in both emulations. Given that we adjusted for some of the most important indications for the studied treatments, it is likely that confounding was not the most important validity threat in our examples.

    Leaving aside a lack of randomization, a major threat to the validity of observational comparative effectiveness analyses is a failure to emulate components of a target trial other than randomization, which may introduce biases such as selection bias and immortal time bias. Explicit emulation of a target trial helps eliminate these problems.6,23

    Similar to our naive analyses, previous uses of SEER-Medicare data to estimate treatment effects that did not explicitly attempt to emulate a target trial may have introduced immortal time bias.40 For the same reasons, prior comparisons of randomized clinical trials and observational studies based on SEER-Medicare41 or other databases42 are difficult to interpret because, in the absence of an explicit emulation of a target trial, any discrepancies may be due to incorrect analysis rather than to inherent limitations of the observational data.

    Limitations

    The use of large observational databases does not guarantee transportable or precise estimates. The requirements of target trial emulation (eg, requiring individuals to participate in Medicare Part D, which was necessary in the erlotinib analysis) may limit the generalizability of our results, although probably not as much as do the eligibility criteria of actual randomized clinical trials. Perhaps the most surprising aspect of our analyses was the small sample sizes despite the apparently large amount of data in the SEER-Medicare database. Because few eligible individuals received the adjuvant therapy of interest, there were fewer than 200 individuals in the fluorouracil strategy after 3 months of follow-up and fewer than 190 individuals in the erlotinib strategy after 12 weeks of follow-up. When, as in our analyses, few individuals follow the treatment strategies of interest, the researchers’ ability to adequately control for confounding and selection bias is limited because not many covariates can be added to the regression models.

    Conclusions

    We emulated 2 target trials for the addition of a cancer treatment by using the SEER-Medicare database. Our findings showed that, in some settings, observational estimates from this database are consistent with those from randomized clinical trials in elderly populations, but only when an appropriate causal inference approach is implemented.

    As the US population ages, understanding how cancer treatments affect elderly individuals will have profound implications for clinical decision-making. Because elderly individuals tend to be underrepresented in randomized clinical trials,2,3 observational health care databases can be used to obtain more precise estimates in that segment of the population or, at the very least, to guide the design of future randomized clinical trials by helping to rule out clearly ineffective or harmful treatment strategies.

    Back to top
    Article Information

    Accepted for Publication: January 8, 2020.

    Published: March 5, 2020. doi:10.1001/jamanetworkopen.2020.0452

    Correction: This article was corrected on April 6, 2020, to remove data from 1 sentence in the Results and to remove or replace data in the Supplement in violation of the SEER-Medicare data use agreement.

    Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2020 Petito LC et al. JAMA Network Open.

    Corresponding Author: Lucia C. Petito, PhD, Feinberg School of Medicine, Division of Biostatistics, Department of Preventive Medicine, Northwestern University, 680 N Lake Shore Dr, Ste 1400, Chicago, IL 60611 (lucia.petito@northwestern.edu).

    Author Contributions: Dr Petito had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

    Concept and design: Petito, García-Albéniz, Howlader, Mariotto, Dahabreh, Hernán.

    Acquisition, analysis, or interpretation of data: Petito, Logan, Hernán.

    Drafting of the manuscript: Petito, Logan, Hernán.

    Critical revision of the manuscript for important intellectual content: Petito, García-Albéniz, Howlader, Mariotto, Dahabreh, Hernán.

    Statistical analysis: Petito, Logan, Dahabreh, Hernán.

    Obtained funding: Howlader, Hernán.

    Administrative, technical, or material support: Logan, Howlader, Mariotto, Hernán.

    Supervision: García-Albéniz, Hernán.

    Conflict of Interest Disclosures: Dr Petito reported that she is currently funded by the following National Institutes of Health grants as a coinvestigator: 5R01HL133860-04 (National Heart, Lung, and Blood Institute [NHLBI]), 5R01HL096875-08 (NHLBI), 4R33AG057383-03 (National Institute on Aging [NIA]), 5R01MD012236-02 (National Institute on Minority Health and Health Disparities), 1R01AG059291-01A1 (NIA), 1R01HL141881-01A1 (NHLBI), 2UL1TR001422-05A1 (National Center for Advancing Translational Sciences). Drs Logan and Hernán reported receiving grants from the National Institutes of Health during the conduct of the study. Dr Dahabreh reported receiving personal fees from Sanofi Pasteur outside the submitted work.

    Funding/Support: This work was supported by the US National Cancer Institute contract HHSN261201700047 and grant P01-CA134294 (to Drs Petito, García-Albéniz, Logan, and Hernán), and by Patient-Centered Outcomes Research Institute awards ME-1306-03758 and ME-1502-27794 (to Dr Dahabreh). The collection of cancer incidence data used in this study was supported by the California Department of Public Health as part of the statewide cancer reporting program mandated by California Health and Safety Code Section 103885; by the Surveillance, Epidemiology and End Results Program of the National Cancer Institute under contract HHSN261201000140C awarded to the Cancer Prevention Institute of California, contract HHSN261201000035C awarded to the University of Southern California, and contract HHSN261201000034C awarded to the Public Health Institute; and by the National Program of Cancer Registries from the Centers for Disease Control and Prevention under agreement U58DP003862-01 awarded to the California Department of Public Health.

    Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or the decision to submit the manuscript for publication.

    Disclaimer: This study used the linked Surveillance, Epidemiology, and End Results (SEER)–Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors. The ideas and opinions expressed herein are those of the authors and endorsement by the California Department of Public Health, the National Cancer Institute, and the Centers for Disease Control and Prevention or their Contractors and Subcontractors is not intended nor should be inferred. Statements in this article do not necessarily represent the views of the Patient-Centered Outcomes Research Institute, its Board of Governors, or the Methodology Committee.

    Additional Contributions: We acknowledge the efforts of the National Cancer Institute; the Office of Research, Development and Information Centers for Medicare and Medicaid Services; Information Management Services Inc; and the SEER Program tumor registries in the creation of the SEER-Medicare database.

    References
    1.
    Visvanathan  K, Levit  LA, Raghavan  D,  et al.  Untapped potential of observational research to inform clinical decision making: American Society of Clinical Oncology Research Statement.  J Clin Oncol. 2017;35(16):1845-1854. doi:10.1200/JCO.2017.72.6414PubMedGoogle ScholarCrossref
    2.
    Hutchins  LF, Unger  JM, Crowley  JJ, Coltman  CA  Jr, Albain  KS.  Underrepresentation of patients 65 years of age or older in cancer-treatment trials.  N Engl J Med. 1999;341(27):2061-2067. doi:10.1056/NEJM199912303412706PubMedGoogle ScholarCrossref
    3.
    Murthy  VH, Krumholz  HM, Gross  CP.  Participation in cancer clinical trials: race-, sex-, and age-based disparities.  JAMA. 2004;291(22):2720-2726. doi:10.1001/jama.291.22.2720PubMedGoogle ScholarCrossref
    4.
    Suissa  S.  Immortal time bias in pharmaco-epidemiology.  Am J Epidemiol. 2008;167(4):492-499. doi:10.1093/aje/kwm324PubMedGoogle ScholarCrossref
    5.
    Hernán  MA, Robins  JM.  Using big data to emulate a target trial when a randomized trial is not available.  Am J Epidemiol. 2016;183(8):758-764. doi:10.1093/aje/kwv254PubMedGoogle ScholarCrossref
    6.
    Hernán  MA, Sauer  BC, Hernández-Díaz  S, Platt  R, Shrier  I.  Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses.  J Clin Epidemiol. 2016;79:70-75. doi:10.1016/j.jclinepi.2016.04.014PubMedGoogle ScholarCrossref
    7.
    Hernán  MA, Alonso  A, Logan  R,  et al.  Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease.  Epidemiology. 2008;19(6):766-779. doi:10.1097/EDE.0b013e3181875e61PubMedGoogle ScholarCrossref
    8.
    Dickerman  BA, García-Albéniz  X, Logan  RW, Denaxas  S, Hernán  MA.  Avoidable flaws in observational analyses: an application to statins and cancer.  Nat Med. 2019;25(10):1601-1606. doi:10.1038/s41591-019-0597-xPubMedGoogle ScholarCrossref
    9.
    Emilsson  L, García-Albéniz  X, Logan  RW, Caniglia  EC, Kalager  M, Hernán  MA.  Examining bias in studies of statin treatment and survival in patients with cancer.  JAMA Oncol. 2018;4(1):63-70. doi:10.1001/jamaoncol.2017.2752PubMedGoogle ScholarCrossref
    10.
    Warren  JLKC, Klabunde  CN, Schrag  D, Bach  PB, Riley  GF.  Overview of the SEER-Medicare data: content, research applications, and generalizability to the United States elderly population.  Med Care. 2002;40(8)(suppl):IV-3-IV-18. doi:10.1097/00005650-200208001-00002PubMedGoogle Scholar
    11.
    Yeh  JM, Tramontano  AC, Hur  C, Schrag  D.  Comparative effectiveness of adjuvant chemoradiotherapy after gastrectomy among older patients with gastric adenocarcinoma: a SEER-Medicare study.  Gastric Cancer. 2017;20(5):811-824. doi:10.1007/s10120-017-0693-xPubMedGoogle ScholarCrossref
    12.
    Fu  S, Wang  M, Zhao  H,  et al.  Comparative effectiveness of chemotherapy, rituximab, and bendamustine in Medicare beneficiaries with mantle-cell lymphoma.  Clin Lymphoma Myeloma Leuk. 2019;19(11):e616-e623. doi:10.1016/j.clml.2019.08.014PubMedGoogle ScholarCrossref
    13.
    Zeidan  AM, Davidoff  AJ, Long  JB,  et al.  Comparative clinical effectiveness of azacitidine versus decitabine in older patients with myelodysplastic syndromes.  Br J Haematol. 2016;175(5):829-840. doi:10.1111/bjh.14305PubMedGoogle ScholarCrossref
    14.
    Gray  R, Barnwell  J, McConkey  C, Hills  RK, Williams  NS, Kerr  DJ; Quasar Collaborative Group.  Adjuvant chemotherapy versus observation in patients with colorectal cancer: a randomised study.  Lancet. 2007;370(9604):2020-2029. doi:10.1016/S0140-6736(07)61866-2PubMedGoogle ScholarCrossref
    15.
    Moore  MJ, Goldstein  D, Hamm  J,  et al; National Cancer Institute of Canada Clinical Trials Group.  Erlotinib plus gemcitabine compared with gemcitabine alone in patients with advanced pancreatic cancer: a phase III trial of the National Cancer Institute of Canada Clinical Trials Group.  J Clin Oncol. 2007;25(15):1960-1966. doi:10.1200/JCO.2006.07.9525PubMedGoogle ScholarCrossref
    16.
    National Cancer Institute: Division of Cancer Control and Population Sciences. SEER-Medicare linked database. https://healthcaredelivery.cancer.gov/seermedicare/. Published 2018. Accessed January 20, 2020.
    17.
    Berger  ML, Mamdani  M, Atkins  D, Johnson  ML.  Good research practices for comparative effectiveness research: defining, reporting and interpreting nonrandomized studies of treatment effects using secondary data sources: the ISPOR Good Research Practices for Retrospective Database Analysis Task Force Report—part I.  Value Health. 2009;12(8):1044-1052. doi:10.1111/j.1524-4733.2009.00600.xPubMedGoogle ScholarCrossref
    18.
    von Elm  E, Altman  DG, Egger  M, Pocock  SJ, Gøtzsche  PC, Vandenbroucke  JP; STROBE Initiative.  The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies.  Ann Intern Med. 2007;147(8):573-577. doi:10.7326/0003-4819-147-8-200710160-00010PubMedGoogle ScholarCrossref
    19.
    Gagne  JJGR, Glynn  RJ, Avorn  J, Levin  R, Schneeweiss  S.  A combined comorbidity score predicted mortality in elderly patients better than existing scores.  J Clin Epidemiol. 2011;64(7):749-759. doi:10.1016/j.jclinepi.2010.10.004PubMedGoogle ScholarCrossref
    20.
    Davidoff  AJ, Gardner  LD, Zuckerman  IH, Hendrick  F, Ke  X, Edelman  MJ.  Validation of disability status, a claims-based measure of functional status for cancer treatment and outcomes studies.  Med Care. 2014;52(6):500-510. doi:10.1097/MLR.0000000000000122PubMedGoogle ScholarCrossref
    21.
    Davidoff  AJ, Zuckerman  IH, Pandya  N,  et al.  A novel approach to improve health status measurement in observational claims-based studies of cancer treatment and outcomes.  J Geriatr Oncol. 2013;4(2):157-165. doi:10.1016/j.jgo.2012.12.005PubMedGoogle ScholarCrossref
    22.
    Dobie  SA, Baldwin  LM, Dominitz  JA, Matthews  B, Billingsley  K, Barlow  W.  Completion of therapy by Medicare patients with stage III colon cancer.  J Natl Cancer Inst. 2006;98(9):610-619. doi:10.1093/jnci/djj159PubMedGoogle ScholarCrossref
    23.
    García-Albéniz  X, Hsu  J, Hernán  MA.  The value of explicitly emulating a target trial when using real world evidence: an application to colorectal cancer screening.  Eur J Epidemiol. 2017;32(6):495-500. doi:10.1007/s10654-017-0287-2PubMedGoogle ScholarCrossref
    24.
    O’Connor  ES, Greenblatt  DY, LoConte  NK,  et al.  Adjuvant chemotherapy for stage II colon cancer with poor prognostic features.  J Clin Oncol. 2011;29(25):3381-3388. doi:10.1200/JCO.2010.34.3426PubMedGoogle ScholarCrossref
    25.
    Weiss  JM, Schumacher  J, Allen  GO,  et al.  Adjuvant chemotherapy for stage II right-sided and left-sided colon cancer: analysis of SEER-Medicare data.  Ann Surg Oncol. 2014;21(6):1781-1791. doi:10.1245/s10434-014-3631-8PubMedGoogle ScholarCrossref
    26.
    Bao  Y, Maciejewski  RC, Garrido  MM, Shah  MA, Maciejewski  PK, Prigerson  HG.  Chemotherapy use, end-of-life care, and costs of care among patients diagnosed with stage IV pancreatic cancer.  J Pain Symptom Manage. 2018;55(4):1113-1121.e3. doi:10.1016/j.jpainsymman.2017.12.335PubMedGoogle ScholarCrossref
    27.
    SEER-Medicare Technical Support, Information Management Services. Documentation for the Patient Entitlement and Diagnosis Summary File (PEDSF). https://healthcaredelivery.cancer.gov/seermedicare/aboutdata/PEDSF.pdf. Published December 2018. Accessed January 20, 2020.
    28.
    National Cancer Institute: Division of Cancer Control and Population Sciences. Measures that are limited or not available in the data. https://healthcaredelivery.cancer.gov/seermedicare/considerations/measures.html. Accessed January 20, 2020.
    29.
    Hernán  MA, Robins  JM.  Per-protocol analyses of pragmatic trials.  N Engl J Med. 2017;377(14):1391-1398. doi:10.1056/NEJMsm1605385PubMedGoogle ScholarCrossref
    30.
    Hernán  MA.  How to estimate the effect of treatment duration on survival outcomes using observational data.  BMJ. 2018;360:k182. doi:10.1136/bmj.k182PubMedGoogle ScholarCrossref
    31.
    Hernán  MA, Hernández-Díaz  S, Robins  JM.  A structural approach to selection bias.  Epidemiology. 2004;15(5):615-625. doi:10.1097/01.ede.0000135174.63482.43PubMedGoogle ScholarCrossref
    32.
    Cole  SR, Hernán  MA.  Adjusted survival curves with inverse probability weights.  Comput Methods Programs Biomed. 2004;75(1):45-49. doi:10.1016/j.cmpb.2003.10.004PubMedGoogle ScholarCrossref
    33.
    Robins  JM, Hernán  MA, Brumback  B.  Marginal structural models and causal inference in epidemiology.  Epidemiology. 2000;11(5):550-560. doi:10.1097/00001648-200009000-00011PubMedGoogle ScholarCrossref
    34.
    Thompson  WA  Jr.  On the treatment of grouped observations in life studies.  Biometrics. 1977;33(3):463-470. doi:10.2307/2529360PubMedGoogle ScholarCrossref
    35.
    D’Agostino  RB, Lee  M-L, Belanger  AJ, Cupples  LA, Anderson  K, Kannel  WB.  Relation of pooled logistic regression to time dependent Cox regression analysis: the Framingham Heart Study.  Stat Med. 1990;9(12):1501-1515. doi:10.1002/sim.4780091214PubMedGoogle ScholarCrossref
    36.
    Toh  S, Hernández-Díaz  S, Logan  R, Robins  JM, Hernán  MA.  Estimating absolute risks in the presence of nonadherence: an application to a follow-up study with baseline randomization.  Epidemiology. 2010;21(4):528-539. doi:10.1097/EDE.0b013e3181df1b69PubMedGoogle ScholarCrossref
    37.
    Henson  KE, Jagsi  R, Cutter  D, McGale  P, Taylor  C, Darby  SC.  Inferring the effects of cancer treatment: divergent results from early breast cancer trialists’ collaborative group meta-analyses of randomized trials and observational data from SEER registries  [retracted in: J Clin Oncol. 2016;34(27):3358-3359].  J Clin Oncol. 2016;34(8):803-809. doi:10.1200/JCO.2015.62.0294PubMedGoogle ScholarCrossref
    38.
     Retraction.  J Clin Oncol. 2016;34(27):3358-3359. doi:10.1200/JCO.2016.69.0875PubMedGoogle ScholarCrossref
    39.
    Yu  JB, Gross  CP, Wilson  LD, Smith  BD.  NCI SEER public-use data: applications and limitations in oncology research.  Oncology (Williston Park). 2009;23(3):288-295.PubMedGoogle Scholar
    40.
    Giordano  SH, Kuo  YF, Duan  Z, Hortobagyi  GN, Freeman  J, Goodwin  JS.  Limits of observational data in determining outcomes from cancer therapy.  Cancer. 2008;112(11):2456-2466. doi:10.1002/cncr.23452PubMedGoogle ScholarCrossref
    41.
    Soni  PD, Hartman  HE, Dess  RT,  et al.  Comparison of population-based observational studies with randomized trials in oncology.  J Clin Oncol. 2019;37(14):1209-1216. doi:10.1200/JCO.18.01074PubMedGoogle ScholarCrossref
    42.
    Ewald  H, Ioannidis  JPA, Ladanie  A, Mc Cord  K, Bucher  HC, Hemkens  LG.  Nonrandomized studies using causal-modeling may give different answers than RCTs: a meta-epidemiological study.  J Clin Epidemiol. 2019;118:29-41. doi:10.1016/j.jclinepi.2019.10.012PubMedGoogle ScholarCrossref
    ×