A Transla)onal Framework for Methodological Rigor to Improve Pa)ent Centered Outcomes in End of Life Cancer Research Francesca Dominici, PhD Senior Associate Dean for Research Professor of Biostatistics Harvard TH Chan School of Public Health K18 HS021991
PROJECT OVERVIEW
Specific Aims 1. Par0cipate in an intense, mentored career development experience in CER with a special focus on cancer 2. Conduct a research project on Glioblastoma 3. Maximize the policy impact of the research
Study Team Nils Arvold, MD Attending Physician, Radiation Oncology Associates, St. Luke s Cancer Center and University of Minnesota Duluth Former Assistant Professor, Harvard Medical School; and Neuro-Radiation Oncologist and Fellowship Director, Department of Radiation Oncology, Dana-Farber/Brigham & Women s Cancer Center Students & Postdocs Matthew Cefalu, PhD Associate Statistician at the RAND Corporation Former Postdoctoral Fellow, Department of Biostatistics, Harvard T.H. Chan School of Public Health Cory Zigler, PhD Assistant Professor of Biostatistics, Department of Biostatistics, Harvard T.H. Chan School of Public Health Danielle Braun, PhD Postdoctoral Fellow, Department of Biostatistics, Harvard T.H. Chan School of Public Health Deborah Schrag, MD, MPH Professor of Medicine at Harvard Medical School and the Chief of the Division of Population Sciences in the Department of Medical Oncology at the Dana-Farber Cancer Institute Joey Antonelli, PhD Doctoral Student, Department of Biostatistics, Harvard University Yun Wang Senior Research Scientist in the Department of Biostatistics at the Harvard T.H. Chan School of Public Health
Formulate Clinical Ques0ons Construct the analy0cal data set from claims and cancer registry data Develop new methods (whenever is necessary) Results Dissemination
Clinical Questions 1. What factors are associated with a high hospitalization burden among brain cancer patients? 2. Does adding chemotherapy to radiation prolong survival for elderly brain cancer patients? 3. Do palliative care interventions among terminal cancer patients reduce hospitalizations, ER visits and other measures of aggressive care? Adjustment for confounding 1 Methodological Challenges 3 Measurement error in the treatment assignment Treatment effect heterogeneity/ personalized medicine 2 Statistical Methods Development 4 Combining heterogeneous sources of data NEW PROJECT
CLINICAL CONTEXT Comparative Effectiveness Research for Medicare Patients with Glioblastoma
Glioblastoma (GBM) 55,000 primary brain tumors per year in U.S. GBM is the most common malignant primary brain tumor Approximately 15,000 GBM cases/year in U.S. Median age at GBM diagnosis: 65 years
Glioblastoma (GBM) Despite half of all GBM pa0ents being elderly, older pa0ents under- represented in RCTs Age 70 y.o. excluded from 2005 landmark RCT Median survival range for GBM: 6 to 18 months Uncertainty about how to treat and how to improve quality of life
Three Clinical Ques0ons 1. What factors are associated with a high hospitalization burden among brain cancer patients? Arvold ND, Wang Y, Zigler C, Schrag D, Dominici F (2014) Hospitalization burden and survival among elderly patients with glioblastoma, Neuro-Oncology 2. Does adding chemotherapy to radiotherapy prolong survival for elderly brain cancer patients? Arvold ND, Cefalu M, Wang Y, Zigler C, Schrag D, Dominici F Radiotherapy with vs. without temozolomide in older patients with glioblastoma, submitted 3. Do palliative care interventions among terminal cancer patients reduce hospitalizations, ER visits and other measures of aggressive care? Routine care settings, populations with little/no clinical trial evidence
Ques0on #2: TMZ/RT vs. RT in Elderly GBM Pa0ents Purpose: To examine overall survival among elderly GBM patients receiving TMZ/RT vs. RT alone Rationale: Concurrent TMZ/RT widely used/recommended for elderly GBM patients, but benefit of TMZ is unclear in this population
Ques0on #3: Pallia0ve Care and Quality at End of Life Is receipt of palliative care associated with reduced aggressiveness of end-oflife care in advanced cancer Medicare patients? Treatment: receiving a palliative care intervention at EOL Outcomes receipt of chemotherapy within 30 days of death (Yes vs No) more than 1 emergency room visit within 30 days of death (Yes vs No) more than 1 hospitalization within 30 days of death (Yes vs No), or alternatively, cumulative # of days hospitalized within 30 days of death death at home (%) enrolled on hospice (%) overall survival from diagnosis date
IN- HOUSE AVAILABLE MEDICARE DATA
In- house Data 100% sample of inpatient claims data (1999-2013) Condition-specific post-acute care data (2009-2011, for pancreases, brain, colon, lung, and bladder cancers) Outpatient Nursing home Hospice (not in our SEER-Medicare data) Home health care Part B, and Durable medical equipment 100% sample of Medicare enrollment data (1999-2013) Condition-specific SEER Medicare (1991-2009) Prostate, stomach, bladder, colorectal, breast, lung, and brain cancers
Post-acute Care, an Important Aspect of Care Poten0al Sequence ader Index Hospitaliza0on Index Admission To Home To SNF Hospice Death Re-Admission
METHODOLOGICAL CHALLENGES
Methodological Challenge #1: Confounding Adjustment Uncertainty Treatments are not randomized and need statistical adjustment to estimate causal effects. Existing approaches assume that the potential confounders are known and measured. Often we have a high-dimensional set of possible confounders in administrative data: Demographics (age, race, sex...) Clinical characteris0cs (tumor size, comorbidi0es,...) Hospital characteris0cs (pa0ent volume, teaching,...) Physician characteris0cs (specialty, case volume,...) In reality, the factors required for adjustment are unknown and must be chosen from a high-dimensional set of possibilities.
Baseline characteristics (% experiencing) and 1-year mortality rate for patients treated with temozolomide plus radiotherapy (n=776) and radiotherapy alone (n=1111). We also report estimated inclusion probabilities defined as the probability that each of these characteristics to be an important confounder. Cefalu M, Dominici F, Arvold N, Parmigiani G (2015) A Model averaged double robust estimator, Biometrics (under review)
Unadjusted 11.7% (7.6-16.0%) Probability of death within one year Adjusted 6.7% (2.4-10.7%) All possible combinations (1000 x 1000) of propensity score models and outcome models based on which confounders they include
Methodological Challenge 2: Treatment Effect Heterogeneity (personalized medicine) Treatments do not affect everyone the same Exis0ng methods es0mate what happens on average. Ideally we would like to iden0fy popula0on subgroups with different treatment effects and es0mate a different effect in each group. How can we use data on high- dimensional pa0ent characteris0cs to iden0fy which subgroups exhibit heterogeneity in treatment response?
Wang C, Dominici F, Parmigiani G, Zigler CM (2015) Accounting for uncertainty in confounder and effect modifier selection when estimating average causal effects in generalized linear models, Biometrics, 20 April 2015. Doi: 10.1111/biom.12315 Important confounders Important effect modifiers
Methodological Challenge 3: Treatment Assignment is measured with error in claims data ICD9 billing codes in claims data inaccurately reflect surgical treatment or any other procedure In SEER data, treatment is ascertained using medical chart reviews Medicare part A claims data are available for n=41,971 (1999-2007) SEER- Medicare data are available for n=5,463 (1999-2007) Sensi0vity=96.8%, specificity=73.8% Goal: doing a CER of surgical resec0on versus biopsy in the whole Medicare popula0on, using SEER- Medicare to correct for error in the treatment assignment
Braun D, Parmigiani G, Arvold N, Gorfine M, Dominici F, Zigler C Propensity Scores with Measurement Error in the Treatment Assignment: a Likelihood-Based Adjustment, submitted Medicare Part A/Seer-Medicare Data Analysis Results Average Treatment Effect [95% CI] SEER-Medicare (Gold -0.03 [-0.07, 0.01] Standard) Medicare Part A No Adjustment -0.13 [-0.13, -0.12] An ATE of -0:16 implies that the probability of dying within one year is 16% larger for those who received a biopsy compared to those who had a resection Medicare Part A (Adjustment) -0.16 [-0.18, -0.14]
Methodological Challenge 4: Combining Medicare and SEER- Medicare to adjust for confounding SEER- Medicare data has an extensive set of confounders (stage of the tumor, loca0on of the tumor, extent of the resec0on, data) Medicare claims data has very limited informa0on on confounders (age, race, zip code of residence) Y Death within 1 year X Surgery vs. biopsy C Age, gender, race, comorbidities and region, from Medicare data U e.g. Tumor number, size, and location, from SEER-Medicare m Goal: CER of surgical resec0on versus biopsy on 1 year mortality in the whole in Medicare popula0on but using all the available measured confounders in SEER m+n n>>m Antonelli J, Dominici F, Using External Validation Data to Adjust for Confounding, in prep
DISSEMINATION
PUBLICATIONS AND ADDITIONAL FUNDED GRANTS
Publications hrp://wordpress.sph.harvard.edu/dominici- lab/research/compara0ve- effec0veness- research- in- cancer/ Arvold ND, Wang Y, Zigler C, Schrag D, Dominici F (2014) Hospitaliza0on burden and survival among elderly pa0ents with Glioblastoma, Neuro- Oncology, 16-11:1530-40. doi: 10.1093/neuonc/nov060. PMCID: PMC4201065. Read more about this paper on Healio.com Zigler CM, Dominici F (2014) Uncertainty in propensity score es0ma0on: Bayesian methods for variable selec0on and model averaged causal effects. JASA, 109(505):95-107. PMCID: PMC3703764 Wang Y, Schrag D, Brooks G, Dominici F (2014) Na0onal trends in pancrea0c cancer outcomes and parern of care among Medicare beneficiaries: 2000-2010, Cancer, 120(7):1050-8. DOI: 10.1002/cncr. 2853. PMCID: PMC4019988. Obermeyer Z, Makar M, Abujaber S, Dominici F, Block S, Cutler DM. (2014) Associa0on between the Medicare hospice benefit and health care u0liza0on and costs for pa0ents with poor- prognosis cancer, JAMA, 312(18):1888-1896. doi:10.1001/jama.2014.14950. PMCID: PMC4274169. Wang C, Dominici F, Parmigiani G, Zigler CM (2015) Accoun0ng for uncertainty in confounder and effect modifier selec0on when es0ma0ng average causal effects in generalized linear models, Biometrics, 20 April 2015. Doi: 10.1111/biom.12315 Cefalu M, Dominici F, Arvold N, Parmigiani G (2015) A Model averaged double robust es0mator, Biometrics (under review) Braun D, Parmigiani G, Arvold N, Gorfine M, Dominici F Zigler C Propensity Scores with Measurement Error in the Treatment Assignment: a Likelihood- Based Adjustment, submired Arvold ND, Cefalu M, Wang Y, Zigler C, Schrag D, Dominici F Radiotherapy with vs. without temozolomide in older pa0ents with glioblastoma, submired
Funded Grants R01 GM111339 (Normand/Dominici) Bayesian Methods for Compara)ve Effec)veness Research with Observa)onal Data Francesca Dominici, PhD Professor of Biostatistics, Department of Biostatistics & Senior Associate Dean for Research, Office of the Dean, Harvard T.H. Chan School of Public Health Sharon-Lise Normand, PhD Professor of Health Care Policy in the Department of Health Care Policy at Harvard Medical School and Mentor of Dr. Dominici K18 Cory Zigler, PhD Assistant Professor of Biostatistics, Department of Biostatistics, Harvard T.H. Chan School of Public Health Sherri Rose, PhD Assistant Professor of Health Care Policy (Biostatistics) in the Department of Health Care Policy at Harvard Medical School
Ques0ons for the Advisory Board Addressing Methodological Gaps: How has this research project advanced the field of CER in cancer and what are some of the questions that will be important to investigate moving forward? Assessing Clinical Impact: Based on the research results, what are the next steps for translating this knowledge into actionable hospital and outpatientbased performance metrics? Stakeholder Engagement & Patient Advocacy: How can we be more responsive to and involve stakeholders in the research project and what is the best approach to engage stakeholders? Dissemination: What are the target populations for the dissemination of research results and what is the most effective way to reach these groups? Policy Impact: Moving forward, how do we maximize the policy impact of the research?