Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Evaluating cancer treatments with health economics methods
(USC Thesis Other)
Evaluating cancer treatments with health economics methods
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
i Evaluating cancer treatments with health economics methods by Xiayu Jiao A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY In the subject of HEALTH ECONOMICS December 2019 ii Table of Contents Chapter 1: Introduction ...............................................................................................................1 References ..................................................................................................................................6 Chapter 2: Outcomes and utilization of adjuvant chemotherapy for stage II colon cancer in the elderly population: a SEER-Medicare analysis ............................................................................6 Abstract ......................................................................................................................................6 Introduction ................................................................................................................................7 Methods ......................................................................................................................................9 Results ...................................................................................................................................... 16 Discussion ................................................................................................................................. 27 References ................................................................................................................................ 30 Chapter 3: One discrete choice experiment study measuring the treatment preferences among patients with NSCLC (Non-Small Cell Lung Cancer) ................................................................ 32 Abstract .................................................................................................................................... 32 Introduction .............................................................................................................................. 33 Methods .................................................................................................................................... 34 Results ...................................................................................................................................... 42 Discussion ................................................................................................................................. 48 References ................................................................................................................................ 51 Chapter 4: Testing methodological advances in willingness to pay (WTP) estimation using a discrete choice experiment: an advanced lung cancer treatment case study ................................ 53 Abstract .................................................................................................................................... 53 Introduction .............................................................................................................................. 54 Method ..................................................................................................................................... 58 Results ...................................................................................................................................... 69 iii Discussion ................................................................................................................................. 78 Appendix .................................................................................................................................. 80 References ................................................................................................................................ 81 Chapter 5: Summary and future research directions ................................................................... 83 iv Acknowledgments I would like to express my deep gratitude to Dr. Joel Hay, my advisor, for his patient guidance, generous help and valuable critique throughout my journey at USC. I am grateful to have Dr. Hay as my advisor, not only for the knowledge he imparted but also for the inspiration he provided with me. Thank you for setting up a role model for me, who works diligently and devotedly for the benefits of people. Also, thanks to your funny jokes that always cheer up the spirits. I would like to sincerely thank Dr. Nieva and Dr. Barzi for their guidance and suggestions from the clinical perspective. Without your help, none of these studies would have been possible. I would also like to thank the patients who participated in my study, thank you for your contributions and inputs. I would also like to acknowledge Dr. Steven Fox, who devoted numerous time and effort to help me grow as a better researcher. I would like to thank Dr. John Romley and Dr. Jason Doctor, for their guidance, suggestions, and comments on my research. I would like to express the gratitude for our beloved Dr. Jeffrey McCombs, thank you for always having an open heart for your students. I also want to acknowledge Patricia St Clair for her gracious effort in helping me with the laborious data problems. I want to thank all my friends for enriching my life and enlightening my world. Special thanks to my roommate Tianyi Lu, I will cherish those happy days that were full of laughter with you. Another special mention to Bo Zhou, for helping me generously and encouraging me every time I lost confidence. Lastly, to my parents: I feel so grateful to be your daughter, thank you for loving me, believing in me, and encouraging me to explore the world as much as I can. 1 Chapter 1: Introduction Cancer has been titled as “the emperor of all maladies” by Dr. Siddhartha Mukherjee in his popular book[1], in which the author traces the history of cancer back 4,600 years ago, from its first identification by the Egyptian physician Imhotep. Nowadays, cancer is still recognized as one of the leading causes of death in the United States (US). In fact, the American Cancer Society projected that in 2019, there will be 1,762,450 cases of new cancer (all sites), and an estimated 606,880 deaths will be attributable to this disease.[2] Fortunately, breakthroughs in scientific research have contributed to an increased number of approved new cancer drug therapies. In addition to conventional chemotherapy, patients with cancer now can be provided with multiple treatment alternatives, including targeted therapy and immunotherapy. In 2018 alone, the U.S. FDA approved 63 new agents or new indications for existing medicines in cancer treatment.[3] With the evolving landscape of cancer treatment, there is a growing need to develop more studies to examine and compare the outcome of cancer therapies. Moreover, cancer treatment imposes a substantial burden on patients and their families, including decreased quality of life and elevated cost burden. Therefore, researchers need to conduct studies to understand and assess patients’ preferences towards treatment, which will help to optimize their treatments and support the best practices of cancer care. The first study is one retrospective study that investigates the role of adjuvant chemotherapy among patients with Stage II colon cancer. Generally, randomized clinical trials (RCT) study is regarded as a good design in examining the effectiveness of a new intervention or treatment. Since the act of randomization can balance both observed and unobserved characteristics between groups, which prevents potential endogeneity problems that may influence the outcome.[4] However, on one hand, the randomization assignment is not ethical in 2 some circumstances. On the other hand, given the homogeneous patient population, highly- structured treatment regimen, and intensive monitoring pattern in RCTs[5], one treatment whose efficacy was demonstrated in clinical trials may perform ambiguously in the real world. Therefore, in order to address this uncertainty, researchers should adopt alternative study designs to investigate the performance of treatment in real-world settings with heterogeneous patient populations. Retrospective studies with statistical methods to control for potential selection biases and other confounding factors, can provide helpful information that complements RCT findings. Moreover, in today’s “big data” era, the observational study can combine with digital technology to include more characteristics/factors in the analysis (such as the abundant data from wearable technology devices) and follow patients in a longer-term duration. Our first study examines the controversial use of adjuvant chemotherapy among patients with stage II colon cancer. In 2004, oxaliplatin and capecitabine were added to the armamentarium of the agents. Evidence documenting the performance and impact of contemporary adjuvant chemotherapy in real-world settings is limited. While oncologists, patients, and their families need real-world evidence to make more informed treatment decisions. Policymakers also rely on information to make policy adjustments for their specific populations. To satisfy these unmet research needs, this study provides an in-depth analysis of SEER- Medicare data for the adoption and benefit of adjuvant chemotherapy in patients with stage II colon cancer who underwent curative resection after the approval of oxaliplatin. The SEER- Medicare database contains substantial Medicare claim records of cancer patients, who were followed by the epidemiologic surveillance system program (SEER) all over the country. In order to control for selection bias, this study uses the instrumental variable regression analysis to not only control for the observable factors such as patient and tumor-related characteristics, but 3 also the unobservable influences on the treatment assignment. By utilizing large national datasets, employing econometric models, and controlling for endogeneity problems, this paper presents an overview of adjuvant chemotherapy use over time (2004-2011) and provides evidence for its effectiveness in the real world. The next two studies employ discrete choice experiments (DCE) to understand the treatment preferences for non-small cell lung cancer (NSCLC). Today, lung and bronchus cancer are still among the top common cancers and the leading causes of cancer deaths for both men and women in the US. In 2019, it is estimated that there will be 228,150 new cases of lung and bronchus cancer and 142,670 deaths are estimated to occur because of the disease.[2] About 52.3% of patients with lung and bronchus cancer were diagnosed at an advanced stage.[6] 5-year relative survival for these patients was only 5.2%.[7] NSCLC is the most common lung cancer, accounting for about 84% of all cases.[8] Fortunately, thanks to the cutting-edge advances in technology, patients with advanced NSCLC who are eligible for targeted therapies or immunotherapies are experiencing improved survival outcomes.[9, 10] In the situation of such severe disease, with more and more treatment options became available, it is important to promote patient-centered care and consider patients’ attitudes/needs in their treatment decisions. The prerequisite for patient-centered care is to gather information about patients’ treatment preferences. There are several methods for understanding preferences. Broadly speaking, these methods can be classified into two groups: revealed preference and stated preference approaches. Revealed preference data is obtained from past behavior. However, it is not possible to use revealed preference approach to value goods/services that are not on the market yet. In contrast, stated preference approaches collect data through surveys. They enable the flexible valuation of 4 goods/services that even not exist. In addition to DCE, there are other commonly used stated- preference elicitation approaches for health care, such as contingent valuation[11] and conjoint analysis[12]. Contingent valuation is a method of directly obtaining a monetary value of a specified good. This method gathers information about the attitude toward a specific scenario. Conversely, the DCE allows researchers to value the effect of attributes, which enable the indirect value estimation when the situation changes. In our case, DCE allows us to measure the relative importance of treatment attributes for patients in their treatment decisions, quantitively. Conjoint analysis, like DCE, is also one multi-attribute method. However, conjoint analysis is not associated with any statistical error theory, which allows the theory to be represented as testable statistical models.[13] On the contrary, DCE is theoretically rooted in well-tested theories, like the classical random utility maximization (RUM).[14] Given DCE has a well- developed theoretical basis, it is more attractive than other approaches in understanding and assessing preferences. In chapter 3, we present a DCE study among 71 patients with NSCLC at Norris Comprehensive Cancer Center. This study estimates and quantifies the relative weights of treatment attributes that patients place on their decision making. These attributes include progression-free survival time, out-of-pocket cost, control of symptoms, and three treatment- related adverse effects: tiredness, skin itching, and hair loss. We use the mixed logit model to allow the attribute coefficients to vary across individuals, accounting for preference heterogeneity. Based on the estimated coefficients, we also calculate the willingness-to-pay (WTP) for improvements in treatment characteristics. Furthermore, we apply our results in the discussion of epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) and we predicate patients’ WTP for osimertinib after the patents of erlotinib and gefitinib expire. 5 A valuable implication from DCE is the derivation of WTP. The WTP obtained from the RUM is irrelevant to the context, which means the amount of money a decision-maker is willing to pay for some treatment improvements is independent of the relative performance of other treatment alternatives in the same choice set. Recently, there is a growing interest in implementing an alternative behavioral framework for understanding choice behavior, the context-dependent random regret minimization (RRM) model[15]. The regret-minimization based models postulate that people aim to minimize anticipated regret when making a choice. In these models, preferences for each alternative depend not only on the performance for that alternative, but also on the relative performance of all the other competing alternatives in one choice set. The derived WTPs from these models are also context-dependent, which allows for a richer interpretation of the trade-offs from decision-makers. In Chapter 4, the third paper focuses on testing methodological advances in WTP estimation using RRM modeling approaches in DCE. We develop another DCE questionnaire which emphasizes the improvements in survival probability rather than the gains in survival time from previous DCE. To be specific, we assess the societal preferences for advanced lung cancer therapies on one-year progression-free survival probability, out-of-pocket cost, and adverse effects including diarrhea, skin rash and tiredness using a US representative sample population from Amazon Mechanical Turk. At last, we empirically compare the WTP estimation in varying scenarios from the results of RUM and a hybrid model in which some attributes are processed in a utility-maximization way while some are processed in a regret-minimization way. Finally, we summarize the overall findings of each analysis in Chapter 5, then provide some direction for future research in these areas. 6 References 1. Mukherjee S. The emperor of all maladies: a biography of cancer: Simon and Schuster; 2010. 2. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA: a cancer journal for clinicians 2019. 3. Food and Drug Administration ,Hematology/Oncology (Cancer) Approvals & Safety Notifications. 4. Roberts C, Torgerson D. Randomisation methods in controlled trials. BMJ 1998;317(7168):1301- 1310. 5. Revicki DA, Frank L. Pharmacoeconomic evaluation in the real world. Pharmacoeconomics 1999;15(5):423-434. 6. Surveillance, Epidemiology, and End Results (SEER) Program (www.seer.cancer.gov) SEER*Stat Database: Incidence - SEER 21 Regs Limited-Field Research Data + Hurricane Katrina Impacted Louisiana Cases, Nov 2018 Sub (2000-2016) <Katrina/Rita Population Adjustment> - Linked To County Attributes - Total U.S., 1969-2017 Counties, National Cancer Institute, DCCPS, Surveillance Research Program, released April 2019, based on the November 2018 submission. In. 7. National Cancer Institute. Surveillance EaERP. Cancer stat facts: lung and bronchus cancer. In: National Cancer Institute Bethesda; 2018. 8. Detterbeck FC, Boffa DJ, Tanoue LT. The new lung cancer staging system. Chest 2009;136(1):260- 271. 9. Reck M, Rodríguez-Abreu D, Robinson A, et al. Updated analysis of KEYNOTE-024: pembrolizumab versus platinum-based chemotherapy for advanced non-small-cell lung cancer with PD-L1 tumor proportion score of 50% or greater. 2019. 10. Zhao D, Chen X, Qin N, et al. The prognostic role of EGFR-TKIs for patients with advanced non- small cell lung cancer. Scientific reports 2017;7:40374. 11. Bayoumi AM. The measurement of contingent valuation for health economics. Pharmacoeconomics 2004;22(11):691-700. 12. Ryan M, Farrar S. Using conjoint analysis to elicit preferences for health care. BMJ 2000;320(7248):1530-1533. 13. Louviere JJ, Flynn TN, Carson RT. Discrete choice experiments are not conjoint analysis. Journal of Choice Modelling 2010;3(3):57-72. 14. McFadden D. Conditional logit analysis of qualitative choice behavior. 1973. 15. Chorus CG. A new model of random regret minimization. European Journal of Transport and Infrastructure Research 2010;10(2). Chapter 2: Outcomes and utilization of adjuvant chemotherapy for stage II colon cancer in the elderly population: a SEER-Medicare analysis Abstract Background 7 Based on existing trials, there is no definitive role for adjuvant chemotherapy (AC) for stage II colon cancer. Since a significant portion of patients with colon cancer are elderly, AC in this population requires a more thorough deliberation of risks and benefits. Using the SEER-Medicare dataset, we explored the utilization and outcome of AC in the population after the FDA approval of oxaliplatin. Methods Patients with stage II colon cancer (2004-2011) who underwent resection were selected for this analysis. Medicare claims data were used to ascertain the administration of AC within 120 days after surgery. The primary endpoint of the analysis was overall survival (OS). We used the Cox proportional hazards model to estimate the effect of AC while adjusting for clinical and sociodemographic variables available in SEER. To adjust for selection bias, we conducted an instrumental variable analysis using the instruments developed from treatment patterns from surgeons and health service areas. Results A total of 16,468 patients were identified and 12.1% received AC. AC delivery was decreasing over time. AC recipients were significantly younger, more likely to be male, non-white, married, and had lower comorbidity index. They were more likely to had the advanced stage, left-sided, and were less differentiated tumors. The hazard ratio from the Cox model showed a statistically significant survival advantage for AC (Hazard Ratio: 0.85, 95% CI:0.78 to 0.92). However, results from the instrumental variable regression analysis indicated that there was no definitive benefit of survival in AC recipients (HR=1.78, 95% CI: 0.93 to 3.41). Conclusion Despite its delivery to healthier and higher risk elderly patients, no survival benefit was observed related to AC. Future studies may elucidate the elderly population who may benefit from AC. Introduction Colon cancer is the third most common cancer in the United States. In 2019 an estimated 101,420 patients will be diagnosed with colon cancer; of those, about one-quarter will have stage II disease.[1] Although adjuvant chemotherapy (AC) with fluoropyrimidine with or without oxaliplatin is an accepted standard for stage III colon cancer, its role for stage II disease remains controversial.[2-4] Despite efforts for risk stratification of stage II patients, no subgroup is found 8 to achieve a definitive benefit from AC. [5] Consensus expert opinion recommends AC for stage II patients with high-risk features, including obstruction, perforation, and lymphovascular invasion.[6] One of the largest trials that investigated the benefit of AC in patients with stage II colon cancer is the United Kingdom QUASAR (QUick And Simple And Reliable Study).[7] In this study, patients with resected stage II cancer were randomly assigned to observation versus adjuvant fluoropyrimidine and leucovorin, with or without levamisole. AC was associated with an 18% reduction in risk of death, which translated into a small, albeit meaningful 3.6% (95% CI, 1.0% to 6.0%) benefit in five-year overall survival. However, due to strict selection criteria, like other RCTs, results of the QUARSA may not be generalizable to the general population.[8, 9] Therefore, it is important to utilize observational data to complement the findings. The SEER (Surveillance, Epidemiology, and End Results) program covers approximately 34.6 percent of the U.S population.[10] SEER registries collect data on patient demographics, primary tumor site, tumor morphology, stage at diagnosis, and cause of death. The SEER- Medicare data links cancer records from SEER with Medicare’s health claims for its beneficiaries from the time of a person’s Medicare eligibility until death.[11] Colon cancer incidence increases with age, and the median age of diagnosis in the US is 70 years. More than 62% of patients with colon cancer are older than 65 at the time of diagnosis. [12] Therefore, the SEER-Medicare database can serve as a platform for studies of colon cancer-related treatment patterns. Previous SEER-Medicare studies did not report significant survival improvements for AC for stage II disease.[13-15] The results of these studies are regarded as biased due to the inability to fully control for factors that contribute to the delivery of AC.[16] 9 Schrag et al. identified patients with resected stage II colon cancer diagnosed between 1991 to 1996 and reported that 27% of the patients received AC.[13] There was no survival advantage for the recipients of AC vs. non-recipients with HR 0.91 (95% CI, 0.77 to 1.09). Similarly, O’Connor et al. utilized SEER-Medicare and examined the AC use pattern for those diagnosed between 1992 to 2005.[14] The percentage of AC use among patients without/with poor prognostic features was 19% and 21%, respectively. They also reported no survival benefit for AC in patients with stage II colon cancer without/with poor prognostic features (HR, 1.02; 95% CI, 0.84 to 1.25 and 1.03; 95% CI, 0.94 to 1.13) respectively. Finally, Weiss et al. reported that 18% of stage II patients with right-sided cancer and 22 % with left-sided cancer received AC from 1992 to 2005.[15] However, no survival benefit was observed for stage II patients with right-sided (HR, 0.97; 95 % CI, 0.87 to 1.09; p = 0.64) or left-sided cancer (HR, 0.97; 95 % CI, 0.84 to 1.12; p = 0.68). These SEER-Medicare studies investigated AC treatment patterns before 2005. The treatment landscape for colorectal cancer has advanced significantly since 2004 after oxaliplatin and capecitabine were added to the armamentarium of drugs for AC. The objectives of this study were: (1) to describe the use of adjuvant therapy with or without oxaliplatin and its effectiveness in SEER-Medicare population in more recent years; (2) to use the instrumental variables regression analysis to account for both observable and unobservable confounders to examine the role of AC on overall survival. Methods Cohort identification We identified colon cancer patients diagnosed between 2004 to 2011, using SEER’s ICD- O-3 (International Classification of Diseases for Oncology, third edition) topographical codes 10 C18.0-C18.9.[17] We restricted our cohort to patients with an adenocarcinoma histology with SEER’s ICD-O-3 morphological codes: 8140, 8263, 8480, 8210, 8000, 8261, 8481, 8010, 8490, 8255, 8262, 8211, 8574, 8020, 8560, 8260, 8244, 8221, 8220, and 8144. We distinguished stage II colon cancer by incorporating information about tumor extension, tumor lymph nodes involvement, and tumor metastasis status according to AJCC (American Joint Committee Cancer) staging guideline including pT3, pT4a, pT4b; N0; M0. We selected patients who were over 65 years at their diagnosis. Only patients covered by Medicare parts A and B during the year prior to their cancer diagnosis were included in this analysis. Patients with enrollment in HMO plans do not have claims data and were therefore excluded from the analysis. We limited our cohort to patients who had surgery within 120 days of their initial diagnosis using inpatient and outpatient ICD-9 procedure codes (45.7x and 45.8x). We excluded patients who died within 90 days after the surgery since these patients were potentially too ill to receive AC. We also searched for chemotherapy claims prior to surgery and excluded patients who received chemotherapy within 180 days before surgery (Figure 2.1). 11 Figure 2.1 Cohort identification Identification of AC We defined AC as chemotherapy with 5-FU, capecitabine, with or without oxaliplatin within 120 days from the date of surgery using ICD-9 diagnosis/procedural codes, healthcare common procedure coding system (HCPCS) codes, and revenue center codes in Medicare claims. Patient with claims for chemotherapy drugs, chemotherapy administration, or medical evaluation of chemotherapy in either hospital outpatient facility (Medpar), or physician supplier (carrier) claims within the 120 days after surgery, were labeled as an AC recipient (Table 2.1). 12 Table 2.1 Definitions of chemotherapy from SEER-Medicare data Chemotherapy ICD-9 diagnosis Codes V58.1, V66.2, V67.2 ICD-9 procedural Codes 99.25 HCPCS codes J9190 (Fluorouracil) J0640 (Leucovorin) J9263 (Oxaliplatin) J8520 J8521 (Capecitabine) Q0084, Q0085 G0355-G0363 C8953 C8954 C8955 S9329 S9330 S9331 96400-96599 Revenue Center 0331 0332 0335 Statistical analysis Demographic and tumor-related variables were extracted from SEER data. These include age at diagnosis, race, gender, marital status, tumor grade, tumor location, number of lymph nodes examined, differentiation grade, poverty level (census tract level data), and registry region. Some of the characteristics are composite and calculated using a set of existing variables. NCI comorbidity index score was calculated using the NCI comorbidity macros.[18] We divided stage II patients into two subgroups: IIA and IIB/C, following AJCC 7 th edition guidelines and stratified our cohort by their cancer location—right-side versus left-side. 13 For the descriptive analysis, we used the Chi-square statistic to compare proportions between recipients of AC vs. non-recipients. Among recipients of AC, we identified those who received an oxaliplatin-containing regimen and reported their clinical characteristics. Furthermore, we assessed the trends in AC and oxaliplatin use from 2004 to 2011. The primary outcome of this study was overall survival, which was defined as the time from 90 days after surgery until death or the end of the observation period (12/31/2013). After evaluating the proportional hazards assumption, we used the Cox proportional hazards model to estimate the overall survival between patients who received versus those who did not receive AC. We included age at diagnosis, NCI comorbidity index score, race, gender, marital status, tumor grade, tumor location, number of lymph nodes examined, differentiation grade, poverty level, and the region as covariates. To determine whether an unmeasured confounding factor was likely to negate the observed result, we calculated the E-value.[19, 20] The E-value measures whether an unmeasured confounder affects the treatment and outcome, while simultaneously considering the measured covariates. We used instrumental variable regression analysis, which controls for potential unmeasured confounders in treatment assignment by estimating treatment effects using only the variation in treatment choices determined by variation related to the instrument, analogous to variations that result from randomization.[21] The instrumental variable analysis adjusts for the effects of both observable and unobservable characteristics. The key step in conducting instrumental variable analysis is identifying instruments, which significantly affect treatment choice (AC or not) but are not directly related to the health outcome. 14 Cox models are nonlinear models, therefore, conventional two-stage least squares (2SLS) procedure is not appropriate. Instead, we implemented the two-stage residual inclusion (2SRI) method. The 2SRI method has been shown to have positive properties when estimating treatment effects using nonlinear regression methods. 19 As with 2SLS, this method requires a two-stage estimation approach. The process includes two stages: The first stage is based on the likelihood of receiving treatment with instruments and other exogenous factors; the second-stage is the estimation of treatment on survival while including the residuals from the first stage and other exogenous factors. Specifically, the first stage is: 𝐴𝐶 = 𝑝𝑟𝑜𝑏𝑖𝑡 (𝑊𝛼 ) + 𝑋𝑢 where AC=adjuvant chemotherapy, α denotes the regression parameters, 𝑊 = [𝑋 0 𝑊 + ] and 𝑊 + = [𝐼𝑛𝑠𝑡𝑟𝑢𝑚𝑒𝑛𝑡𝑠 ]. 𝑋 0 represents other observable confounders The residuals from this regression are: 𝑋𝑢 ̂ = 𝐴𝐶 − 𝑝𝑟𝑜𝑏𝑖𝑡 (𝑊 𝛼 ̂ ) where 𝛼 ̂ is the column of consistently estimated parameters, 𝑋𝑢 ̂ denotes the difference between the actual value of the treatment choice and the predicted probability generated by the previous probit model. The second stage is: ℎ(𝑡 ) = ℎ 0 (𝑡 )𝑒𝑥𝑝 (𝐴𝐶 𝛽 𝐴𝐶 ̂ + 𝑋 0 𝛽 0 ̂ + 𝑋𝑢 ̂ 𝛽 𝑢 ̂ ) where 𝛽 𝐴𝐶 ̂ is the consistent estimation of the effect of AC on survival outcome. While this approach can generate asymptotically unbiased estimates of the true coefficient values, the standard errors cannot be obtained directly from the output based on a 15 statistical package. We used a bootstrapping approach to approximate the asymptotically correct standard errors for coefficients (500 replications). Our instrumental variables were constructed based on the treatment pattern in two levels: surgeons and health service areas. We identified the surgeon of record from claims data. We also grouped patients into health service areas that were defined by SEER. Then we calculated the monthly lagged and accumulated proportions of patients who received AC from patients who were treated by the same surgeon on record or in the same health service area. These instrumental variables generate a measure of treatment effects for “marginal patients”, like the patients who were indifferent to choosing between AC or observation, but their actual treatment choices were driven by the treatment preferences from the health service area or the surgeon that treated the patient.[22-24] Because unobservable characteristics cannot be “observed”, the validity of the key assumption that IVs are uncorrelated with unobservable factors cannot be verified directly. The development of these instruments was based on two criteria: first, whether they had a statistically significant impact on the acceptance of AC. Second, whether they balanced the observable characteristics. For the first criteria, we measured the joint significance of instrumental variables from the first-stage regression. For the second criteria, we split our cohort by the median value of each instrumental variable into two parts, respectively: above-median group and below-median group. Associations between the below/above group and demographic and clinical characteristics were analyzed. Tests of association were performed with the Chi-square tests. We plotted the non-parametric survival curves between patients who received AC and those who did not with the methods of Kaplan-Meier. We also generated the predicted survival curves based on a spline-based Royston-Parmar model.[25] We employed the parametric 16 standardized survival curves[26] to display the differences in survival outcomes under two counterfactual scenarios: everyone in the population has AC and everyone does not have AC. All statistical tests were two-sided and assessed for significance at the 5% level. We used STATA (StataCorp. 2017. Stata Statistical Software: Release 15. College Station, TX: StataCorp LLC) for the Survival analysis and SAS (SAS Institute Inc., Cary NC, USA) for all other analyses. Results We identified 16,468 patients with Stage II colon cancer between 2004 to 2011 who met eligibility criteria, 12.1% of them received AC (n=1,996). The median age of the cohort was 80 and the median age of the patients who received AC was 73. Compared with patients who did not receive AC, the recipients were statistically significantly younger, more likely to be male, non-white, and married. Moreover, they were more likely to have lower comorbidity scores, have left-sided tumors, advanced-stage, poorly differentiated or undifferentiated tumors, and fewer lymph nodes examined (Table 2.2). Table 2.2 Demographic and Clinical Characteristics for patients with Stage II Colon Cancer Diagnosed between 2004 and 2011 N=16,468 Without AC N=14,472 With AC N=1,998 P-value for Chi-square statistic Age at diagnosis 0.000 65-74 4890(29.7%) 3735(25.8%) 1155(57.9%) 75-85 7514(45.6%) 6755(46.7%) 759(38.0%) 86+ 4064(24.7%) 3982(27.5%) 82(4.1%) NCI Comorbidity Index 0.000 0 7597(46.1%) 6511(45.0%) 1086(54.4%) 1-3 5993(36.4%) 5306(36.7%) 687(34.4%) >3 2878(17.5%) 2655(18.3%) 223(11.2%) Gender 0.000 17 Male 7079(43.0%) 6100(42.2%) 979(49.0%) Female 9389(57.0%) 8372(57.8%) 1017(51.0%) Marital status 0.000 Married 7823(47.5%) 6624(45.8%) 1199(60.1%) Single, Separated, Divorced 2426(14.7%) 2094(14.5%) 332(16.6%) Windowed 5549(33.7%) 5145(35.6%) 404(20.2%) Unknown 668(4.1%) 607(4.2%) 61(3.1%) Race 0.012 White 14453(87.8%) 12734(88.0%) 1719(86.1%) Black 1295(7.9%) 1132(7.8%) 163(8.2%) Other 696(4.2%) 587(4.1%) 109(5.5%) Unknown 24(0.1%) 19(0.1%) 5(0.3%) Poverty Level 0.015 0%-<5% poverty 4105(24.9%) 3616(25.0%) 489(24.5%) 5% to <10% poverty 4432(26.9%) 3899(26.9%) 533(26.7%) 10% to <20% poverty 4710(28.6%) 4106(28.4%) 604(30.3%) 20% to 100% poverty 2927(17.8%) 2575(17.8%) 352(17.6%) Unknown 294(1.8%) 276(1.9%) 18(0.9%) Urban/Rural 0.004 Big Metro 8689(52.8%) 7580(52.4%) 1109(55.6%) Metro 4752(28.9%) 4207(29.1%) 545(27.3%) Urban 1048(6.4%) 954(6.6%) 94(4.7%) Less Urban 1600(9.7%) 1400(9.7%) 200(10.0%) Rural 378(2.3%) 330(2.3%) 48(2.4%) Region 0.001 west 6259(38.0%) 5576(38.5%) 683(34.2%) South 4029(24.5%) 3507(24.2%) 522(26.2%) Midwest 2531(15.4%) 2187(15.1%) 344(17.2%) Northeast 3649(22.2%) 3202(22.1%) 447(22.4%) Tumor Location 0.000 Left-sided colon cancer 4949(30.1%) 4209(29.1%) 740(37.1%) Right-sided colon cancer 11519(69.9%) 10263(70.9%) 1256(62.9%) Tumor Stage 0.000 IIA 14472(87.9%) 12911(89.2%) 1561(78.2%) IIB 1114(6.8%) 910(6.3%) 204(10.2%) IIC 882(5.4%) 651(4.5%) 231(11.6%) Tumor grade 0.001 well differentiated 1208(7.3%) 1067(7.4%) 141(7.1%) moderately differentiated 11673(70.9%) 10322(71.3%) 1351(67.7%) 18 poorly differentiated 2977(18.1%) 2559(17.7%) 418(20.9%) undifferentiated 296(1.8%) 248(1.7%) 48(2.4%) not determined 314(1.9%) 276(1.9%) 38(1.9%) Lymph node examined 0.000 <12 lymph node examined 5008(30.4%) 4297(29.7%) 711(35.6%) >=12 lymph node examined 11460(69.6%) 10175(70.3%) 1285(64.4%) Diagnosis Year 0.000 2004 2750(16.7%) 2300(15.9%) 450(22.5%) 2005 2693(16.4%) 2348(16.2%) 345(17.3%) 2006 2671(16.2%) 2300(15.9%) 371(18.6%) 2007 2353(14.3%) 2092(14.5%) 261(13.1%) 2008 2114(12.8%) 1899(13.1%) 215(10.8%) 2009 1419(8.6%) 1274(8.8%) 145(7.3%) 2010 1285(7.8%) 1178(8.1%) 107(5.4%) 2011 1183(7.2%) 1081(7.5%) 102(5.1%) Among recipients of AC, 43.4% (n=866) received an oxaliplatin-containing regimen. Patient characteristics are reported in Table 2.3. Table 2.3 Characteristics of recipients of oxaliplatin With AC N=1,996 Without oxaliplatin N=1,130 With oxaliplatin N=866 P-value for Chi-square statistic Age at diagnosis 0.000 65-74 1155(57.9%) 567(50.2%) 588(67.9%) 75-85 759(38.0%) 494(43.7%) 265(30.6%) 86+ 82(4.1%) 69(6.1%) 13(1.5%) NCI Comorbidity Index 0.027 0 1086(54.4%) 589(52.1%) 497(57.4%) 1-3 687(34.4%) 400(35.4%) 287(33.1%) >3 223(11.2%) 141(12.5%) 82(9.5%) Gender 0.429 Male 979(49.0%) 563(49.8%) 416(48.0%) Female 1017(51.0%) 567(50.2%) 450(52.0%) Marital status 0.475 Married 1199(60.1%) 667(59.0%) 532(61.4%) 19 Single, Separated, Divorced 332(16.6%) 185(16.4%) 147(17.0%) Windowed 404(20.2%) 242(21.4%) 162(18.7%) Unknown 61(3.1%) 36(3.2%) 25(2.9%) Race 0.745 White 1719(86.1%) 973(86.1%) 746(86.1%) Black 163(8.2%) 93(8.2%) 70(8.1%) Other 109(5.5%) 60(5.3%) 49(5.7%) Unknown 5(0.3%) 4(0.4%) 1(0.1%) Poverty Level 0.350 0%-<5% poverty 489(24.5%) 281(24.9%) 208(24.0%) 5% to <10% poverty 533(26.7%) 303(26.8%) 230(26.6%) 10% to <20% poverty 604(30.3%) 342(30.3%) 262(30.3%) 20% to 100% poverty 352(17.6%) 190(16.8%) 162(18.7%) Unknown 18(0.9%) 14(1.2%) 4(0.5%) Urban/Rural 0.683 Big Metro 1109(55.6%) 623(55.1%) 486(56.1%) Metro 545(27.3%) 312(27.6%) 233(26.9%) Urban 94(4.7%) 48(4.2%) 46(5.3%) Less Urban 200(10.0%) 120(10.6%) 80(9.2%) Rural 48(2.4%) 27(2.4%) 21(2.4%) Region 0.027 west 683(34.2%) 372(32.9%) 311(35.9%) South 522(26.2%) 278(24.6%) 244(28.2%) Midwest 344(17.2%) 206(18.2%) 138(15.9%) Northeast 447(22.4%) 274(24.2%) 173(20.0%) Tumor Location 0.020 Left-sided colon cancer 740(37.1%) 394(34.9%) 346(40.0%) Right-sided colon cancer 1256(62.9%) 736(65.1%) 520(60.0%) Tumor Stage 0.000 IIA 1561(78.2%) 944(83.5%) 617(71.2%) IIB 204(10.2%) 95(8.4%) 109(12.6%) IIC 231(11.6%) 91(8.1%) 140(16.2%) Tumor grade 0.954 well differentiated 141(7.1%) 79(7.0%) 62(7.2%) moderately differentiated 1351(67.7%) 773(68.4%) 578(66.7%) poorly differentiated 418(20.9%) 230(20.4%) 188(21.7%) undifferentiated 48(2.4%) 27(2.4%) 21(2.4%) not determined 38(1.9%) 21(1.9%) 17(2.0%) Lymph node examined 0.027 20 <12 lymph node examined 711(35.6%) 426(37.7%) 285(32.9%) >=12 lymph node examined 1285(64.4%) 704(62.3%) 581(67.1%) Diagnosis Year 0.000 2004 450(22.5%) 360(31.9%) 90(10.4%) 2005 345(17.3%) 224(19.8%) 121(14.0%) 2006 371(18.6%) 194(17.2%) 177(20.4%) 2007 261(13.1%) 113(10.0%) 148(17.1%) 2008 215(10.8%) 99(8.8%) 116(13.4%) 2009 145(7.3%) 62(5.5%) 83(9.6%) 2010 107(5.4%) 39(3.5%) 68(7.9%) 2011 102(5.1%) 39(3.5%) 63(7.3%) The use of AC declined over time: from 16.4% in 2004 to 8.6% in 2011 (Figure 2.2). However, among patients who received AC, the percentages of receipt of oxaliplatin increased from 20.0% in 2004 to 61.8% in 2011. Figure 2.2 The use of AC over time 21 The Kaplan-Meier survival curves are shown in Figure 2.3. The 3-year overall survival rate was 81.0% in the AC group versus 73.1% in the untreated group. The Cox model indicated that the hazard ratio of AC while adjusting for SEER variables was 0.85 (95% CI:0.78 to 0.92). The point estimator of the E-value for AC was 1.49 (upper limit of the confidence interval estimator: 1.32). This means that an unmeasured covariate beyond SEER covariates could bias the AC treatment benefit as measured in COX model with a relative risk association of at least 1.49 with both AC and survival outcomes. The E-value for the upper limit of the confidence interval was 1.32. Based on hazard ratios for covariates from the Cox model, the E-values for some known factors that associated with survival outcome were 1.79 (CI:1.69) for being a female, 1.57 (CI:1.47) for being married, and 2.06 (CI:1.93) for having advanced stage tumor (Table 2.4). It is likely that there is some unobservable confounder that would have a greater effect on survival outcome and AC by having an E-value over 1.49. Thus, it is possible that some unmeasured confounding could bias these findings on AC treatment effects. This analysis endorses the necessity of performing an instrumental variable regression analysis. Table 2.4 E-value E-Value (Point Estimator) E-Value (Confidence Interval Estimator) AC 1.49 1.32 Gender (Ref: male) female 1.79 1.69 Marital (Ref: Not married) married 1.57 1.47 Tumor stage (Ref: stage IIB/C) stage IIA 2.06 1.93 Tumor site (Ref: left) right 1.25 1.08 # of lymph nodes examined: (Ref: < 12) 22 >= 12 1.59 1.49 Tumor grade (Ref: well/moderately) poorly/not 1.30 1.15 Poverty level (Ref:5%-100% poverty) 0%-5% poverty 1.29 1.14 Using instrumental variables analysis, the estimated risk of mortality trended higher in those who received AC (HR=1.78, 95% CI: 0.93 to 3.41). The predicted standard survival curves (Figure 2.4) indicated that the survival improvements due to AC was uncertain. The 95% CI of predicted survival curves for patients who used AC was wide and the upper parts overlapped with the survival curves for patients who didn’t receive AC. For instrumental variable analysis, our sample was not the same as the previous Cox model since the patients who were treated in January 2004 did not have data about previous treatment patterns for either surgeon or health service area (152 patients were removed from the analysis). Results from unadjusted and instrumental variable estimates of the Cox proportional hazard model are displayed in Table 2.5. 23 Figure 2.3 Kaplan Meier survival curves Figure 2.4 Standardized survival curves (with 95% CI) between patients who would have received AC and patients who would not have AC 24 Table 2.5 Survival Analysis with Cox proportional hazard model and instrumental variable regression analysis Cox Model N=16,468 IV Analysis + Cox Model N=16,316 HR P value HR (95% CI) HR P value HR (95% CI) Adjuvant chemotherapy 0.85 0.000 ( 0.78 , 0.92 ) 1.78 0.083 ( 0.93 , 3.41 ) Age at diagnosis 1.06 0.000 ( 1.05 , 1.06 ) 1.06 0.000 ( 1.06 , 1.07 ) NCI index 1.21 0.000 ( 1.19 , 1.22 ) 1.22 0.000 ( 1.20 , 1.23 ) Gender (ref: male) female 0.73 0.000 ( 0.70 , 0.77 ) 0.73 0.000 ( 0.69 , 0.77 ) Marital status (ref: not married) married 0.82 0.000 ( 0.78 , 0.86 ) 0.80 0.000 ( 0.76 , 0.84 ) Race (ref: non-white) white 1.04 0.249 ( 0.97 , 1.12 ) 1.04 0.328 ( 0.96 , 1.13 ) Tumor stage (ref: stage IIB/C) stage IIA 0.64 0.000 ( 0.60 , 0.68 ) 0.69 0.000 ( 0.62 , 0.75 ) Tumor site (ref: left) right 0.94 0.022 ( 0.90 , 0.99 ) 0.96 0.147 ( 0.91 , 1.01 ) # of lymph nodes examined: (ref: < 12) >= 12 0.81 0.000 ( 0.77 , 0.85 ) 0.83 0.000 ( 0.78 , 0.87 ) Tumor grade (ref: well/moderately) poorly/not 1.08 0.004 ( 1.03 , 1.15 ) 1.07 0.045 ( 1.00 , 1.13 ) Poverty level (ref:5%-100% poverty) 0%-5% poverty 0.93 0.005 ( 0.88 , 0.98 ) 0.93 0.006 ( 0.88 , 0.98 ) Region (ref: West) South 1.15 0.000 ( 1.09 , 1.22 ) 1.15 0.000 ( 1.09 , 1.22 ) Midwest 1.02 0.561 ( 0.95 , 1.09 ) 0.99 0.882 ( 0.93 , 1.07 ) Northeast 1.05 0.135 ( 0.99 , 1.11 ) 1.03 0.340 ( 0.97 , 1.10 ) Residual term NA NA NA 0.47 0.025 ( 0.24 , 0.91 ) 25 The strength of the instrumental variables was examined by their joint significance in the first-stage equation. Our instrumental variables were highly statistically significant in the first- stage equation: the Chi-square score is 26.7 (P<0.001). Table 2.6 display how our instrumental variables balanced the observed characteristics. Although the Chi-square tests indicated that there were still some differences, grouping patients by the value of the instruments, indeed, narrowed the differences in the observed characteristics. Moreover, the residual generated from the first stage was significant (p=0.025) which is equivalent to Hausman tests of endogeneity in the linear setting. Table 2.6 Comparison of samples grouped by the median value of instrumental variables Health Service Area Provider Total N=16,316 Below Median N=8,186 Above Median N=8,130 Chi- square statistic Total N=16,316 Below Median N=8,257 Above Median N=8,059 Chi- square statistics Age at diagnosis 0.537 0.602 65-74 4845(29.70%) 2413(29.50%) 2432(29.90%) 4845(29.70%) 2481(30.00%) 2364(29.30%) 75-85 7445(45.60%) 3723(45.50%) 3722(45.80%) 7445(45.60%) 3745(45.40%) 3700(45.90%) 86+ 4026(24.70%) 2050(25.00%) 1976(24.30%) 4026(24.70%) 2031(24.60%) 1995(24.80%) NCI Comorbidity Index 0.000 0.000 0 7525(46.10%) 3915(47.80%) 3610(44.40%) 7525(46.10%) 3960(48.00%) 3565(44.20%) 1-3 5938(36.40%) 2958(36.10%) 2980(36.70%) 5938(36.40%) 2938(35.60%) 3000(37.20%) >3 2853(17.50%) 1313(16.00%) 1540(18.90%) 2853(17.50%) 1359(16.50%) 1494(18.50%) Gender 0.145 0.001 Male 7024(43.00%) 3478(42.50%) 3546(43.60%) 7024(43.00%) 3451(41.80%) 3573(44.30%) Female 9292(57.00%) 4708(57.50%) 4584(56.40%) 9292(57.00%) 4806(58.20%) 4486(55.70%) Marital status 0.000 0.002 Married 7754(47.50%) 3906(47.70%) 3848(47.30%) 7754(47.50%) 3893(47.20%) 3861(47.90%) Single, Separated, Divorced 2406(14.70%) 1187(14.50%) 1219(15.00%) 2406(14.70%) 1248(15.10%) 1158(14.40%) Windowed 5493(33.70%) 2818(34.40%) 2675(32.90%) 5493(33.70%) 2824(34.20%) 2669(33.10%) Unknown 661(4.10%) 274(3.30%) 387(4.80%) 661(4.10%) 291(3.50%) 370(4.60%) Race 0.000 0.000 White 14318(87.80%) 7288(89.00%) 7030(86.50%) 14318(87.80%) 7227(87.50%) 7091(88.00%) Black 1286(7.90%) 528(6.50%) 758(9.30%) 1286(7.90%) 614(7.40%) 672(8.30%) Other 689(4.20%) 355(4.30%) 334(4.10%) 689(4.20%) 400(4.80%) 289(3.60%) Unknown 23(0.10%) 15(0.20%) 8(0.10%) 23(0.10%) 16(0.20%) 7(0.10%) Poverty Level 0.546 0.000 0%-<5% poverty 4058(24.90%) 2020(24.70%) 2038(25.10%) 4058(24.90%) 1871(22.70%) 2187(27.10%) 26 5% to <10% poverty 4405(27.00%) 2186(26.70%) 2219(27.30%) 4405(27.00%) 2171(26.30%) 2234(27.70%) 10% to <20% poverty 4663(28.60%) 2362(28.90%) 2301(28.30%) 4663(28.60%) 2440(29.60%) 2223(27.60%) 20% to 100% poverty 2898(17.80%) 1460(17.80%) 1438(17.70%) 2898(17.80%) 1606(19.50%) 1292(16.00%) Unknown 292(1.80%) 158(1.90%) 134(1.60%) 292(1.80%) 169(2.00%) 123(1.50%) Urban/Rural 0.000 0.000 Big Metro 8613(52.80%) 3308(40.40%) 5305(65.30%) 8613(52.80%) 3773(45.70%) 4840(60.10%) Metro 4705(28.80%) 2955(36.10%) 1750(21.50%) 4705(28.80%) 2491(30.20%) 2214(27.50%) Urban 1038(6.40%) 744(9.10%) 294(3.60%) 1038(6.40%) 685(8.30%) 353(4.40%) Less Urban 1584(9.70%) 948(11.60%) 636(7.80%) 1584(9.70%) 1059(12.80%) 525(6.50%) Rural 375(2.30%) 230(2.80%) 145(1.80%) 375(2.30%) 248(3.00%) 127(1.60%) Region 0.000 0.000 west 6199(38.00%) 3407(41.60%) 2792(34.30%) 6199(38.00%) 3579(43.30%) 2620(32.50%) South 3986(24.40%) 1810(22.10%) 2176(26.80%) 3986(24.40%) 1889(22.90%) 2097(26.00%) Midwest 2512(15.40%) 826(10.10%) 1686(20.70%) 2512(15.40%) 1233(14.90%) 1279(15.90%) Northeast 3619(22.20%) 2143(26.20%) 1476(18.20%) 3619(22.20%) 1556(18.80%) 2063(25.60%) Tumor Location 0.518 0.522 Left-sided colon cancer 4903(30.10%) 2441(29.80%) 2462(30.30%) 4903(30.10%) 2500(30.30%) 2403(29.80%) Right-sided colon cancer 11413(69.90%) 5745(70.20%) 5668(69.70%) 11413(69.90%) 5757(69.70%) 5656(70.20%) Tumor Stage 0.123 0.076 IIA 14340(87.90%) 7153(87.40%) 7187(88.40%) 14340(87.90%) 7216(87.40%) 7124(88.40%) IIB 1100(6.70%) 580(7.10%) 520(6.40%) 1100(6.70%) 592(7.20%) 508(6.30%) IIC 876(5.40%) 453(5.50%) 423(5.20%) 876(5.40%) 449(5.40%) 427(5.30%) Tumor grade 0.000 0.011 well differentiated 1194(7.30%) 604(7.40%) 590(7.30%) 1194(7.30%) 624(7.60%) 570(7.10%) moderately differentiated 11570(70.90%) 5717(69.80%) 5853(72.00%) 11570(70.90%) 5865(71.00%) 5705(70.80%) poorly differentiated 2943(18.00%) 1531(18.70%) 1412(17.40%) 2943(18.00%) 1478(17.90%) 1465(18.20%) undifferentiated 296(1.80%) 182(2.20%) 114(1.40%) 296(1.80%) 122(1.50%) 174(2.20%) not determined 313(1.90%) 152(1.90%) 161(2.00%) 313(1.90%) 168(2.00%) 145(1.80%) Lymph node examined 0.681 0.000 <12 lymph node examined 4929(30.20%) 2485(30.40%) 2444(30.10%) 4929(30.20%) 2784(33.70%) 2145(26.60%) >=12 lymph node examined 11387(69.80%) 5701(69.60%) 5686(69.90%) 11387(69.80%) 5473(66.30%) 5914(73.40%) Diagnosis Year 0.000 0.000 2004 2598(15.90%) 1479(18.10%) 1119(13.80%) 2598(15.90%) 1833(22.20%) 765(9.50%) 2005 2693(16.50%) 1383(16.90%) 1310(16.10%) 2693(16.50%) 1578(19.10%) 1115(13.80%) 2006 2671(16.40%) 1342(16.40%) 1329(16.30%) 2671(16.40%) 1348(16.30%) 1323(16.40%) 2007 2353(14.40%) 1124(13.70%) 1229(15.10%) 2353(14.40%) 1065(12.90%) 1288(16.00%) 2008 2114(13.00%) 1048(12.80%) 1066(13.10%) 2114(13.00%) 914(11.10%) 1200(14.90%) 2009 1419(8.70%) 685(8.40%) 734(9.00%) 1419(8.70%) 572(6.90%) 847(10.50%) 2010 1285(7.90%) 601(7.30%) 684(8.40%) 1285(7.90%) 488(5.90%) 797(9.90%) 2011 1183(7.30%) 524(6.40%) 659(8.10%) 1183(7.30%) 459(5.60%) 724(9.00%) 27 Discussion Our results of the instrumental variable regression analysis suggest that it is unlikely that AC provides a survival benefit in the elderly Medicare population. This is despite the fact that recipients of AC were younger and healthier (lower comorbidity scores) and had higher risk disease (more advanced stage, poorly/not differentiated tumors). The findings are in line with other published studies, including three SEER-Medicare analyses [13-15] and a systematic review[27], in the era of 5-FU based chemotherapy. We further show that despite a decline in the use of AC, the aggressiveness of treatment (defined as the use of oxaliplatin-containing regimen) is increasing. To our knowledge, this is the first instrumental variable analysis based on a large population-based cohort for addressing observed and unobserved confounders in the treatment of stage II colon cancer. We took advantage of the SEER-Medicare database variation in lagged treatment patterns across 187 different health service areas in developing our instrumental variables. Several studies have produced similar instrumental variables that were based on the variation across health service areas.[28-30] We also developed one instrumental variable from the surgeon of record. Although surgeons are not the decision-maker for AC, their referral pattern to a medical oncologist may influence the treatment delivery. The strong associations between our instrumental variables and treatment choices indicated the good reliability of our instrumental variables. While in the Cox model the hazard ratio of AC was 0.85 (95% CI: 0.78 to 0.92), the hazard ration in the instrumental variable model was 1.78 (95% CI: 0.93 to 3.42). This discrepancy is attributable to differences in the area and surgeon level treatment preference controlled by the instrumental analysis. 28 The results of our analysis contrast with the QUASAR trial. Several patients and disease- related factors may explain this difference. Firstly, the median age in QUASAR enrollees was 63 (IQR 56-68), and in our cohort was 79 (IQR 73-84)—the subgroup analysis from the QUASAR trial indicated that patients younger than 70 years were more likely to benefit from AC. Secondly is the adequacy of staging, 64% of patients enrolled in the QUASAR had fewer than 12 lymph nodes examined, in contrast, 70% of our patients had more than 12 lymph nodes examined. Inadequately staged patients may be stage III and thus more likely to benefit from chemotherapy.[31] Lastly, the QUASAR trial included patients with rectal cancer, while our cohort was only composed of patients with colon cancer. Our study is also in conflict with a retrospective analysis from the National Cancer Database (NCDB). In that study, 153,110 stage II colon cancer patients that were diagnosed from 1998 to 2006 were identified and adjuvant treatment was associated with improved survival (HR, 0.76; P < 0.001) adjusted for age.[32] However, NCDB data only has the intent for chemotherapy and the number of cycles is not one of their data variables. Additionally, the authors did not limit the timing of chemotherapy, so it is possible that some patients with metastatic disease were included in their cohort. In contrast, we only included patients who started adjuvant therapy within 120 days from their date of surgery and the median duration of chemotherapy, based on claims, was 5.1 months (IQR 2.8-5.7). SEER-Medicare and NCDB studies were observational and the potential for bias limits the causal relationship between AC and survival outcome. Even though they all employed the propensity score matching (PSM) to decrease the impact of selection bias, the method can only create matched pairs based on observed characteristics. If unobservable factors were a major source of bias, then bias would remain an issue and PSM would not eliminate the bias. 29 Moreover, researchers using simulation design reported that PSM could increase imbalance, model dependence, and bias.[33, 34] In principle, the instrumental variable analysis is more robust than the propensity score matching since it adjusts for both observable and unobservable potential sources of bias. Our study has several limitations. First, our sample only included patients who were older than 65 years, which limits the generalizability of our findings to the younger population. Second, we did not have information about patients’ microsatellite instability (MSI) status. MSI is an important factor in treatment decisions for AC in patients with stage II colon cancer, since patients with MSI-H tumors may not benefit from 5-FU AC.[35] Third, the instrumental variable analysis does not guarantee that all potential bias has been eliminated. In conclusion, our study is the first to incorporate the instrumental variable analysis in investigating the effect of AC among patients with stage II colon cancer using the SEER- Medicare dataset. We show no definitive survival benefit for AC with fluoropyrimidine with or without oxaliplatin in the elderly patients with stage II disease. Future studies need to focus on the identification of the elderly population with stage II who may benefit from AC. 30 References 1. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA: a cancer journal for clinicians 2019;69(1):7-34. 2. André T, Boni C, Mounedji-Boudiaf L, et al. Oxaliplatin, fluorouracil, and leucovorin as adjuvant treatment for colon cancer. New England Journal of Medicine 2004;350(23):2343-2351. 3. André T, Boni C, Navarro M, et al. Improved overall survival with oxaliplatin, fluorouracil, and leucovorin as adjuvant treatment in stage II or III colon cancer in the MOSAIC trial. J Clin Oncol 2009;27(19):3109-3116. 4. Schmoll H-J, Tabernero J, Maroun J, et al. Capecitabine plus oxaliplatin compared with fluorouracil/folinic acid as adjuvant therapy for stage III colon cancer: final results of the NO16968 randomized controlled phase III trial. Journal of Clinical Oncology 2015;33(32):3733-3740. 5. Tournigand C, André T, Bonnetain F, et al. Adjuvant therapy with fluorouracil and oxaliplatin in stage II and elderly patients (between ages 70 and 75 years) with colon cancer: subgroup analyses of the Multicenter International Study of Oxaliplatin, Fluorouracil, and Leucovorin in the Adjuvant Treatment of Colon Cancer trial. Journal of clinical oncology 2012;30(27):3353-3360. 6. Benson Iii AB, Schrag D, Somerfield MR, et al. American Society of Clinical Oncology recommendations on adjuvant chemotherapy for stage II colon cancer. Journal of clinical oncology 2004;22(16):3408-3419. 7. Quasar Collaborative G. Adjuvant chemotherapy versus observation in patients with colorectal cancer: a randomized study. The Lancet 2007;370(9604):2020-2029. 8. Denicoff AM, McCaskill-Stevens W, Grubbs SS, et al. The National Cancer Institute–American Society of Clinical Oncology cancer trial accrual symposium: summary and recommendations. Journal of oncology practice 2013;9(6):267-276. 9. Van Spall HGC, Toren A, Kiss A, et al. Eligibility criteria of randomized controlled trials published in high-impact general medical journals: a systematic sampling review. Jama 2007;297(11):1233-1240. 10. National Cancer Institute. Overview of the SEER Program. https://seer.cancer.gov/about/overview.html. 11. Warren JL, Klabunde CN, Schrag D, et al. Overview of the SEER-Medicare data: content, research applications, and generalizability to the United States elderly population. Medical care 2002:IV3-IV18. 12. Howlader NN, Krapcho AM, Miller M, et al. KA (Eds.)(2016). SEER cancer statistics review, 1975– 2013, National Cancer Institute. Bethesda, MD. In; 2016. 13. Schrag D, Rifas-Shiman S, Saltz L, et al. Adjuvant chemotherapy use for Medicare beneficiaries with stage II colon cancer. Journal of clinical oncology 2002;20(19):3999-4005. 14. O'Connor ES, Greenblatt DY, LoConte NK, et al. Adjuvant chemotherapy for stage II colon cancer with poor prognostic features. Journal of clinical oncology 2011;29(25):3381. 15. 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. Annals of surgical oncology 2014;21(6):1781-1791. 16. Giordano SH, Kuo YF, Duan Z, et al. Limits of observational data in determining outcomes from cancer therapy. Cancer: Interdisciplinary International Journal of the American Cancer Society 2008;112(11):2456-2466. 17. Fritz A PC, Jack A, et al. International Classification of Diseases for Oncology (ed 3). In. Geneva, Switzerland: World Health Organization, 2000 18. Klabunde CN, Potosky AL, Legler JM, et al. Development of a comorbidity index using physician claims data. Journal of clinical epidemiology 2000;53(12):1258-1267. 19. Haneuse S, VanderWeele TJ, Arterburn D. Using the E-value to assess the potential effect of unmeasured confounding in observational studies. Jama 2019;321(6):602-603. 31 20. VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Annals of internal medicine 2017. 21. Cameron AC, Trivedi PK. Microeconometrics: methods and applications: Cambridge university press; 2005. 22. Angrist JD, Imbens GW, Rubin DB. Identification of causal effects using instrumental variables. Journal of the American statistical Association 1996;91(434):444-455. 23. Heckman J. Instrumental variables: A study of implicit behavioral assumptions used in making program evaluations. Journal of human resources 1997;32(3):441-462. 24. Harris KM, Remler DK. Who is the marginal patient? understanding instrumental variables estimates of treatment effects. Health services research 1998;33(5 Pt 1):1337. 25. Royston P, Lambert PC. Flexible parametric survival analysis using Stata: beyond the Cox model: Stata College Station, Texas; 2011. 26. Lambert P. STPM2_STANDSURV: Stata module to obtain standardized survival curves after fitting an stpm2 survival model. 2018. 27. Figueredo A, Charette ML, Maroun J, et al. Adjuvant therapy for stage II colon cancer: a systematic review from the Cancer Care Ontario Program in evidence-based care’s gastrointestinal cancer disease site group. Journal of clinical oncology 2004;22(16):3395-3407. 28. Hadley J, Yabroff KR, Barrett MJ, et al. Comparative effectiveness of prostate cancer treatments: evaluating statistical adjustments for confounding in observational data. Journal of the National Cancer Institute 2010;102(23):1780-1793. 29. McDowell BD, Chapman CG, Smith BJ, et al. Pancreatectomy predicts improved survival for pancreatic adenocarcinoma: results of an instrumental variable analysis. Annals of surgery 2015;261(4):740. 30. Xu H, Xia Z, Jia X, et al. Primary tumor resection is associated with improved survival in stage IV colorectal cancer: an instrumental variable analysis. Scientific reports 2015;5:16516. 31. Hutchins G, Southward K, Handley K, et al. Value of mismatch repair, KRAS, and BRAF mutations in predicting recurrence and benefits from chemotherapy in colorectal cancer. J Clin Oncol 2011;29(10):1261-1270. 32. Casadaban L, Rauscher G, Aklilu M, et al. Adjuvant chemotherapy is associated with improved survival in patients with stage II colon cancer. Cancer 2016;122(21):3277-3287. 33. Brooks JM, Ohsfeldt RL. Squeezing the Balloon: Propensity Scores and Unmeasured Covariate Balance. Health Services Research 2013;48(4):1487-1507. 34. King G, Nielsen R. Why propensity scores should not be used for matching. Copy at http://j. mp/1sexgVw Download Citation BibTex Tagged XML Download Paper 2016;378. 35. Sargent DJ, Marsoni S, Monges G, et al. Defective mismatch repair as a predictive marker for lack of efficacy of fluorouracil-based adjuvant therapy in colon cancer. Journal of Clinical Oncology 2010;28(20):3219. 32 Chapter 3: One discrete choice experiment study measuring the treatment preferences among patients with NSCLC (Non-Small Cell Lung Cancer) Abstract Introduction The treatment decisions for NSCLC (Non-Small Cell Lung Cancer) have become more complex and require more trade-offs from patients’ perspectives. Understanding patients’ treatment preferences can help to optimize their treatments and support the best practices of cancer care. Using a discrete choice experiment (DCE), we evaluated the relative weights that patients placed on one NSCLC treatment based on specific characteristics. We also assessed patients’ willingness to pay (WTP) for several treatment improvements. Methods We designed a discrete choice experiment survey and interviewed NSCLC patients at Norris Comprehensive Cancer Center (Los Angeles). In the survey, each patient was presented with 18 hypothetical scenarios. In each scenario, two treatment profiles were compared across the six following dimensions: progression-free survival time, monthly out-of-pocket cost, control of symptoms, and three treatment-related adverse effects: tiredness, skin itching, and hair loss. The responses were analyzed using a mixed logit model to elicit the relative importance of each treatment attribute and to calculate patients’ willingness to pay for each attribute improvement. We also estimated patients’ WTP for osimertinib after the patents of erlotinib or gefitinib expire. Results We collected 71 responses. All treatment attributes were found to be significant in patients’ decision-making. Progression-free survival time was the most influential factor. Furthermore, the treatment’s ability to control symptoms and out-of-pocket cost were also important factors. Among adverse effects, severe skin-itching was the principal barrier to treatment. Patients were willing to pay $ 3,290 (95% CI: $2,550 to $4,030) for a one-month increase in progression-free survival time. Based on our estimated coefficients, we predicted patients would be willing to pay up to $ 1,350 per month for osimertinib. Conclusions This study demonstrates the utility of one NSCLC treatment depends on the improvements in progression-free survival, the severity of disease symptoms and adverse effects, and the amount of out-of-pocket cost. 33 Introduction Lung cancer is the leading cause of cancer-related death in both men and women in the United States. In 2019, the number of new cases of lung cancer is estimated to be 228,150 and nearly 142,670 people will die of this disease.[1] More than 86% of lung cancer patients have non-small cell lung cancer (NSCLC)[2], which is a group of diseases with significant heterogeneity. Thanks to the cutting-edge developments in technology, several therapeutic targets have been discovered. Nowadays, patients with NSCLC can be provided with multiple treatment options. Usually, treatment decisions are made by physicians alone, or occasionally, by physicians and patients together. To further improve patient-centered care, physicians should take into consideration patients’ treatment preferences. A crucial step towards understanding patients’ treatment preferences is to determine how patients evaluate treatment characteristics when they make decisions. This information can aid physicians when counseling their patients. The treatment of patients with NSCLC and locally advanced EGFR-sensitizing mutations had changed significantly since the discovery of “epidermal growth factor receptor” (EGFR). For these patients, based on clinical trials that showed improved survival outcomes compared with platinum-based chemotherapy, EGFR tyrosine kinase inhibitors (erlotinib, gefitinib, afatinib, dacomitinib, and osimertinib) were recommended as first-line therapy[3-7]. In a recent phase 3 trial (FLAURA), the third generation EGFT TKI—osimertinib—was compared with the first- generation TKI—either erlotinib or gefitinib—in the first-line therapy. With longer PFS (osimertinib vs erlotinib/gefitinib: 18.9 months vs 10.2 months), the National Comprehensive Cancer Network (NCCN) labeled osimertinib as “preferred”[8]. However, the price of osimertinib is extremely high. One cost-effectiveness analysis (CEA) indicated that, when 34 compared to erlotinib/gefitinib, osimertinib was not cost-effective in neither the United States or Brazil[9]. Moreover, in the US, the patent of erlotinib and the exclusivity of gefitinib for the first-line treatment of patients with NSCLC will expire on Nov 9, 2020, and July 13, 2022, respectively[10]. The prices of these two drugs are expected to decline rapidly once the generic drugs enter the market. If the acquisition cost of osimertinib remains substantial, the incremental cost-effectiveness ratio (ICER) of this drug will become more unacceptable. However, the CEA study was conducted based on the payer’s perspective. The valuation of osimertinib from patients’ perspectives remains unknown. This study aimed at quantitatively presenting the impacts of NSCLC treatment characteristics on patients’ treatment decision-making. We implemented one discrete choice experiment, which is known as one stated preference approach that provides information on the relative merits of complex outcomes. Previous studies had utilized DCE to measure treatment preferences for NSCLC in the United Kingdom[11] and Germany[12]. However, the DCE study related to NSCLC treatment in the United States is limited. This DCE was designed to determine the relative importance of treatment characteristics in NSCLC patients’ decision-making. We estimated NSCLC patients’ willingness to pay for each incremental improvement provided by the treatment. Based on the expressed preferences, we also predicted patients’ willingness to pay for osimertinib after the patents of erlotinib and gefitinib expire. Methods Discrete choice experiment Discrete choice experiment (DCE) is a stated preference technique increasingly used for estimating health-related preferences.[13] It combines Lancaster’s consumer theory[14] and 35 McFadden’s random utility maximization (RUM) theory[15]. The DCE postulates that a good or service can be fully described by its attributes (characteristics) and these attributes can be further distinguished by various levels or ranks. The RUM assumes that a decision-maker (in our case, the NSCLC patient) maximizes his/her utility by systematically choosing one alternative (treatment choice) which provides the highest utility; this utility is made up of the combination of observable attributes and associated levels, along with unobservable idiosyncratic error. To be specific, in a given DCE, participants are presented with several hypothetical scenarios (choice sets) that contain several choice options. Each option is illustrated by several attributes with pre- specified levels. In each scenario, patients are asked to select a single option from each choice set. Then the chosen alternative is assumed to have the highest utility for that participant, based on both observable and unobservable components. By analyzing a set of respondents’ resulting data using econometric models pioneered by McFadden, the relative importance of each observable attribute and its associated levels can be evaluated. There are several well-organized guidelines available for constructing DCEs[16-19]. Our study was designed and implemented with the help of these guidelines. After determining research objectives, we developed attributes and levels, constructed experimental design, conducted pilot-testing, revised survey questionnaires, collected responses, performed econometric analysis and interpreted the results. Attribute and level development Identifying attributes and their levels is a fundamental step in a DCE since participants (in our case, patients) use this information to evaluate the utility for each option. Due to task complexity, only a few attributes and levels can be included. The validity of DCE depends on 36 how complex information about health policies or interventions is transformed into a limited number of relevant attributes.[20] The results of DCEs are likely to be biased or meaningless if the research team fails to identify and include key attributes and levels[19]. In our study, the development of attributes and levels was based on a comprehensive literature review, discussion with experts, twelve one-on-one short survey-based interviews and one focus group with NSCLC patients. A systematic literature search was conducted (PubMed, Cochrane, Web of Science, and Google Scholar) to identify studies that have used a range of methodologies to assess preference for NSCLC treatment attributes. The literature review generated a list of potential attributes that influenced NSCLC patients’ stated preferences for their treatment. The possible attributes were classified into four broad groups: survival outcome, health conditions (disease symptoms and treatment-related adverse effects), out-of-pocket cost, and administration mode. The treatment attributes extracted from the literature search were then tested through interviews with two oncologists. Progression-free survival time, monthly out-of-pocket cost, and symptoms including coughing, chest pain and shortness of breath were included. Next, we interviewed 12 NSCLC patients on a one-on-one basis and asked them to rank the treatment- related adverse effects from the most to least important to their treatment decision making. Using patients’ rankings, we included three treatment-related adverse effects: tiredness, skin itching, and hair loss. The levels of progression-free survival time and monthly out-of-pocket cost were developed through discussions with oncologists and health economists. By reviewing the National Cancer Institute’s Common Terminology Criteria for Adverse Events (CTCAE) as guidance,[21] levels for adverse effects were derived. 37 In a subsequent focus group discussion with four NSCLC patients, previously identified treatment attributes were confirmed. Also, patients were encouraged to name any additional relevant treatment attributes. These subjects also validated the specified levels of each attribute and the comprehensibility of attribute and level descriptions. The final overview of the treatment attributes, associated levels, graphical icons, and related descriptions are delineated in Figure 3.1. Figure 3.1 Overview of treatment attributes and levels Experimental Design Since we planned to use a generic design in this experiment (treatments were presented without actual drug names or brands), there were 2^1* 4^1 * 3^4 = 648 possible combinations for one potential treatment alternative (four three-level attributes, one four-level attribute, and one two-level attribute). These combinations yielded 209,628 (648*(648-1) *1/2) possible 38 scenarios in comparing two different treatment alternatives. It is not possible to present a single participant with 209,628 choice tasks, we decided to employ the fractional factorial design to select a fraction of the total number of treatment combinations.[22] According to our pilot testing among 16 participants, the number of scenarios in the survey was set to 18. With the help of SAS’s (SAS Institute Inc., Cary, NC, USA) built-in macros, and by following instructions from a SAS support report, we implemented the computerized search algorithm to construct the choice designs which optimize D-efficiency. [23] We employed the “optimal generic design”[23] methods to create a design by shifting alternatives in orthogonal arrays. This method helped to minimize the overlaps of levels between treatment alternatives, which required participants to consider all trade-offs across attributes in their treatment decisions. To be noted, the design contained one dominant alternative in a choice set, in which treatment A is better on all attribute levels when compared with treatment B. This alternative served as an internal validity test to evaluate participants’ comprehensibility of the study. Study Sample and setting We surveyed 71 patients with NSCLC who visited the Norris Comprehensive Cancer Center (Los Angeles) from April to June 2018. Patients were invited to participate in the survey after their treatment consultation. The survey was presented to patients with one survey software called SurveyMonkey (iPad tablet-assisted). The 18 choice tasks were presented to each patient in a randomized order to minimize order bias. An example of one choice task is illustrated in Figure 3.2. This study was approved by the University of Southern California Health Sciences Campus Institutional Review Board. 39 Figure 3.2 An example of treatment choice task Data analysis Respondents’ demographic characteristics including age, gender, race, and income level were assessed using descriptive statistics. Respondents’ time to complete the survey was also 40 recorded. The frequency of selecting the dominant alternative among all respondents was calculated. The DCE was presented to the respondent with three options: treatment A, treatment B and no treatment at all. However, respondents seldom chose the “no treatment at all” option (only 5 of 71 patients ever selected the “no treatment at all” option). It is consistent with the fact that the survey respondents were patients who were already being treated, thus implicitly having a high preference for “treatment” rather than “no treatment at all”. One patient’s comment revealed this preference clearly: “There is nothing in the world that can get me to give up my treatment”. Therefore, the DCE was analyzed in a setting of the binary dependent variable. A mixed logit model was employed to understand the heterogeneous treatment preferences across respondents. Compared with the traditional conditional logit model, the mixed logit model allows the parameters of attributes to vary across the population and relaxes the assumption of independence of irrelevant alternatives (IIA).[24]Assuming that all attributes have an independent impact on patient’s treatment preference, the main-effects model below was estimated: 𝑈 𝑛𝑖 = 𝑉 𝑛𝑖 + 𝜖 𝑛𝑖 = 𝛽 1 ∗ 𝑃𝐹𝑆 + 𝛽 2 ∗ 𝐶𝑜𝑠𝑡 + 𝛽 3 ∗ 𝑆𝑦𝑚𝑝𝑡𝑜 𝑚 𝑆𝑒𝑣𝑒𝑟𝑒 + 𝛽 4 ∗ 𝑆𝑦𝑚𝑝𝑡𝑜 𝑚 𝑀𝑜𝑑𝑒𝑟𝑎𝑡𝑒 + 𝛽 5 ∗ 𝑇𝑖𝑟𝑒𝑑𝑛𝑒𝑠 𝑠 𝑆𝑒𝑣𝑒𝑟𝑒 + 𝛽 6 ∗ 𝑇𝑖𝑟𝑒𝑑𝑛𝑒𝑠 𝑠 𝑀𝑜𝑑𝑒𝑟𝑎𝑡𝑒 + 𝛽 7 ∗ 𝑆𝑘𝑖 𝑛 _𝑖𝑡𝑐 ℎ𝑖𝑛 𝑔 𝑆𝑒𝑣𝑒𝑟𝑒 + 𝛽 8 ∗ 𝑆𝑘𝑖𝑛 _𝑖𝑡𝑐 ℎ𝑖𝑛 𝑔 𝑀𝑜𝑑𝑒𝑟𝑎𝑡𝑒 + 𝛽 9 𝐻𝑎𝑖𝑟𝐿𝑜𝑠𝑠 + 𝜖 𝑛𝑖 𝑃 𝑛𝑖 = ∫ 𝑒𝑥𝑝 (𝑉 𝑖 ) ∑𝑒𝑥𝑝 (𝑉 𝑗 ) 𝑗 𝑓 (𝛽 )𝑑𝛽 Where 41 𝑈 𝑛𝑖 is the utility when choosing treatment alternative i for individual n 𝑉 𝑛𝑖 is the observed utility when choosing treatment alternative i for individual n β1- β9 are the coefficients that represent the relative importance 𝜖 𝑛𝑖 is the unobserved utility or error term, assumed to be iid extreme type I distributed The effects of PFS (Progression-free survival) and cost (out-of-pocket cost) were treated as continuous variables, whereas symptom severity and adverse effect severity were regarded as dummy variables, with “no symptom” or “no adverse effect” as a reference level. The mixed logit model was implemented using Stata, Version 14.0 (StataCorp. 2015. Stata Statistical Software: Release 14. College Station, TX: StataCorp LP.). The sign of a coefficient indicates whether this attribute has a positive or negative effect on perceived utility. While the absolute values of β1- β9 have no direct interpretation, they present the relative importance of each effect. The trade-offs that the respondents were willing to make between effects were estimated by the marginal substitution rate between two effects (the ratio of two coefficients). Most importantly, the value of coefficient β2 (out-of-pocket cost) can be used to calculate the willingness to pay (WTP) for other effects. We set β2 to be fixed in our mixed logit model, while keeping other variables as random parameters. For example, the WTP to avoid severe symptom was estimated as – (β3/ β2), which means the respondents are willing to pay the amount of – (β3/ β2) per month to avoid experiencing the severe symptom. The confidence intervals were estimated using a Stata module with the delta method. [25] In addition, according to the sample size requirements proposed by de Bekker-Grob et al[26], we calculated the minimum sample size for our study in R software. The result indicated that a sample size of 71 was sufficient. 42 Results Respondents Respondents had a mean age of 66.5 years (SD 11.4, range 32-89). Among respondents, nearly half (47.9%) were male, the most common race (52.1%) was White/Caucasian, and the second most common race (33.8%) was Asian/Pacific Islander. An annual household income level of “$50,000-$74,999” was the most frequently reported (Table 3.1). Table 3.2 Demographic characteristics of respondents, n=71. N=71 Percentages (%) Age (mean, max, min, sd) 66.5 89 32 11.4 Gender Male 34 47.9% Female 37 52.1% Race Asian / Pacific Islander 24 33.8% Black or African American 4 5.6% Hispanic 6 8.5% White / Caucasian 37 52.1% Household annual income level $0-$24,999 12 17.4% $25,000-$49,999 6 8.7% $50,000-$74,999 17 24.6% $75,000-$99,999 9 13.0% $100,000-$124,999 8 11.6% $125,000-$149,999 6 8.7% $150,000-$174,999 1 1.5% 43 $175,000-$199,999 5 7.3% $200,000 and up 5 7.3% SD: standard deviation Two participants refused to indicate their income levels The majority of respondents completed the survey in a range of 8-16 minutes (Table 3.2), which indicated that respondents did spend time on considering trade-offs, rather than making hasty decisions. All respondents passed the internal validity test (chose the dominant alternative), which indicated that respondents had a good comprehension of what they were being asked. Table 3.3 Survey Completion details The time used for the survey N=71 Percentages (%) Mean (minutes) 11 <8min 10 14.08% 8-16 min 53 74.65% >16 min 8 11.27% Able to understand the task Select the dominant alternative 71 100% DCE results The estimated coefficients and associated P-values from the mixed logit model are shown in Table 3.3. All coefficients except for “moderate tiredness” were significant (P<0.05) and all signs of these coefficients were consistent with expectation: The positive sign of the coefficient for “progression-free survival time" indicated that respondents preferred treatment with longer time of disease control, when controlling everything else equal. The negative sign of the 44 coefficient “Monthly out-of-pocket cost” showed that respondents preferred a less expensive treatment. The negative signs for the quality-of-life-related coefficients (symptoms and treatment-related adverse effects) indicated that respondents’ treatment utility would decrease when experiencing these worsened health conditions. Table 3.4 Respondents' preferences for NSCLC treatment Effects Coefficient estimate (95% CI) P-value Progression-free survival time (Months) 0.55 (0.41, 0.68) 0.000 Monthly out-of-pocket cost ($1,000) -0.17 (-0.20, -0.13) 0.000 Severe symptoms -0.34 (-0.41, -0.27) 0.000 Moderate symptoms -0.18 (-0.23, -0.13) 0.000 Adverse effect: severe tiredness -0.17 (-0.22, -0.13) 0.000 Adverse effect: moderate tiredness -0.03 (-0.07, 0.00) 0.074 Adverse effect: severe skin itching -0.35 (-0.42, -0.27) 0.000 Adverse effect: moderate skin itching -0.08 (-0.12, -0.04) 0.000 Adverse effect: hair loss -0.10 (-0.16, -0.05) 0.000 45 The magnitude of these coefficients showed the relative importance among treatment attributes on treatment preferences. The only positive coefficient “Progression-free Survival time” implied a 0.55 increase in utility with a one-month increase in progression-free survival time. On the other hand, the magnitude of all of the negative coefficients is smaller than 0.55, which indicated that a month increase in progression-free survival time could compensate for the disutility of a $1,000 increase in out-of-pocket cost or experiencing these worsened health conditions. Undergoing severe disease symptoms and severe skin itching were regarded as the most influential effects on patients’ treatment preferences, and these two effects had similar influence: -0.34 (95% CI: -0.41, -0.27) vs -0.35(95% CI: -0.42, -0.27). The coefficient of moderate tiredness was not significant, indicated that this effect was not significantly related to patients’ treatment preferences. However, the coefficient for hair loss was significant (-0.10, P<0.01), which means that patients certainly cared about the changes in their appearance in their treatment decision-making. However, the impact of hair loss was less than one-fifth of the satisfaction of improvement in progression-free survival. We noted that women cared slightly more about hair loss than men when making treatment decisions: -0.11 (95% CI: -0.20, -0.02) vs -0.09 (95% CI: -0.16, -0.02). Willingness to pay We estimated willingness-to-pay (WTP) by calculating the marginal rate of substitution between attributes in monetary units. The calculated WTP and associated 95% confidence intervals are shown in Table 3.4. We only estimated WTP for those effects that were significant (excluded “Adverse effect: moderate tiredness”). Compared with coefficients of effects, the WTP estimates are more easily interpretable for the relative importance that patients place on 46 each attribute. Given that, “progression-free survival time” had the most dominant effect on patients’ decision making, our results indicated that the WTP was the highest for having a month increase in progression-free survival time: $ 3,290 (95% CI: $2,550 to $4,030) out-of-pocket cost. Patients were willing to pay more than $2,000 per month to avoid severe symptoms or severe skin itching. Patients were willing to pay around $1,000 a month to avoid the moderate level of symptoms or severe level of tiredness. Compared with the WTP for a one-month increase in progression-free survival time, patients’ WTP dropped to 20%, about $600 to avoid moderate skin itching or hair loss. 47 Table 3.4 Willingness to pay for treatment characteristics 48 We predicted the amount of money that patients would be willing to pay for osimertinib after erlotinib and gefitinib lost their patents. Based on results from the phase III clinical trial, FLAURA,[8] we made the following assumptions. (1) Patients who take osimertinib will have 18.9 months PFS, and patients who take erlotinib/gefitinib will have 10.2 months PFS. (2) After the patents of erlotinib /gefitinib expire, the out-of-pocket cost of erlotinib /gefitinib will become $100 per month. (3) Patients who receive osimertinib have a 57% chance of experiencing moderate skin itching, while this chance for patients who receive erlotinib/gefitinib is 71%. With the help of our estimated coefficients, for osimertinib, we predicted that patients would be willing to pay up to $1,350 out-of-pocket payment per month for osimertinib: WTP for osimertinib = (10.2∗0.55−0.1∗10.2∗0.17−10.2∗0.71∗0.18)−(18.9∗0.55−18.9∗0.57∗0.18) −0.17∗18.9 Discussion We developed a discrete choice experiment to convey useful preference information on how NSCLC patients valued trade-offs in their treatment decision making. Consistent with results from previous studies in Germany and the United Kingdom [11, 12], our study showed that the improvements in progression-free survival also had the most substantial effect on treatment decisions among patients. Moreover, our study also indicated that patients considered severe disease symptoms and severe skin itching as the most intolerable health conditions. The increase of utility due to a one-month increase in “progression-free survival time” was 1.6 times larger than the decrease in utility due to severe symptoms or severe skin itching, 3.2 times larger 49 than the utility decrease due to severe tiredness, and 5.5 times larger than the utility decrease due to hair loss. Furthermore, in contrast to previous DCE studies, our study is the first to include “hair loss” into NSCLC treatment attributes. Our results showed that " hair loss" played a significant role in patients’ decision-making. Moreover, our study illustrated that, in ranking treatment- related adverse effects, patients typically cared more about hair loss, in contrast with their indifference to moderate levels of tiredness. Our study evaluated patients' WTP for improvements in various aspects of NSCLC treatment. Since progression-free survival was the key attribute, patients were willing to pay an average $3,290 out-of-pocket cost for a one-month increase in progression-free survival (95% CI: $2,050, $4,030). Our study also predicted patients’ WTP for osimertinib after erlotinib and gefitinib lose their patent protection. Incorporating clinical trial information with estimated coefficients in our model, we expected patients would be willing to pay up to $1,350 out-of- pocket payment per month for osimertinib after erlotinib and gefitinib lose patents. Our study used the same interviewer across all interviews, which helped minimize potential differences in the interview manner. Further, all patients choose the dominant alternative when it was presented, showing internal response validity. However, our study had some limitations. First, due to power considerations, we only included three treatment-related adverse effects in our research. Therefore, we could not assess the implications of diarrhea, nausea, low appetite, or other adverse effects. We developed our attributes through qualitative analysis. However, this procedure still could not guarantee that the attributes in our study represented all of the relevant factors related to patients’ preferences for 50 NSCLC treatment. Second, we only collected responses from patients who visited their oncologist for advice, thus we failed to capture the preferences for the patients who had already chosen not to treat their NSCLC. Third, limited by our sample size, we employed the main effects model for our data analysis. The interactions among attributes and levels remained unidentified. Future studies can investigate how patients’ demographics or clinical characteristics impacted their treatment decision-making. Understanding patients’ preferences for NSCLC treatment is vital to improving their quality of life. Patient-centered care tied to their actual preferences will satisfy their needs and protect their dignity when challenged by such a life-threatening disease. In conclusion, our study showed that all included treatment attributes proved to be crucial for patients' treatment decision making. The improvements in survival outcomes, disease symptoms control, treatment-related adverse effects, and out-of-pocket cost were crucial determinants of their treatment preferences. 51 References 1. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA: a cancer journal for clinicians 2019;69(1):7-34. 2. Detterbeck FC, Boffa DJ, Tanoue LT. The new lung cancer staging system. Chest 2009;136(1):260-271. 3. Rosell R, Carcereny E, Gervais R, et al. Erlotinib versus standard chemotherapy as first-line treatment for European patients with advanced EGFR mutation-positive non-small-cell lung cancer (EURTAC): a multicentre, open-label, randomised phase 3 trial. The lancet oncology 2012;13(3):239-246. 4. Mok TS, Wu Y-L, Thongprasert S, et al. Gefitinib or carboplatin–paclitaxel in pulmonary adenocarcinoma. New England Journal of Medicine 2009;361(10):947-957. 5. Yang JC-H, Wu Y-L, Schuler M, et al. Afatinib versus cisplatin-based chemotherapy for EGFR mutation-positive lung adenocarcinoma (LUX-Lung 3 and LUX-Lung 6): analysis of overall survival data from two randomised, phase 3 trials. The lancet oncology 2015;16(2):141-151. 6. Wu Y-L, Cheng Y, Zhou X, et al. Dacomitinib versus gefitinib as first-line treatment for patients with EGFR-mutation-positive non-small-cell lung cancer (ARCHER 1050): a randomised, open-label, phase 3 trial. The Lancet Oncology 2017;18(11):1454-1466. 7. Mok TS, Wu Y-L, Ahn M-J, et al. Osimertinib or platinum–pemetrexed in EGFR T790M–positive lung cancer. New England Journal of Medicine 2017;376(7):629-640. 8. Soria J-C, Ohe Y, Vansteenkiste J, et al. Osimertinib in untreated EGFR-mutated advanced non– small-cell lung cancer. New England journal of medicine 2018;378(2):113-125. 9. Aguiar PN, Haaland B, Park W, et al. Cost-effectiveness of Osimertinib in the first-line treatment of patients with EGFR-Mutated advanced non–Small cell lung Cancer. JAMA oncology 2018;4(8):1080- 1084. 10. Accessdata.fda.gov. (2019). Orange Book: Approved Drug Products with Therapeutic Equivalence Evaluations . [online] Available at: https://www.accessdata.fda.gov/scripts/cder/ob/index.cfm [Accessed 13 Oct. 2019]. 11. Bridges JFP, Mohamed AF, Finnern HW, et al. Patients’ preferences for treatment outcomes for advanced non-small cell lung cancer: a conjoint analysis. Lung Cancer 2012;77(1):224-231. 12. Mühlbacher AC, Bethge S. Patients’ preferences: a discrete-choice experiment for treatment of non-small-cell lung cancer. The European Journal of Health Economics 2015;16(6):657-670. 13. Clark MD, Determann D, Petrou S, et al. Discrete choice experiments in health economics: a review of the literature. Pharmacoeconomics 2014;32(9):883-902. 14. Lancaster KJ. A new approach to consumer theory. Journal of political economy 1966;74(2):132- 157. 15. McFadden D. Conditional logit analysis of qualitative choice behavior. 1973. 16. Lancsar E, Louviere J. Conducting discrete choice experiments to inform healthcare decision making. Pharmacoeconomics 2008;26(8):661-677. 17. Johnson FR, Lancsar E, Marshall D, et al. Constructing experimental designs for discrete-choice experiments: report of the ISPOR conjoint analysis experimental design good research practices task force. Value in Health 2013;16(1):3-13. 18. de Bekker‐Grob EW, Ryan M, Gerard K. Discrete choice experiments in health economics: a review of the literature. Health economics 2012;21(2):145-172. 19. Coast J, Al‐Janabi H, Sutton EJ, et al. Using qualitative methods for attribute development for discrete choice experiments: issues and recommendations. Health economics 2012;21(6):730-741. 20. Mangham LJ, Hanson K, McPake B. How to do (or not to do)… Designing a discrete choice experiment for application in a low-income country. Health policy and planning 2009;24(2):151-158. 52 21. US Department of Health and Human Services. "Common terminology criteria for adverse events (CTCAE) version 4.0." National Institutes of Health NCI, no. 03 (2009). In. 22. Louviere JJ, Hensher DA, Swait JD. Stated choice methods: analysis and applications: Cambridge university press; 2000. 23. Kuhfeld WF. Marketing Research Methods in SAS: Citeseer; 2003. 24. Train KE. Discrete choice methods with simulation: Cambridge university press; 2009. 25. Hole A. WTP: Stata module to estimate confidence intervals for willingness to pay measures. In; 2007. 26. de Bekker-Grob EW, Donkers B, Jonker MF, et al. Sample size requirements for discrete-choice experiments in healthcare: a practical guide. The Patient-Patient-Centered Outcomes Research 2015;8(5):373-384. 53 Chapter 4: Testing methodological advances in willingness to pay (WTP) estimation using a discrete choice experiment: an advanced lung cancer treatment case study Abstract Background In discrete choice experiments (DCE), the willingness to pay (WTP) estimation based on the conventional random utility maximization (RUM) is not context-dependent. One alternative behavior approach called random regret minimization (RRM) can generate choice-specific WTP estimates. Recently, a generalized RRM model (µRRM) has been proposed. Methods We designed a DCE study that aimed at assessing preferences for advanced lung cancer treatment. Responses were collected from the Amazon Mechanical Turk platform. Using decision data, we first estimated the attribute-specific scale factor (µ) for survival probability and cost in the pairwise attribute-level regret function. Then, informed by the estimated scale factors, we developed a hybrid model. We estimated the WTP for a 10% improvement in progression-free survival probability in varying scenarios. Lastly, we empirically compared the WTP estimates from the RUM and the hybrid models. Results The estimated scale factors for survival probability and cost were estimated to be smaller than one and larger than five, respectively. This suggested that, in our case, decision-makers were more likely to process the “cost” attribute using utility-maximization decision rule and manage the “survival probability” attribute using regret-minimization decision rule. We then developed a hybrid model in which only “survival probability” was analyzed assuming an underlying regret minimization process. The hybrid model significantly improved model fit compared to RRM. Derived WTP from the hybrid model was sensitive to changes in the composition of the choice set. Conclusion Our empirical analysis confirmed that the WTP derived from RUM was constant across different scenarios and the WTP estimates from the hybrid model were context-dependent. Our results suggest that regret-based models have the potential to offer new insights to explain context-specific choice behavior. 54 Introduction The landscape of treatment options for patients with cancer has evolved tremendously over the past years. There has been an increase in the number of new drugs approved by the US Food and Drug Administration (FDA) for life-threatening cancer treatment[1]. However, these new cancer therapies are priced extremely high[2]. Consequently, the cost burden of cancer treatment is snowballing. Furthermore, there are some controversies around the marginal improvements in survival for these newly approved therapies. Clinical trials usually use median and mean progression-free or overall survival time to represent the effectiveness of treatment. However, even though these average survival gains were statistically significant, the magnitude of these improvements was not large. Fojo et al. (2014) summarized the therapies approved for solid tumors between 2002 and 2014, and they reported that “the median gains in progression- free and overall survival (OS) were a very modest 2.5 and 2.1 months, respectively”.[3] Take lung cancer treatment as one example, the checkpoint immunotherapies have gained popularity in recent years. Patients and oncologists are excited about this treatment option, not only because it had improved average survival gains. From the results of a Phase 3 clinical trial (Keynote 189), in which pembrolizumab plus chemotherapy was compared with placebo plus chemotherapy for advanced non-small-cell lung cancer [4], the survival improvement in terms of median progression-free survival was not very exciting: 8.8 months in the pembrolizumab-combination group versus 4.9 months in the placebo-combination group. Nevertheless, even though checkpoint immunotherapies do not help as many people, a portion of patients lived longer. More importantly, research had shown that, in the settings of treatment for severe disease, people might care about more than just the average or median survival gains associated with a treatment, which means they may value the “hope” of a survival gain, independent of a therapy’s 55 average gain. [5] [6] One contingent valuation study had indicated that the framing of survival gains as an increase in the probability of survival rather than an increase in median survival time increased the perceived value of cancer care.[7] Therefore, in addition to average survival gains, assessments of the economic value of end-of-life cancer therapies from decision-makers’ perspectives should incorporate preferences on survival probability. In this study, we want to estimate the perceived willingness to pay (WTP) for improvements in survival probability in the setting of immunotherapies for the treatment of advanced lung cancer. Discrete choice experiment (DCE) is one widely employed tool in understanding preferences. The advantage of DCE is that it can quantify the relative importance of crucial factors in decision making. Unlike the contingent valuation approach that elicits WTP for specific goods, DCE estimates WTP based on the marginal rates of substitution between generic attributes. Broadly speaking, in a DCE, there are several treatment scenarios. The treatment is represented by a limited number of attributes. Moreover, these attributes take a few values (called levels). The combination of attributes and associated levels determines the profile of each hypothetical treatment (alternative). Participants are asked to state their prefered alternative in each scenario. Researchers can then uncover the relative importance of attributes and elicit the preferences by analyzing these responses with choice models.[8, 9] The majority of discrete choice experiment studies explain choice behavior with the classical utility maximization decision process. The process is based on the principle that, while choosing alternatives, decision-makers tend to maximize their utility.[10, 11] McFadden introduced the random utility maximization models, which assumes that in addition to the systematic part, there is a random part in the utility.[12] The random part is assumed to be independently and identically distributed (i.i.d) with the extreme value type one distribution, and 56 then the predicted probability of selecting one alternative can be represented by the logit format. Researchers have developed well-organized guidelines for applying the RUM models.[8, 9] Recently, there is a growing interest in implementing an alternative behavioral framework for understanding choice behavior, the random regret minimization (RRM).[13, 14] Before the introduction of the RRM, the literature showed that the anticipated regret had influenced choice decision-making[15, 16]. In contrast with RUM’s maximizing expected utility, the RRM supposes that decision-makers want to minimize their anticipated regret.. Specifically, the anticipated regret is the feeling when comparing the chosen alternative against non-chosen alternatives, and the non-chosen alternative turns out to be more attractive than the chosen one. While similar to RUM, RRM also assumes that anticipated regret can be composed of two parts, the systematic part, and the random part. Then the random part is also assumed to be extreme value type one distributed. Therefore, selection probabilities can also be described with the multinomial logit model. The RRM, which was initially applied in the domain of transportation choices[17]. Researchers then extended its application in explaining and predicting choices beyond transportation, such as to choice contexts in environmental policy[18], automobile fuel[19], and healthcare[20]. The comparisons of the empirical performance between RRM and RUM have been widely discussed. Chorus et al. summarized 43 empirical comparisons between RRM and the traditional RUM, and they reported that "RRM and hybrid RRM–RUM models outperform RUM counterparts in a majority of cases, in terms of model fit and predictive ability"[21]. Consequently, RRM has become a useful tool in understanding decision choice behavior. More recently, a variant of the RRM has been proposed: µRRM[22]. The µRRM generalizes the RRM by introducing a scale factor in the attribute-level regret function, which further increases its 57 flexibility. More importantly, the scale factor determines the shape of the attribute-level regret function: a small µ indicates that the attribute is more likely being processed using a regret minimization rule, a large µ suggests that the attribute is more likely being processed using a utility maximization rule. Classical RRM is a special case of µRRM when scale factors for all attributes take the value of one. To our knowledge, although several studies have employed the µRRM in the field of transportation[23] [24], and energy programs[25], no published study had applied the µRRM in understanding preferences in health care decisions, especially for end-of-life treatment. Furthermore, the fundamental difference between RUM and RRM is the “compensatory” assumption in the RUM and the “semi-compensatory” assumption in the regret minimization decision role[26, 27]. To be specific, in the RUM, the loss in utility because of poor performance on one attribute for one alternative can be fully compensated by better performance on another attribute. However, the RRM-related models postulate that “improve an alternative in terms of an attribute on which it already performs well relative to other alternatives generates only small decreases in regret, whereas deteriorating to a similar extent the performance on another alternative may generate substantial increases in regret.”[21] This “semi-compensatory” behavior in RRM-related models result in choice-set-specific WTP estimation. This indicates that the amount of WTP for improvements in survival probability for a specific treatment depends on the relative performance among treatment alternatives. For example, if a treatment is already with relatively high survival probability when compared with other treatment alternatives, then the people may not be willing to pay much more for further improvement in survival probability of this treatment. 58 In this study, first, we want to understand and quantify the treatment preferences on the increase in the probability of survival for advanced lung cancer using one discrete choice experiment. Second, we want to explore the role of µRRM modeling in distinguishing which attribute is processed using utility-maximization decision rule and which attribute is processed using regret-minimization decision rule. Third, we want to empirically investigate the context- specific WTP estimates obtained from RRM-related models. Method Study participants We collected responses from workers on Amazon Mechanical Turk (MTurk), an online crowdsourcing platform run by Amazon.com. This platform provides fast, convenient, and inexpensive online access to researchers.[28] On this platform, requesters can create and publish tasks (known as “Human Intelligence Tasks” – HITs), and workers are paid upon successful completion of each HIT. Several validation studies have shown the usefulness and reliability of MTurk[29, 30]. We also conducted our own pilot testing on MTurk with the same survey that we used to interviewe patients (survey in Chapter 3). The results indicated that the estimated WTPs from the MTurk workers were consistent with the ones from the patients (Table 4.1). Collectively, the evidence suggested that MTurk is a reliable sampling frame for our study. Table 4.1 WTP estimation obtained from patients and Amazon MTurk workers Willingness to Pay (95% CI) Attribute Patients (n=71) Amazon mTurk (n=20) 59 Progression-free survival time (Months) $3,300 ($2,500, $4,000) $3,200 ($2,500, $3,900) Severe symptoms $2,000 ($2,400, $1,700) $2,300 ($3,000, $1,600) Moderate symptoms $1,100 ($1,300, $800) $900 ($1,300, $600) Adverse effect: severe tiredness $1,000 ($1,300, $700) $1,000 ($1,400, $500) Adverse effect: moderate tiredness $200 ($400, $0) $400 ($700, $100) Adverse effect: severe skin itching $2,100 ($2,600, $1,600) $1,900 ($2,600, $1,300) Adverse effect: moderate skin itching $500 ($800, $200) $300 ($600, $0) Adverse effect: hair loss $600 ($900, $300) $400 ($700, $200) We used additional restrictions on the MTurk participants to guarantee quality of the responses: (1) they must have completed over 100 HITs; (2) they must have higher than an 80% HIT approval rating for their previous HITs. We only recruited US residents into our survey population, since this study investiages WTP in the US. Participants who completed the survey received $1.00 compensation. DCE design We developed the most relevant treatment attributes based on a review of the literature and discussion with experts. We selected the “one-year progression-free survival probability” as the attribute for survival outcome, and we described it as “chance of being alive with cancer in remission at one year” since in the real world one treatment would be discontinued if the disease progressed. We included the “total out-of-pocket cost” to represent the cost burden on decision- makers. The trade-offs between this attribute and the survival outcome enabled us to estimate the 60 amount of willingness to pay for incremental improvement on survival probability. For the level development, we conducted one retrospective data analysis using the Optum database (2016 quarter 1 to 2018 quarter 2), and we estimated the total out-of-pocket cost on utilization of three immune checkpoint inhibitors (nivolumab, pembrolizumab, and atezolizumab) for NSCLC treatment (Table 4.2). We initially set the levels of total out-of-pocket cost as “$0, $3,000 and $6,000”. In one pilot test (n=20), we detected that participants always selected the treatment alternative with the higher survival probability, regardless of the cost level. Therefore, we modified the cost levels as “$0, $4,000 and $8,000”. We included three common adverse effects of immune checkpoint inhibitors into the design: diarrhea, skin rash, and tiredness.[31] We described these adverse effects according to the Common Terminology Criteria for Adverse Events (CTCAE) as guidance[32]. The overview of our attributes and levels is presented in Table 4.3. Table 4.2 level development for total out-of-pocket cost from a retrospective data analysis on Optum Nivolumab Pembrolizumab Atezolizumab Sample size 4,477 2,693 196 Number of patients who had “None-zero” out-of-pocket cost 2,747 1,686 126 Among patients who had “None-zero” out-of-pocket cost Mean $4,797 $3,865 $3,152 61 Median $2,713 $2,280 $2,929 Standard Deviation $7,480 $5,110 $2,768 Our DCE contained five attributes: two three-level attributes and three two-level attributes. The were 72 possible combinations of all these attributes and levels for one treatment alternative (3 * 3 * 2 * 2 * 2). If we were to represent all the possible combinations of these treatment alternatives, we would have to present 72*71/2= 2556 different scenarios. Given that it Table 4.3 Overview of treatment attributes and levels 62 is not possible to provide a single participant with all these scenarios, we used a fractional factorial design to generate an orthogonal main-effects design. This design was constructed based on a design that maximizes D-efficiency using SAS macros. The DCE was presented to the participants via a survey administered via Qualtrics (www.qualtrics.com). The questionnaire consisted of two parts: demographic information and 18 treatment scenarios. For each treatment scenario, participants were provided with two hypothetical treatment options and an “opt-out” alternative (i.e. “No treatment”) to mimic real-life situations. One example of the scenarios is presented in Figure 4.1. 63 Figure 4.3 An example of a choice set Evolution of RRM-related models The first RRM[17] (2008) was developed on the notion of minimizing the maximum loss from the principle proposed by Savage[33]. In key component of this model is the attribute-level regret function. The attribute-level regret of comparing the alternative i with another alternative j on attribute m (𝑅 𝑖 ↔𝑗 𝑚 ) was represented by: max {0, [𝛽 𝑚 (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )]}. The 𝛽 𝑚 represents the weights (taste) of attribute m. In addition, the minimax regret behavior is described by:min 𝑖 〈 max 𝑖 ≠𝑗 {∑ max [0, 𝛽 𝑚 (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )] 𝑚 } + 𝜀 𝑖 〉. Then, the RRM (2010)[13] was revised on two dimensions: 1) it developed a smooth attribute-regret function instead of the non-smooth max operator, which provided the solution for difficulties in the derivation of marginal effects and the model estimation through maximum likelihood. The 𝑅 𝑖 ↔𝑗 𝑚 took the form 64 of ln{1 + exp [𝛽 𝑚 ∗ (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )] }. 2) It assumed that regret is “potentially experienced with respect to each foregone alternative that performs well,” rather than with respect to only the best of foregone alternative. Therefore, this model assumed the decision-maker would minimize the sum of regret: ∑ ∑ ln{1 + exp [β m ∗ (x jm − x im )] } m j≠i . This function represented the sum of all pairwise regret between the chosen alternative and all non-chosen alternatives across all attributes. It should be noted that the pairwise regret function was not chosen arbitrarily, it was developed by adding two i.i.d (extreme value type I-distributed) errors 𝜐 𝑚 1 and 𝜐 𝑚 2 in the non- smooth function, where 𝑅 𝑖 ↔𝑗 𝑚 = 𝐸 〈max {0 + 𝜐 𝑚 1 , [𝛽 𝑚 (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 ) + 𝜐 𝑚 2 ]}〉. Then, the integration of 𝑅 𝑖 ↔𝑗 𝑚 over f (𝜐 𝑚 1 ,𝜐 𝑚 2 ) resulted in this function: ln{1 + exp [𝛽 𝑚 ∗ (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )] }. Moreover, in the process of integration, the RRM (2010) assumed that the scale factor (µ 𝑚 ) for the error terms 𝜐 𝑚 1 and 𝜐 𝑚 2 was normalized to one (the variance of 𝜐 𝑚 1 , 𝜐 𝑚 2 is 𝜋 2 6 ). In 2015, Cranenburgh[22] relaxed this assumption and allowed these scale factors to be estimated (the variance of 𝜐 𝑚 1 , 𝜐 𝑚 2 is 𝜋 2 6 ∗ µ 𝑚 2 ). The 𝛽 𝑚 and µ 𝑚 jointly measured the importance of attribute (m): 𝑅 𝑖 ↔𝑗 𝑚 = µ 𝑚 × ln {1 + exp [ 𝛽 𝑚 µ 𝑚 (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )] }. Therefore, this model is more flexible, and it enables researchers to explore the attribute-level decision rule. The estimated attribute-specific scale factor provides deeper insights into the underlying decision process (utility-maximization or regret-minimization) for an attribute. This information contributes to the development of a hybrid model in which some attributes are processed by the decision-makers in a utility- maximization way, while others are processed in a regret-minimization way. Model framework 65 Suppose a decision-maker faces J (j=1, …, J) alternatives in a choice set. Each alternative is described by M (m=1, …, M) attributes. RUM approach[12] The RUM assumes that the utility of choosing one alternative (i), 𝑈 𝑖 , is composed of two parts: the systematic and the random. The systematic part (𝑉 𝑖 ) reflects the representative tastes, which is the linear-additive function of attributes. Specifically, is ∑ 𝛽 𝑚 × 𝑥 𝑚 𝑚 . The random part (𝜀 𝑖 ) is unknown to researchers, which is the heterogeneous tastes or measurement error. 𝑈 𝑖 = 𝑉 𝑖 + 𝜀 𝑖 = ∑ 𝛽 𝑚 × 𝑥 𝑚 𝑚 + 𝜀 𝑖 The decision-maker is assumed to choose the alternative that maximizes utility. Namely, the probability that the decision-maker chooses alternative i, rather than any other alternative, is the probability that the utility of choosing i is the highest among the choice set: 𝑃 (𝑖 ) = 𝑃 (𝑈 𝑖 > 𝑈 𝑗 , ∀𝑗 ≠ 𝑖 ) = 𝑃 (𝑉 𝑖 + 𝜀 𝑖 > 𝑉 𝑗 + 𝜀 𝑗 ) = 𝑃 (𝜀 𝑗 − 𝜀 𝑖 < 𝑉 𝑖 − 𝑉 𝑗 ) Given 𝜀 𝑖 and 𝜀 𝑗 are assumed to be i.i.d. extreme value type I distributed, 𝑃 (𝑖 ) is represented by a logit-type formula: 𝑃 (𝑖 ) = 𝑒𝑥𝑝 (𝑉 𝑖 ) ∑ 𝑒𝑥𝑝 (𝑉 𝑗 ) 𝑗 =𝐽 𝑗 =1 µRRM approach[22] As presented above, the µRRM presumes that when facing alternatives, the decision- maker aims to minimize the anticipated regret rather than maximize the utility. The anticipated regret is the sum of all attribute-level regrets between the chosen alternative and all the non- 66 chosen alternatives. The attribute-level regret is R i↔j m = µ m × ln {1 + exp [ β m µ m (x jm − x im )] }. Similarly, the µRRM assumes that the regret can be partitioned into a systematic component (𝑅 𝑅 𝑖 ) and a random component (𝜀 𝑖 ). Mathematically, the regret of choosing alternative i is: 𝑅 𝑅 𝑖 = 𝑅 𝑖 + 𝜀 𝑖 = ∑ ∑ µ 𝑚 ∗ 𝑙𝑛 {1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 (𝑥 𝑗𝑚 − 𝑥 𝑖𝑚 )] } + 𝜀 𝑖 𝑚 𝑗 ≠𝑖 The decision-maker now is assumed to choose the alternative that minimizes the regret. That is to say, the probability that the decision-maker chooses alternative i, nor the other alternatives, is the probability that the regret of choosing i is the lowest among the choice set: 𝑃 (𝑖 ) = 𝑃 (𝑅 𝑅 𝑖 < 𝑅 𝑅 𝑗 , ∀𝑗 ≠ 𝑖 ) = 𝑃 (−(𝑅 𝑅 𝑖 ) > −(𝑅 𝑅 𝑗 ), ∀𝑗 ≠ 𝑖 ) = 𝑃 (−(𝑅 𝑖 + 𝜀 𝑖 ) > −(𝑅 𝑗 + 𝜀 𝑗 ), ∀𝑗 ≠ 𝑖 ) = 𝑃 [(−𝜀 𝑗 ) − (−𝜀 𝑖 ) < (−𝑅 𝑖 ) − (−𝑅 𝑖 ))] In the same way, µRUM assumed −𝜀 𝑖 and −𝜀 𝑗 are i.i.d. extreme value type I distributed. Given minimizing anticipated regret is identifical to maximizing the negative of the anticipated regret, the probability of choosing alternative i is given by a similar logit-type formula: 𝑃 (𝑖 ) = 𝑒𝑥𝑝 (−𝑅 𝑖 ) ∑ 𝑒𝑥𝑝 (−𝑅 𝑗 ) 𝑗 =𝐽 𝑗 =1 Hybrid approach As described above, for one attribute (m), if the estimated scale factor µ m is large ( >5), then we would assume this attribute is treated in a utility-maximization manner. If the estimated scale factor µm is small, then we assume this attribute is treated in a regret-minimization manner. 67 In a hybrid model, suppose there are X attributes being managed through the RUM function, Y attributes being managed through the RRM function, and the random component is assumed to be extreme value type I distributed, then the utility of this hybrid model is: 𝑈 i = V i + ε i = ∑ β m × x m m=1..X − ∑ ∑ µ m ∗ ln {1 + exp [ β m µ m (x jm − x im )] } + ε i m=1..Y j≠i Consequently, the probability of choosing alternative i for this hybrid model is represented by a logit-type formula: 𝑃 (𝑖 ) = 𝑒𝑥𝑝 (𝑉 𝑖 ) ∑ 𝑒𝑥𝑝 (𝑉 𝑗 ) 𝑗 =𝐽 𝑗 =1 Model estimation All models mentioned above were estimated using maximum likelihood. This study employed PandasBiogeme to analyze the decision data with each approach. Biogeme is one open-source software and is specifically designed for discrete choice modeling. PandasBiogeme is the third version of Biogeme. This package combines Python and C++ and is distributed on the “Python Package Index” repository. Biogeme’s developer (Michel Bierlaire) provided plentiful resources in programming on the related website.[34] Caspar Chorus and Sander van Cranenburgh also offered valuable tutorials on the estimation of RRM-related models [14, 35] Model fit We used the log-likelihood as the measure of goodness-of-fit to decide which model (RUM, RRM, or Hybrid) fits the data better. We implemented the Ben-Akiva and Swait test[36] to compare which model significantly fits the data better since these models are not nested. This 68 test enables us to obtain an upper bound on the probability that the Akaike Likelihood ratio index will choose the incorrect model when comparing two non-nested specifications of a model. This upper bound can be considered a conservative proxy for the significance of a difference in model fit between two non-nested models. Akaike Likelihood Ratio index for model X: ρ x = 1 − LL(X)−K LL(0) Upper bound probability when model A achieves a lower log-likelihood than non-nested model B: Pr [(𝜌 𝐴 − 𝜌 𝐵 ) ≥ 𝑧 ] ≤ 𝛷 [−(−2 ∗ 𝑧 ∗ 𝐿𝐿 (0) + 𝐾 𝐴 − 𝐾 𝐵 ) 0.5 ] Where K is the number of parameters; LL(0) is the initial log-likelihood; Z is the model fit difference. Parameters and willingness to pay (WTP) The implication of the sign of a coefficient is consistent across models: for one effect, a positive sign indicates the decision-maker prefers higher value on this effect; negative sign suggests the decision-maker prefers lower value on this effect. However, the interpretation of a coefficient's magnitude differs between different decision rules: in the RUM, the magnitude reflects the utility change of a unit increase in an attribute’s value; in the RRM and hybrid model, the extent of change in attribute-level regret that is caused by a unit increase in an attribute, is determined by the joint effect of ß m and µm, and the relative performance between these two alternatives on this attribute (m). Specifically, the parameter ß m captures the slope of the 69 attribute-level regret function, and the parameter µm captures the level of regret aversion for this attribute. The WTPs in all models can be derived based on the marginal substitution rate between the outcome attribute and the cost attribute, parsimoniously. The details of the partial derivative of the µRRM were explained in the appendix. For each WTP estimation, we implemented the non-parametric bootstrap method (draws=100) to obtain the distribution. Results Descriptive statistics From 09/02/2019 to 09/16/2019, 482 Amazon MTurk workers participated in our task. Among these workers, 473 finished the task. Fifteen participants failed to pass the internal validity test. Finally, we collect 458 responses. The demographic information for these participants is presented in Table 4.4. The majority of respondents had an age in the range of 25-34 and 35-44. About 52% of the respondents were male. The most common race was white (77%). Nearly half of the participants reported the annual household income level less than $50,000. Moreover, most of the respondents were employed. Table 4.4 Demographic Characteristics for Amazon mTurk workers participated in the survey Characteristics N=458 Percentage (%) Age 18-24 37 8.1% 25-34 141 30.8% 35-44 159 34.7% 45-54 75 16.4% 70 55-64 39 8.5% 65-74 7 1.5% Gender Male 239 52.3% Female 218 47.7% Race White or Caucasian 352 77.0% Black or African American 39 8.5% Hispanic or Latino 23 5.0% Asian or Asian American 29 6.3% Other 14 3.2% Income $0-$24,999 83 18.1% $25,000-$49,999 142 31.0% $50,000-$74,999 110 24.0% $75,000-$99,999 62 13.5% $100,000-$124,999 29 6.3% $125,000 and up 32 7.0% Employment Working (paid employee) 313 68.3% Working (self-employed) 92 20.1% Not working 53 11.6% Model output We reported the output of five model specifications: RUM, RRM, µRRM_generic, µRRM_specific and hybrid in Table 4.5. In the µRRM_generic model, we assumed that all attributes shared the same scale factor µ0. The estimated µ0 was larger than 5 (upper bound suggested by Cranenburgh[22]), which indicated that overall, the choices in the data were best described by the utility-maximization behavior. In other words, attribute-level regret functions were linear, and the model generated the same choice probabilities as the RUM model. Mathematically, for binary effects, RUM and RRM process these variables equivalent.[26] In the µRRM_specific model, we assumed a generic scale factor µ0 for all binary 71 effects, and we assigned attribute-specific scale factors to continuous effects: cost (µ cost) and survival probability (µ 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 ). As expected, the estimated µ0 was larger than 5. The estimated µ 𝑠𝑢𝑟 𝑣 𝑖𝑣𝑎𝑙 was smaller than 1, and the estimated µ cost was larger than 5. This result indicated the “survival” attribute was processed using the regret minimization rule, and the “cost” attribute was processed using the utility maximization rule. Finally, we developed a hybrid model, which assumed only the “survival” was analyzed using an underlying regret minimization process, while other effects were coded according to a linear-additive utility maximization rule. In this model, the estimated µ 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 was 0.114. We implemented a bootstrapping approach to obtain the asymptotically correct standard errors for coefficients. The standard error for µ 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 was 0.003. This scale parameter implied that the degree of regret aversion for the “survival” attribute was larger than the one imposed by the classical RRM model implicitly, in which µ is assumed to be 1. The signs of the coefficients in all the models were consistent with expectations, which also endorsed the theoretical validity of all different models. All coefficients were significant at the 0.01 significance level. For the model fit, with the largest log-likelihood, the RUM model had a better fit with the data when compared with both the RRM and the hybrid model. Even though these differences in fit were small, the Ben-Akiva and Swait test indicated that the RUM fitted the data significantly better than the RRM and the hybrid model, while the hybrid model fitted the data significantly better than the RRM model (Table 4.6). 72 Table 4.5 Estimation results 73 Table 4.6 Log likelihood test results Log-likelihood Test Ben-Akiva and Swait test Interpretation RUM vs. RRM Z=-12.82 P <0.000 RUM fits data significantly better than RRM RUM vs. Hybrid Z=-12.32 P <0.000 RUM fits data significantly better than hybrid Hybrid vs. RRM Z=-3.52 P <0.000 Hybrid fits data significantly better than RRM Willingness to pay estimation We compared the WTP estimation for a 10% increase in survival probability from two models: the RUM and the hybrid model. The WTP from the RUM model was a constant ratio of two coefficients: (−1) ∗ 𝛽 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 𝛽 𝑐𝑜𝑠𝑡 . On the contrary, the WTP derived from the hybrid model was (−1) ∗ ∑ −𝛽 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 exp[− 𝛽 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 µ 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 ∗(𝑋 𝑗 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 −𝑋 𝑖 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 )]+1 𝑗 ≠𝑖 𝛽 𝑐𝑜𝑠𝑡 . Noticeably, the WTP obtained from the hybrid model was context-dependent. Namely, the WTP was a function of the composition of the choice tasks. The WTP estimates were determined by the differences between levels on the “survival” attribute between alternatives. It’s important to note that, since the “cost” attribute was treated in a utility-maximization manner, the denominator of the WTP formula was not context-dependent, which translated to an identical change in utility that was caused by the marginal change in cost among different situations. 74 We generated two hypothetical scenarios (Scenario A and Scenario B) to better illustrate the WTP’s context-dependent feature in the hybrid model. In Scenario A, the one-year progression-free survival probability for Treatment A was assumed to be 75%. In Scenario B, the one-year progression-free survival probability for Treatment A was assumed to be 90%. In both scenarios, the one-year progression-free survival probability for Treatment B was changed, from 40% to 70%. Then we plotted a graph (Figure 4.2) in which the vertical axis was the WTP for a 10% improvement in one-year progression-free survival probability; the horizontal axis was the one-year progression-free survival probability for Treatment B. Briefly speaking, the graph indicated that: the WTP derived from RUM was constant across different scenarios; the WTP derived from the hybrid model was context-dependent. Specifically, the WTP derived from the hybrid model depended on the relative performance of the considered alternative and the competing alternatives. When the survival probability of the considered alternative (Treatment B) was relatively low, there was a larger WTP for an improvement in the survival probability. As the performance of the considered alternative improved relative to the completing alternative (Treatment A), the WTP for incremental survival probability decreased. Theoretically speaking, when the survival probability of the considered alternative was relatively low, there was considerable pairwise regret associated with the comparison with the competing alternative. An improvement in survival probability resulted in a sizeable reduction in regret. Therefore, when there was a large difference between the considered alternative and the competing alternatives in terms of survival outcomes, decision-makers were willing to pay more for a marginal increase in survival probability. In contrast, when the survival outcome for the considered alternative was close to its competitor, there was little regret associated with the pairwise comparison. At this point, further improvements only led to very 75 small reductions in regret. Hence, when there was a small difference between the considered alternative and the competing alternatives on survival outcome, decision-makers were willing to pay less for an incremental increase. Figure 4.2 Willingness to pay for a 10% increase in one-year progression-free survival probability Sub-group analysis: willingness-to-pay estimation $9,484 $9,258 $9,047 $8,821 $8,549 $8,198 $7,737 $9,153 $8,773 $8,348 $7,836 $7,205 $6,440 $5,556 $9,000 $5,000 $5,500 $6,000 $6,500 $7,000 $7,500 $8,000 $8,500 $9,000 $9,500 $10,000 40% 45% 50% 55% 60% 65% 70% One-year progression-free survival probability for treatment B WTP for a 10% increase in one-year progression-free survival probability Hybrid-Scenario B Hybrid-Scenario A RUM 76 We performed subgroup analysis between two groups: participants with annual household income levels <$50,000 and annual household income levels > $50,000. The result indicated that the predicted WTP increased with respondents’ income, regardless of which model was implemented. Not surprisingly, our results showed that higher-income decision-makers were willing to pay more than lower-income people. The validity of the hybrid model was confirmed. The WTP pattern that we described above was also detected in both groups: the WTP derived from the hybrid model depended on the relative performance of the considered alternative and the competing alternatives. The better the competing alternative, the more the decision-makers were willing to pay for the improvement in the considered alternative. The better the considered alternative, the less the decision-makers were willing to pay for the improvement in this considered alternative. The estimated WTPs from different scenarios are presented in Figure 4.3 77 Figure 4.3 Sub-group analysis 78 Discussion In this study, we assessed the treatment preferences for advanced lung cancer using a discrete choice experiment. We empirically demonstrated the methodological advances in estimating the WTP by implementing an alternative behavior approach. To our knowledge, our study is the first healthcare-related discrete choice experiment study that implemented the µRRM in explaining the decision behavior. We also introduce using the µRRM approach to obtain insights into the development of a hybrid model framework. The estimation of the attribute-specific scale factor allows researchers to distinguish which attributes should be analyzed using the conventional utility maximization rules, and which attributes are better be processed using the regret minimization rules. The attribute-specific scale factor captures the degree of regret aversion to that attribute. In our case, the estimated scale-factor for “survival probability” was significantly smaller than one, which indicated that the regret aversion level for “survival probability” was larger than the level imposed by the classical RRM model. Using hybrid modeling, even though the model fit improved slightly from the RRM, the model fit improvement was significant. Model fit was not the primary model property that we were interested in. We cared more about the implications from the model, the WTP estimation. In this study, we tested methodological advances in obtaining context-dependent WTP estimation. In the hybrid model, the “survival probability” attribute was processed in a regret-based fashion. This resulted in the choice-set specific WTP measures for this attribute. More precisely, the WTP measures depended on the composition of the choice set in terms of relative performance. This approach allows for a richer interpretation of the trade-offs decision-makers make as a function of the composition of the choice set they face. 79 We empirically demonstrated how the WTP derived from the hybrid model for a marginal improvement in survival probability was strongly impacted by the relative performance of the competing alternative (Treatment A) and considered alternative (Treatment B) in Figure 4.2. As the survival outcome of Treatment B increased, the model predicted that the WTP for a 10% increase in one-year progression-free survival probability decreased. When keeping the performance of the considered alternative constant, as the survival outcome of the competing alternative increased (Scenario A to Scenario B), the model showed that the WTP for a 10% increase in one-year progression-free survival probability increased. In contrast, the RUM model-based WTP estimation remained unchanged, irrespective of the setting in different scenarios. This is the evidence of the context-dependent estimation of WTP, which allows researchers and policymakers to predict the WTP based on a particular situation. Our study suggests that we can provide a more precise assessment of the treatment preferences by combining the information of specific treatment options, and the regret-based processing rules. Our study had some limitations. First, we processed the survival probability proportionally. However, study indicated that decision-makers may not use linear probability weighting.[37] Future studies should allow probability weighting in modeling. Second, our study only contained stated-preference data. Chorus [21]showed that in the context of revealed preference data, RRM models also had a comparable model fit performance with the RUM model. It would be interesting to further test the µRRM approach to revealed preference data and then explore the attribute specific scale factor in regret function. Third, since the one-year overall or progression-free survival rate is one commonly reported survival outcome measure for advanced lung cancer, in this study we only investigated the preferences on “one-year 80 progression-free survival probability”. However, it would be worthwhile to assess the stated preferences for treatments with the expanded time horizon. Appendix The partial-derivation of µRRM 𝜕 𝑅 𝑖 µ𝑅𝑅𝑀 𝜕 𝑋 𝑚𝑖 = 𝜕 𝜕 𝑋 𝑖𝑚 µ 𝑚 ∑ 𝑙𝑛 {1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )]} 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝜕 𝜕 𝑋 𝑖𝑚 𝑙𝑛 {1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )]} 𝑗 ≠𝑖 = µ 𝑚 ∑ 1 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝜕 𝜕 𝑋 𝑖𝑚 {1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )]} 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝜕 𝜕 𝑋 𝑖𝑚 (1) + 𝜕 𝜕 𝑋 𝑖𝑚 {𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )]} 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝜕 𝜕 𝑋 𝑖𝑚 {𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )]} 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] ∗ 𝜕 𝜕 𝑋 𝑖𝑚 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] ∗ 𝛽 𝑚 µ 𝑚 ∗ 𝜕 𝜕 𝑋 𝑖𝑚 (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 ) 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝑗 ≠𝑖 = µ 𝑚 ∑ 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] ∗ 𝛽 𝑚 µ 𝑚 ∗ (−1) 1 + 𝑒𝑥𝑝 [ 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] 𝑗 ≠𝑖 = ∑ −𝛽 𝑚 𝑒𝑥𝑝 [− 𝛽 𝑚 µ 𝑚 × (𝑋 𝑗𝑚 − 𝑋 𝑖𝑚 )] + 1 𝑗 ≠𝑖 81 References 1. Johnson JR, Ning Y-M, Farrell A, et al. Accelerated approval of oncology products: the food and drug administration experience. JNCI: Journal of the National Cancer Institute 2011;103(8):636-644. 2. Vivot A, Jacot J, Zeitoun JD, et al. Clinical benefit, price and approval characteristics of FDA- approved new drugs for treating advanced solid cancer, 2000–2015. Annals of Oncology 2017;28(5):1111-1116. 3. Fojo T, Mailankody S, Lo A. Unintended consequences of expensive cancer therapeutics—the pursuit of marginal indications and a me-too mentality that stifles innovation and creativity: the John Conley Lecture. JAMA Otolaryngology–Head & Neck Surgery 2014;140(12):1225-1236. 4. Gandhi L, Rodríguez-Abreu D, Gadgeel S, et al. Pembrolizumab plus chemotherapy in metastatic non–small-cell lung cancer. New England journal of medicine 2018;378(22):2078-2092. 5. Becker G, Murphy K, Philipson T. The value of life near its end and terminal care. In: National Bureau of Economic Research; 2007. 6. Lakdawalla DN, Romley JA, Sanchez Y, et al. How cancer patients value hope and the implications for cost-effectiveness assessments of high-cost cancer therapies. Health Affairs 2012;31(4):676-682. 7. Lin P-J, Concannon TW, Greenberg D, et al. Does framing of cancer survival affect perceived value of care? A willingness-to-pay survey of US residents. Expert review of pharmacoeconomics & outcomes research 2013;13(4):513-522. 8. Louviere JJ, Hensher DA, Swait JD. Stated choice methods: analysis and applications: Cambridge university press; 2000. 9. Hensher DA, Rose JM, Greene WH. Applied choice analysis: a primer: Cambridge University Press; 2005. 10. Manski CF. The structure of random utility models. Theory and decision 1977;8(3):229-254. 11. Marschak J. Binary-choice constraints and random utility indicators (1960). In. Economic Information, Decision, and Prediction: Springer; 1974, 218-239. 12. McFadden D. Conditional logit analysis of qualitative choice behavior. 1973. 13. Chorus CG. A new model of random regret minimization. European Journal of Transport and Infrastructure Research 2010;10(2). 14. Chorus CG. Random Regret-based discrete choice modeling: A tutorial: Springer Science & Business Media; 2012. 15. Zeelenberg M. Anticipated regret, expected feedback and behavioral decision making. Journal of behavioral decision making 1999;12(2):93-106. 16. Tversky A, Kahneman D. Loss aversion in riskless choice: A reference-dependent model. The quarterly journal of economics 1991;106(4):1039-1061. 17. Chorus CG, Arentze TA, Timmermans HJP. A random regret-minimization model of travel choice. Transportation Research Part B: Methodological 2008;42(1):1-18. 18. Thiene M, Boeri M, Chorus CG. Random regret minimization: exploration of a new choice model for environmental and resource economics. Environmental and resource economics 2012;51(3):413- 429. 19. Hensher DA, Greene WH, Chorus CG. Random regret minimization or random utility maximization: an exploratory analysis in the context of automobile fuel choice. Journal of Advanced Transportation 2013;47(7):667-678. 20. de Bekker-Grob EW, Chorus CG. Random regret-based discrete-choice modelling: an application to healthcare. PharmacoEconomics 2013;31(7):623-634. 82 21. Chorus C, van Cranenburgh S, Dekker T. Random regret minimization for consumer choice modeling: Assessment of empirical evidence. Journal of Business Research 2014;67(11):2428-2436. 22. van Cranenburgh S, Guevara CA, Chorus CG. New insights on random regret minimization models. Transportation Research Part A: Policy and Practice 2015;74:91-109. 23. Belgiawan PF, Dubernet I, Schmid B, et al. Context-dependent models (CRRM, MuRRM, PRRM, RAM) versus a context-free model (MNL) in transportation studies: a comprehensive comparisons for Swiss and German SP and RP data sets. Transportmetrica A: Transport Science 2019;15(2):1487-1521. 24. Sharma B, Hickman M, Nassir N. Park-and-ride lot choice model using random utility maximization and random regret minimization. Transportation 2019;46(1):217-232. 25. Boeri M, Longo A. The importance of regret minimization in the choice for renewable energy programmes: Evidence from a discrete choice experiment. Energy Economics 2017;63:253-260. 26. Chorus C. Random regret minimization: an overview of model properties and empirical evidence. Transport reviews 2012;32(1):75-92. 27. Chorus CG, Rose JM, Hensher DA. Regret minimization or utility maximization: it depends on the attribute. Environment and Planning B: Planning and Design 2013;40(1):154-169. 28. Goodman JK, Cryder CE, Cheema A. Data collection in a flat world: The strengths and weaknesses of Mechanical Turk samples. Journal of Behavioral Decision Making 2013;26(3):213-224. 29. Huff C, Tingley D. “Who are these people?” Evaluating the demographic characteristics and political preferences of MTurk survey respondents. Research & Politics 2015;2(3):2053168015604648. 30. Buhrmester M, Kwang T, Gosling SD. Amazon's Mechanical Turk: A new source of inexpensive, yet high-quality, data? Perspectives on psychological science 2011;6(1):3-5. 31. Friedman CF, Proverbs-Singh TA, Postow MA. Treatment of the immune-related adverse effects of immune checkpoint inhibitors: a review. JAMA oncology 2016;2(10):1346-1353. 32. US Department of Health and Human Services. "Common terminology criteria for adverse events (CTCAE) version 4.0." National Institutes of Health NCI, no. 03 (2009). In. 33. Savage LJ. The theory of statistical decision. Journal of the American Statistical association 1951;46(253):55-67. 34. Bierlaire M. BIOGEME: a free package for the estimation of discrete choice models. 35. van Cranenburgh S. Choice Tutorial 12: Using Latent GOLD choice to Estimate Random Regret Minimization-based Discrete Choice Models. 2016. 36. Ben-Akiva M, Swait J. The Akaike likelihood ratio index. Transportation Science 1986;20(2):133- 136. 37. Kahneman D. Tversky A.(1979). Prospect theory: an analysis of decision under risk 1979:263- 292. 83 Chapter 5: Summary and future research directions This dissertation presents three studies that contribute to the evaluation of cancer treatments. In chapter 2, we performed one retrospective data analysis using the SEER-Medicare database. Our results indicated that the use of adjuvant chemotherapy among patients with stage II colon cancer was decreasing over time (2004 to 2011). Moreover, using instrumental variable regression analysis, we showed no definitive survival benefit of this treatment. It is encouraging to observe that the use of this controversial treatment was declining, as providers were cautious in not delivering this treatment to frail patients. Future studies should investigate this treatment pattern in more recent years. Moreover, in our study, due to the restriction of the encrypted data, we developed our instruments based on surgeons and health service areas to capture the unobservable treatment preferences. We suggest future studies request the restricted data, such as the zip codes for patients and the NPI for physicians. We expect that, based on these variables, the instruments would be more closely related to patients’ treatment decisions. In Chapter 3, we conducted one discrete choice experiment and assessed the treatment preferences among patients with non-small cell lung cancer. We quantitatively presented the relative importance that patients placed on five treatment characteristics: one-year progression- free survival time, monthly out-of-pocket costs, control of symptoms, and three treatment-related adverse effects: tiredness, skin itching, and hair loss. We estimated that patients were willing to pay $ 3,290 (95% CI: $2,550 to $4,030) for a one-month increase in progression-free survival time. Based on our estimated coefficients, we predicted that patients would be willing to pay up to $ 1,350 per month for osimertinib after the patents of erlotinib and gefitinib expire. Future work in this area could enlarge sample size and integrate complex models to investigate the 84 relationship between patients’ sociodemographic characteristics and their treatment preferences. Such a study would provide more detailed information for policymakers to understand patients’ heterogeneous treatment preferences based on their sociodemographic characteristics. In Chapter 4, we conducted another discrete choice experiment and assessed the treatment preferences for advanced lung cancer among members on Amazon Mechanical Turk. We empirically tested the µRRM’s role in differentiating which attribute was processed using the utility-maximization rule and which attribute was processed using the regret-minimization role. Then we presented the process of develop a hybrid model following guidance from the estimated scale factors. Our empirical analysis confirmed that the WTP derived from the RUM was constant across different scenarios, and the WTP estimates from the hybrid model were context- dependent. Future studies could explore more flexible models to explain decision behavior with psychological insights, such as the relative advantage maximization (RAM) approach. Both DCE studies present stated treatment preferences, we suggest future study combine revealed and stated preference data and investigate the treatment preferences using a richer framework.
Abstract (if available)
Abstract
Cancer is one of the leading causes of death in the United States (US). It has been estimated that in 2019, there will be 1,762,450 cases of new cancer (all sites), and 606,880 deaths will be attributable to this disease. Fortunately, breakthroughs in scientific research have contributed to an increased number of approved new cancer drug therapies. In addition to conventional chemotherapy, patients with cancer now can be provided with multiple treatment alternatives, including targeted therapy and immunotherapy. With the evolving landscape of cancer treatment, there is a growing need to develop more studies to examine and compare the outcome of cancer therapies. Moreover, cancer treatment imposes a substantial burden on patients and their families, including decreased quality of life and elevated cost burden. Therefore, researchers need to conduct studies to understand and assess patients’ preferences towards treatment, which will help to optimize their treatments and support the best practices of cancer care. This dissertation aims to evaluate the outcomes of cancer treatments, better understand the patient’s treatment preferences, and investigate new approaches in explaining treatment decision-making. ❧ The first study provided an in-depth analysis of SEER-Medicare data for adoption and benefit of adjuvant chemotherapy in patients with stage II colon cancer who underwent curative resection after approval of oxaliplatin. The results showed a decline in the use of adjuvant chemotherapy from 2004 to 2011. Using instrumental variable regression analysis, this study found no definitive survival benefit in adjuvant chemotherapy recipients. The second study employed one discrete choice experiment to help understand the treatment preferences among patients with Non-Small Cell Lung Cancer at Norris Comprehensive Cancer Center (Los Angeles). The results showed that progression-free survival time, out-of-pocket cost and quality of life outcomes significantly impacted patients’ decision-making. The study estimated that patients were willing to pay $ 3,290 (95% CI: $2,550 to $4,030) for a one-month increase in progression-free survival time. The third study investigated the role of a new decision-making model—random regret minimization—in generating context-dependent willingness to pay (WTP) estimations using one discrete choice experiment on the Amazon Mechanical Turk platform. The empirical analysis confirmed that the WTP derived from the random utility maximization (RUM) was constant across different scenarios and the WTP estimates from the new approach were context-dependent. ❧ Together, the studies suggested that the management of cancer treatments could be adjusted to meet clinical goals and individual needs. The research design and empirical methods presented here could be appropriate for further research regarding treatment optimization in other cancer types.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Health-related quality of life correlates of the treatment decision making process of newly diagnosed prostate cancer patients
PDF
An empirical evaluation of breast cancer treatment decision making
PDF
Health economics and outcomes research for informed decision making in rapidly evolving therapeutic areas
PDF
Documenting the impact of under-treatment for chronic disorders
PDF
Evaluating treatment options for metastatic, castration-resistant prostate cancer: a comprehensive value assessment
PDF
The value of novel antihyperlipidemic treatments in the U.S. healthcare system: Reducing the burden of cardiovascular diseases and filling the gap of low adherence in statins
PDF
Evaluating approaches to reduce inappropriate antibiotic use in the United States
PDF
Molecular signature of aggressive disease and clonal diversity revealed by single-cell copy number analysis of prostate cancer cells across multiple disease states
PDF
US biosimilar market entry and uptake: US commercial payer preferences in access decision making
PDF
Survival trends and related outcomes of survivors of childhood and young adult cancer
PDF
Association between informed decision-making and mental health-related quality of life in long term prostate cancer survivors
PDF
Mining an ePR system using a treatment plan navigator for radiation toxicity to evaluate proton therapy treatment protocol for prostate cancer
PDF
Economic, clinical, and behavioral outcomes from medical and pharmaceutical treatments
PDF
Addressing federal pain research priorities: drug policy, pain mechanisms, and integrative treatment
PDF
New approaches using probabilistic graphical models in health economics and outcomes research
PDF
Investigating racial and ethnic disparities in patient experiences with care and health services use following colorectal cancer diagnosis among older adults with comorbid chronic conditions
PDF
Burden of illness in hemophilia A: taking the patient’s perspective
PDF
Childhood cancer survivorship: parental factors associated with survivor's follow-up care behavior and mental health
PDF
Outcomes of antibiotic use among children with acute respiratory tract infections
PDF
Racial/ethnic differences in colorectal cancer patient experiences, health care utilization and their association with mortality: findings from the SEER-CAHPS data
Asset Metadata
Creator
Jiao, Xiayu
(author)
Core Title
Evaluating cancer treatments with health economics methods
School
School of Pharmacy
Degree
Doctor of Philosophy
Degree Program
Health Economics
Publication Date
12/17/2019
Defense Date
10/08/2019
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
cancer treatments,discrete choice experiment,Health Economics,OAI-PMH Harvest,treatment decision-making
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Hay, Joel (
committee chair
), Fox, Steven (
committee member
), Nieva, Jorge (
committee member
)
Creator Email
jiaoxiayu@gmail.com,xjiao@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c89-256212
Unique identifier
UC11673181
Identifier
etd-JiaoXiayu-8092.pdf (filename),usctheses-c89-256212 (legacy record id)
Legacy Identifier
etd-JiaoXiayu-8092.pdf
Dmrecord
256212
Document Type
Dissertation
Rights
Jiao, Xiayu
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
cancer treatments
discrete choice experiment
treatment decision-making