Patients and treatment regimens
We analyzed patient-level data from a multi-institutional, open-label RCT (JFMC37-0801)  wherein patients aged 20–80 years with curatively resected stage III (T1–4, N1–2, M0) colon cancer were randomly assigned to either 12 or 6 months of capecitabine. Capecitabine (1250 mg/m2) was orally administered twice daily after meals for 14 consecutive days, followed by 7 days of rest; this 3-week period comprised one course. The study group (12 M group) and the control group (6 M group) received 16 and 8 courses, respectively. The assigned treatment was started within 8 postoperative weeks. After completing the treatment protocol, patients were followed up according to a predefined surveillance schedule. The study was conducted in accordance with the Declaration of Helsinki and the Ethical Guidelines for Clinical Research in Japan; it was approved by the institutional review boards of each participating institutions. Written informed consent was received from all patients before their enrollment. Additional details, including dose modifications, were provided previously .
Framework of the economic analysis
The cost-effectiveness analysis of the 12-month capecitabine compared with the 6-month regimens was conducted in accordance with the guidelines on cost-effectiveness evaluation disseminated by Japan’s Ministry of Health, Labour and Welfare (MHLW) . The health outcomes of each intervention were evaluated in terms of quality-adjusted life-years (QALYs). The analysis was conducted from the Japanese public healthcare payer’s perspective and included only direct medical costs. A lifetime horizon was used, and the discount rate was set at 2% for both costs and health outcomes. The cost-effectiveness was determined by the incremental cost-effectiveness ratio (ICER), and the ICER reference value to be judged as cost-effective was determined to be JPY 5 million (USD 46,296) per QALY, referring to the willingness-to-pay threshold in Japan [14,15,16]. Unit costs were based on the 2018 Japanese fee schedule and drug tariffs, each defined by the MHLW at an exchange rate of USD 1 = JPY 108 (December 2019), as reported by the Bank of Japan . Analyses were performed using SAS® 9.4 (SAS Institute Inc., Cary, NC, USA.) and Microsoft Excel® for Office 365 (Microsoft Corp., Redmond, WA, USA).
The analysis was conducted using a partitioned survival analysis, comprising area under the curve interpretation from a set of mutually exclusive survival curves to determine state membership, exploiting the unidirectional nature of transitions in a progressive model . The analysis model considered three health states: stable disease, post-metastasis, and death. Stable disease was defined as without metastasis or death from any cause. In the JFMC37-0801 study, RFS was defined as survival from randomization without recurrence or death from any cause. Thus, OS can be divided into pre-metastasis and post-metastasis with consideration of RFS and OS.
Although some colon cancer patients develop metastases, some cancers can be completely cured. The population was thus a mixture of two subpopulations of cured patients and uncured patients, and RFS and OS survival were expected to differ between the two subpopulations. RFS and OS must be estimated with the assumption that they are different for each subpopulation; however, standard survival models do not assume two different populations. The cure model enables covariates to have different influences on cured patients and uncured patients . The cure model approach was applied to estimate the RFS and OS survival curves. The survival function can be shown in the following way:
S (t) = p + (1 − p) S(t) (where p is the probability of cure).
The cure rate was estimated by the RFS data from the JFMC37-0801 study. The definitions of OS and RFS in the JFMC37-0801 study included other causes of death. Immediately after the start of the study, the proportion of cancer-related events such as cancer recurrence and cancer-related death is expected to be high, while the proportion of other causes of death is expected to increase over time. OS and RFS in non-cured patients were estimated separately for “cancer-related events” and “other causes of death” because the breakdown of events defined as OS and RFS in non-cured was expected to change over time. The survival function for non-cured patients was investigated in four distributions: Weibull, log-logistic, log-normal, and exponential. The suitability of the fitted models was assessed by a combination of statistical fit, clinical plausibility, and visual examination of the parametric model curves . Statistical goodness-of-fit was assessed based on the Akaike information criterion (AIC) and Bayesian information criterion (BIC), with smaller values indicating a better fit. For the statistically well-fit models, the clinical plausibility of the slope and the curve tails were assessed. To account for increased mortality with aging, the risk of other causes of death after 60 months was assumed to be equivalent to that of the general population of the same age, and a nonparametric survival function with life tables for the general population was applied. For cancer-related deaths, the estimated parametric curve was applied throughout the lifetime. The mortality in cured patients was given as a nonparametric mortality using life tables for the general population.
The expected QALY obtained for each treatment was calculated as the survival proportion for each health state multiplied by the corresponding utility value. The utility value for stable disease was taken from an auxiliary study to the JFMC37-0801 study . Patients who agreed to participate in a health-related QOL (HRQOL) survey were evaluated using the EuroQOL 5 dimensions 3-level (EQ-5D-3L) tool at the start of the protocol treatment, and after 3, 6, 9, 12, 15, 18, 24, 36, 48, and 60 months. When a patient relapsed, the patient was excluded for ethical reasons. The EQ-5D-3L scores were converted to utility values using Japanese tariffs . After we calculated basic statistics, the utility values were applied to a linear mixed model of repeated measures. Utility values were estimated for each treatment group up to 12 months because HRQOL might vary between the treatment groups due to divergent frequencies and severity of AEs.
After 12 months, the estimation was performed without distinguishing between treatment groups. The utility values were adjusted with respect to baseline score, treatment, time, and treatment-by-time interaction during the treatment period, and they were only adjusted with respect to baseline score and time after 12 months. The random effects of individual patient factors were included in the model. Relapse and death were censored, and other missing values were supplemented by the last observation carried forward method. After the study period (60 months), the utility value of stable disease was set using values from the general population  because the estimated utility at 60 months was higher than that of the general population, and patients with a recurrence-free period greater than 5 years were assumed comparable to the general population.
Since the QOL data were not captured after patients relapsed, the utility value for the post-metastasis period, set at 0.705, was taken from a published report by Huxley et al.  through a targeted literature review.
Medical resource utilization
The cost parameters were based on patient-level data from the JFMC37-0801 study  and its auxiliary study. Capecitabine costs were calculated using the actual dose data recorded in the JFMC37-0801 study’s case report form. The costs of stable disease and post-metastasis were calculated using the claims data obtained from the patients who agreed to participate in the HRQOL survey. The medical fee reimbursement claimed by medical institutions was selected as each state cost, which included the monthly total billed amounts for all inpatient and outpatient medical interventions [visit costs, tests, treatments, including drugs (excluding capecitabine) and treatment for AEs].
The annual costs of stable disease were calculated separately by treatment group from baseline to 2 years. The total cost of months 1–12 for each treatment group was defined by the cost of stable disease in the first cycle, and months 13–24 were used to define the cost of stable disease in the second cycle. Thereafter, both groups were tabulated together because the observational costs during the stable disease period were assumed to be equal in both groups. The costs of stable disease were not considered after 5 years, as Japanese guidelines recommended a monitoring period of 5 postoperative years . The expected cost of the stable disease obtained from each treatment group was calculated as the survival proportion in the RFS state multiplied by the costs of stable disease.
The costs of the post-metastasis period were calculated using data from relapsed patients tabulated on a monthly basis. There was a trend toward higher costs immediately after recurrence and lower costs thereafter, indicating that annual costs are not constant. Given that the model was a partitioned survival model, it was not possible to change the cost for patients with recurrence from year to year. The monthly cost was multiplied by the survival proportion in each month for relapse (up to 60 months), and the event cost for relapse cases was set (undiscounted).
Uncertainty over the post-metastasis costs were assessed using one-way sensitivity analysis (JPY 1–8 million) because it could only be calculated from 25 patients, and the discount rate was not considered in the analytical model, resulting in high uncertainty. In addition, since the post-metastasis utility value was not available in the JFMC37-0801 study, and value reported overseas by Huxley et al. was used, a one-way sensitivity analysis was also performed for the post-metastasis utility value (0.600–0.800). Since data on the variability of the values were not available, a wide range of assumed values was used in the analysis. Furthermore, to assess the impact of the distribution selection of the survival function, a scenario analysis was conducted in which different distributions were selected.
The bootstrap method (10,000 resamples) was used for the probabilistic sensitivity analysis, and a cost-effectiveness acceptability curve was created . In the bootstrap method, all parameters obtained from JFMC37-0801 were resampled and subjected to sensitivity analysis, except for the utility for disease progression taken from the literature. The utility for disease progression was randomly sampled by probability distribution following a beta distribution.