Model structure
A decision analysis Markov model for EC with 11 health states was built in Treeage Pro (2019). Only ESCC was modeled. The health states included normal, LGIN, IC, SM, moderate cancer stage (Mod), advanced cancer stage (Adv), disease-free survival of IC (DFS_IC), disease-free survival of SM (DFS_SM), disease-free survival of moderate stage (DFS_Mod), disease progression free of advanced stage (PFS_Adv), and death. Here, in the nonscreening cohorts, “normal health” was assumed to be the state of non-EC, while it was assumed to be the state of health without LGIN and EC in the screening cohorts. IC included high-grade intraepithelial neoplasia, while moderate stage included stage IB, stage II, and stage III. Stage IV was classified as an advanced cancer stage. Overall, IC and SM both constituted the early EC stage, while moderate and advanced stages were identified as the invasive EC stage. eFigure 1 summarizes the state transition processes, with the arrows presenting the transitions between states.
Individuals aged 40 to 69 years were assumed to be the participants. They were classified into six age groups separated by five-year intervals (ages 40-44, 45-49, 50-54, 55-59, 60-64, and 65-69 years). Cohort simulation was performed until the cohort age reached 79 years or until hypothetical death. The model assumed a one-year cycle length. A hypothetical cohort was generated with 100,000 individuals assigned to each age group. Three different strategies were assessed for each cohort: (1) nonscreening, a strategy that assumes that all the individuals were not screened, and no endoscopic follow-up for LGIN was conducted. Patients were not diagnosed until clinical symptoms appeared, which is referred to as passive treatment. Over 90% of the patients in these cohorts had progressed into moderate and advanced stages. (2) Screening with annual follow-up for LGIN, a strategy that assumes all the individuals undergo the one-time standard endoscopic screening. The screening could identify the LGIN and EC. Approximately 90% of EC cases were diagnosed at an early stage. Moreover, we assume that all LGINs underwent annual endoscopic surveillance. (3) Screening without follow-up, a strategy that assumes all the individuals underwent the one-time standard endoscopic screening, which could identify the LGIN and EC. Most of the EC was in the early stage. However, endoscopic surveillance of LGIN individuals was not required. It was assumed that all the patients diagnosed by the three strategies had the correct diagnosis and adhered to the standardized treatment scenarios, including the standardized clinical follow-up recommended by the Guidelines for the Diagnosis and Treatment of EC 2018 in China.
Effectiveness outcome
QALY represented the effectiveness, and the incremental cost-effectiveness ratio (ICER) served as the economic evaluation indicator. ICER indicates the additional cost per unit of additional effectiveness. The calculation formula was Cost-effectiveness analyses were used for the comparisons between the competing strategies, including “absolutely dominated strategy”, an option that had both more costs and less effectiveness; “extended dominated strategy”, an option that was less costly and less effective than the alternative but had a higher ICER; and the “undominated best strategy”, an option that was cost effective with an ICER between 1 and 3 GDP per capita based on the criterion recommended by the WHO[21]. Other outcomes assessed included costs, EC cumulative incidence, and mortality. Moreover, the willingness-to-pay (WTP) was set as three times the gross domestic product per capita (GDP) (USD 51,000) in 2017 in China.
Cost and utility information
Costs were estimated from a social perspective. The estimation consisted of screening costs, treatment costs, transportation, and wage loss of patients and relatives due to hospital visits. Screening costs were calculated using the data of the screening program in Zhejiang Province, while treatment costs were extracted from the electronic medical record at Zhejiang Cancer Hospital. The average annual costs of the disease-free survival states were computed by the following formula: where C was the annual costs of disease-free states, was the average cost per visit, N was the annual average number of hospital visits, i was the number of physical examination items, and p was the price per item. N and i were suggested by the “Guidelines for the Diagnosis and Treatment of Esophageal Cancer 2018 in China”, and p was based on the price of medical services in provincial public hospitals in Zhejiang Province[22]. Every patient was assumed to have one accompanying relative. All costs were measured in the 2017 Chinese currency and were converted into US dollars using the purchasing power parity of 3.506 in 2017[23]. All the items related to costs in the study will inflate at the same inflation rate of 4.7%[24]. State-specific utilities were extracted from published papers and the results of the screening. A discount rate of 5% was used for both costs and effectiveness [25, 26]. eTable 5 and eTable 6 display the state-specific costs and utilities.
Data analysis
Cohort initial probabilities
The initial cohorts’ probabilities for EC states in nonscreening cohorts were calculated according to the 2012 age-specific incidence of EC in Zhejiang Province, multiplied by the stage distribution at the time of diagnosis that was obtained from the hospital-based retrospective study[27]. And the probabilities for EC states in screening cohorts were computed by the age-specific EC detection rate multiplied by the stage distribution that was distinguished by screening, which was implemented in a total of 29762 residents aged 40-69 years from seven counties between 2010 and 2017 in Zhejiang Province. The initial probabilities of LGIN in screening cohorts were the age-specific LGIN detection rate obtained from the screening program, while it was assumed to be zero in the nonscreening cohorts since no LGIN would be diagnosed without screening. The probabilities for the normal state were one minus the sum of the probabilities for the other states for both scenarios. eTable 1 displays the age-specific incidence and detection rate of EC and the age-specific detection rate of LGIN, while table 2 addresses the stage distributions for both scenarios.
Transition probability between Markov states
The EC annual incidence was used to calculate the probabilities of transition from the normal state to the EC state for nonscreening cohorts, while in screening cohorts, the following formula was used to compute the adjusted EC annual incidence. The formula was Ia=Ip×RR, where Ip was the EC annual incidence, Ia was the adjusted EC annual incidence, and RR was the annual incidence probability ratio that was computed through the formula for conversion between rate and probability using the cumulative incidence of ESCC in the screening population versus the nonscreening group [16, 28]. Moreover, the stage distribution at the time of diagnosis by passive treatment was used to identify EC states in both scenarios.
The probabilities of transferring from LGIN to EC states were computed by the EC annual incidence among the LGIN (eTable 1) multiplied by the EC stage distribution. The major difference was that the EC stage distribution was identified by the results of the screening programs for follow-up cohorts, while it was identified by passive treatment for cohorts without follow-up. In addition, it was assumed that a proportion of LGIN patients transferred to a normal state for the follow-up strategies[29-35], while it was assumed that no individuals would transfer from LGIN to a normal state for strategies without follow-up. The calculation of EC incidence among the LGIN was the adjusted incidence according to the risk ratio of the EC incidence among LGIN compared to that of the true healthy population combined with the detected proportion of LGIN during screening. The risk ratio was 3.66, which was summarized from published papers[36-39]. No individuals would transfer from LGIN to EC states under the nonscreening scenario since no LGIN would be diagnosed due to the asymptomatic characteristics without the implementation of screening. Other probabilities of transitioning between EC states were collected from various kinds of literature (eTable 4).
The age-specific annual death probabilities for the normal state were defined as the difference between the all-cause mortality and EC death probabilities. All-cause mortality was obtained from the sixth population service survey, while EC mortality was obtained from the 2012 age-specific EC mortality in Zhejiang Province[27, 40]. LGIN was considered a precancerous lesion. The EC-specific 5-year survival was 100% for IC and DFS_IC; therefore, people with LGIN or IC were less likely to die from EC. Consequently, the mortality for LGIN, IC, and DFS_IC was assumed to be the same as that of the normal state (eTable 3). Death probabilities of SM and invasive cancer were identified from published papers, while we adjusted the mortality risk according to age (eTable 3).
Moreover, a cycle length of one year was chosen; therefore, all the probabilities of transitions between states are presented as one-year probabilities. Given the different follow-up periods in the various data sources, we used two-step calculations. First, we converted the t-year follow-up probabilities into one-year rates. Then, we calculated one-year probabilities using one-year rates. The following formula was used to calculate the relationship between rate and probability. where r indicates rate, p indicates probability, and t indicates the follow-up years [28].
Sensitivity analysis
Sensitivity analyses for cost-effectiveness screening strategies were conducted. Probabilistic sensitivity analyses (PSA) permit the joint uncertainty across all the parameters in the model to be assessed at the same time, which involves sampling model parameter values from the distribution imposed on variables in the model. Initial cohort probabilities and death probabilities are assumed to be beta distributions, while the discount rate and inflation rate are identified as triangular distributions. A gamma distribution was set for costs. Moreover, beta and Dirichlet were assigned to transition probabilities. One-way sensitivity analyses were simulated to observe that the impact of each parameter estimate varied independently and singly on the model results. The range of each parameter is assumed to be simply a “plausible’’ range. The following assumption was made. Initial probabilities, state transition probabilities, risk ratios, and health utility varied by±20% of the base case value, while costs changed by±30% of the base case value. In addition, 0-8% was simulated for the discount rate, and 3.2- 6.2% was used for the inflation rate.