Discrete choice experiment

Understanding the impact of the crucial attributes of treatments on patients’ decision-making is more informative than merely understanding their preference of the actual product as a whole. Attributes may have different weights that influence patients’ evaluation of the trade-offs between treatments. Although traditional surveys can be used to obtain quantitative preferences about products as a whole, it is difficult to compare the relative importance of these individual attributes. DCE is a mature method to easily measure the level of risk/benefit that patients would take/give up for one attribute in exchange for the benefit of another. By using this approach, we may be able to ascertain the dominating factors that influence patients when choosing treatments, as well as the price they are prepared to spend to be free of or obtain specific attributes. Therefore, the results could assist clinicians, pharmaceutical companies, governments, and other stakeholders in adjusting treatment plan, improving products, and offering more reasonable pricing, which could improve medication adherence. This study used a DCE design to quantitatively assess patients’ preferences of the most important attributes for oral anticoagulants. The entire experiment adhered to the recommendations of the International Society for Pharmacoeconomics and Outcome Research protocol for conducting DCE [18].

Target population and sample size

A convenience sampling approach was used to recruit patients who were (i) diagnosed with AF; (ii) aged ≥18 years old; (iii) with congestive heart failure (C), hypertension (H), age ≥ 75 (A₂), diabetes mellitus (D), prior stroke or transient ischemic attack or thromboembolism (S₂), vascular disease (V), age between 65 and 74 (A), sex category (Sc) (CHA₂DS₂-VASc) ≥1 for men and CHA₂DS₂-VASc ≥2 for women; and (iv) without communication impediments.

Attribute and level identification

A total of 7 attributes of oral anticoagulants (Table 1) were identified based on the evidence from previous studies [12,16,17,19–21], in addition to the ranking by an expert panel with 40 cardiologists and 40 eligible patients. Two levels were assigned to the attributes of antidote and food–drug interaction. Patients who use warfarin drugs are subject to monthly blood monitoring although this is not necessary for NOAC users. We used 4 levels, every 1/3/6 month(s) or not required/necessary, for this attribute in order to determine the preferred monitoring frequency for clinical application. Four levels were used for the risk of nonfatal major bleeding (0.7/3.1/5.5/7.8[%]), the risk of nonfatal stroke or systemic embolism (0.6/3.2/5.8/8.4[%]), and the risk of nonfatal acute myocardial infarction (AMI) (0.2/1.0/1.8/2.5[%]) [22–27]. The highest and lowest rates of the aforementioned 3 risk events when using oral anticoagulants reported from observational studies were used as the ceiling and baseline levels of attributes, respectively, with the ranges equally divided into 4. The monthly out-of-pocket cost of NOAC for a patient with medical insurance is approximately 150 to 300 RMB (23 to 47 USD). As most patients would have medical insurance, the levels for this attribute were expressed as 0/120/240/360 RMB.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Selected attributes of oral anticoagulants.

https://doi.org/10.1371/journal.pmed.1003730.t001

DCE design

To prevent an impractically large sample size reaching sufficient power, we used a D-Efficient design to generate a set of choices consisting of 16 combinations from a full factorial design, which had 2 × 2 × 4 × 4 × 4 × 4 × 4 = 4,096 hypothetical profiles for each alternative in this study. In order to give a manageable workload to each patient while harnessing the statistical properties, 16 scenarios with different combinations of hypothetical profiles of alternatives were randomly divided into 2 blocks, with each block including 8 scenarios. Eligible patients were randomly allocated to one of the 2 blocks.

The names of alternatives could distract patients’ attention from the targeting attributes and bias their preferences, especially when the population has a different understanding about the products [28,29]. Therefore, patients were asked to choose one preferred oral anticoagulant from 2 unlabeled hypothetical treatments. One “opt-out” option (i.e., no anticoagulant therapy) would be chosen if patients disliked either treatments based on the provided attribute levels. This could provide a more realistic preference [30].

A testing scenario with a dominated alternative was included at the beginning to assess patients’ understanding (S1 File). Investigators would immediately explain the meaning of levels of attributes to those who did not choose the dominated alternative. Those patients were allowed to continue answering the questionnaire only when their understanding were confirmed by investigators. DCEs were generated by Ngene V1.1.2 (ChoiceMetrics, W201/599 Pacific Highway, St Leonards, Sydney, NSW 2065, AU) (S2 and S3 Files).

Sample size

According to Orme’s rule of thumb formula, minimum sample size = [31]. Given the conditions of 2 alternatives (excluding opt-out), and maximum 4 levels for attributes, as well as 8 scenarios presented to each patient to obtain the largest minimum required sample size, at least 125 participants in total are required. To ensure sufficient power, we aimed to recruit 300 patients for each site [31].

Data collection

Eligible patients identified between May 2018 and December 2019 by cardiologists were transferred to trained investigators for the subsequent investigation. The purpose of this study, basic information about oral anticoagulant therapy, and how to complete the DCE questionnaire were explained by the site investigator. All patients who signed the consent forms were included in the analyses. Face-to-face interviews were conducted via paper-based questionnaires (S4 File). Patient demographic characteristics, such as age, sex, occupation, education level, monthly income after tax, medical insurance type, history of cardiovascular medication, history of diseases, self-evaluated health score, were also collected for analyses (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of recruited patients.

https://doi.org/10.1371/journal.pmed.1003730.t002

Analyses

Mixed logit regression model was used for the main analysis to estimate patients’ preferences on each attribute. Considering the assumption of independence for outcomes in mixed logit regression model would have been violated because the selection of “opt-out” option would be highly dependent on the selection of “drug A” and “drug B”, scenarios with opt-out selected were removed from the main analysis. Antidote and food–drug interaction were treated as binary; the frequency of blood monitoring was treated as ordinal; all of the risk attributes and the monthly out-of-pocket cost were treated as continuous variables. Factors that could influence patients’ understandings for questionnaire, such as education level, income level, and city, were adjusted. Factors that could represent patients’ previous experiences on anticoagulant therapy, such as self-evaluated health score, histories of cardiovascular disease/other vascular disease/any stroke/any bleeding, and use of anticoagulant/antiplatelet, were adjusted. As common confounders, age and sex were also adjusted. Random effects for each attribute were taken into consideration to minimize the effects from residual confounders that induced heterogeneity in patients, except for out-of-pocket cost, which was treated as a fixed effect to calculate WTP. The correlation between any pair of attributes was also adjusted in the model. McFadden Pseudo R² was used to evaluate the goodness of fit for models [32]. Marginal willingness to pay (mWTP) was calculated by assessing the ratio of the preference of other attributes to the preference of out-of-pocket cost (Table 3). Subgroup analyses by sex, education level (under or above high school), income level (under or above 5,000 RMB [775 USD] per month), with or without history of any stroke/bleeding, and cities were performed (Table 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Preference weights estimated by mixed logit regression model and mWTP (RMB^*) in main analysis.

https://doi.org/10.1371/journal.pmed.1003730.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Preference weights estimated by mixed logit regression model and mWTP (RMB^*) in subgroup analyses.

https://doi.org/10.1371/journal.pmed.1003730.t004

Several sensitivity analyses were performed to test the robustness of our results. Firstly, patients who failed the testing scenario were removed from the analysis to assess the quality of our data (S5 File). Furthermore, to determine more specific preferences, the frequency of blood monitoring was treated as a categorical variable (S6 File). Block of scenarios was adjusted in mixed logit regression model to assess the potential impact of blocks on preferences (S7 File). An analysis stratified by outpatient and inpatient setting was conducted to assess the potential influence from different settings (S8 File). As medical insurance may influence patients’ preferences, an analysis with the exclusion of all patients who did not have any medical insurance was performed (S9 File). Two additional statistical models, nested logit model and multinomial logit model, were used to assess the influence of the opt-out option (S10 and S11 Files).

The main analysis, subgroup analyses, and sensitivity analyses (to test the robustness of different models) were prespecified, conducted, and reported as planned (S10 and S11 Files). Additional sensitivity analyses (S5–S9 Files) that had been added to test the robustness of the results against the main results were data driven. All statistical analyses were conducted using R 3.6.3 (R Core Team 2020, Austria). Patient characteristics were summarized as mean (SD) for continuous variables and in frequencies (percentage) for categorical variables. All hypothesis tests were 2 sided, with a significance level of 0.05. This study was approved by the ethics committees of both Chinese People’s Liberation Army General Hospital and The University of Hong Kong-Shenzhen Hospital.

Characteristics of patients (Table 2)

After excluding one patient who did not answer any question, a total of 506 patients (299 from Beijing and 207 from Shenzhen), with mean age 70.3 years, of which 42.1% were women, were successfully recruited and completed consent forms. Regardless of testing scenario, each patient was given 8 scenarios with 2 treatment alternatives and one opt-out option to choose from, and a result of 12,144 (506 × 8 × 3) completed records were expected. After removing 201 empty responses, 11,943 records were kept for further analyses. The first and second blocks were answered by 254 and 252 patients, respectively (S4 File). A total of 36 out of 506 patients (7.1%) did not choose the dominant drug in the testing scenario.

Main analysis (Table 3)

The positive and negative values of coefficients represent patients’ preference and dislike for the corresponding attribute, respectively. The greater the absolute value of the coefficient, the more weight for changes per unit of that attribute were taken into consideration when patients evaluated the trade-offs.

Compared to other attributes, the relatively higher weights for the risk of nonfatal AMI (β: −1.03; 95% CI: −1.31, −0.75; p < 0.001), the risk of nonfatal stroke or systemic embolism (β: −0.81; 95% CI: −0.90, −0.73; p < 0.001), and the risk of nonfatal major bleeding (β: −0.69; 95% CI: −0.78, −0.60; p < 0.001) indicate that the issues patients had the most concern about during anticoagulant therapy were safety and effectiveness. The impact of the frequency of blood monitoring on decision-making was the weakest (β: −0.31; 95% CI: −0.39, −0.24; p < 0.001), and patients disliked frequent hospital visits. Patients preferred treatment with the antidote (β: 0.43; 95% CI: 0.17, 0.69; p = 0.001), while avoiding food–drug interactions (β: −0.63; 95% CI: −0.88, −0.38; p < 0.001).

Patients were willing to pay most of the cost for oral anticoagulant therapy: 798 RMB (β: −798; 95% CI: −1,247, −518), 632 RMB (β: −632; 95% CI: −880, −488), and 536 RMB (β: −532; 95% CI: −762, −403), respectively, to reduce 1% of the risk of nonfatal AMI, nonfatal stroke or systemic embolism, and nonfatal major bleeding. They began to hesitate about paying more to be free of food–drug interaction (β: −489; 95% CI: −764, −288) or to obtain the extra antidote for side effects and reactions (β: 338; 95% CI: 125, 639) and were willing to pay only the minimum to reduce the frequency of blood monitoring (β: −244; 95% CI: −347, −180).

Subgroup analyses (Table 4)

Men were slightly more concerned about almost all of the attributes compared to women and were willing to pay double that of women. The impact of food–drug interaction on men (β: −0.96; 95% CI: −1.36, −0.56; p < 0.001) was much stronger than that on women (β: −0.34; 95% CI: −0.73, 0.05; p = 0.09), and they were willing to pay over 4 times more than women (805 versus 174 RMB [125 versus 27 USD]).

As is widely known, education level correlates with income level. The similarity in patterns of preferences on all attributes could also be observed for both subgroups; patients with lower educational attainment (at or under high school) or who had a lower income (≤5,000 RMB/month) were less concerned about all of the attributes and were only willing to pay the minimum to obtain extra benefits or get rid of disadvantages. mWTP varied widely among patients with higher levels of education and income.

History of any bleeding or stroke influenced patients’ preference. Anyone with a history valued all attributes at the same high weights and were willing to pay up to 5 times of those without a history. For this group, food–drug interaction was the most concerning attribute (β: −2.47; 95% CI: −3.92, −1.02; p < 0.001).

We stratified the analysis by cities to test heterogeneity. Compared to Shenzhen, the effect size of all attributes on patients from Beijing was chiefly doubled. The food–drug interaction had the strongest impact on patients from Beijing (β: −1.82; 95% CI: −2.56, −1.07; p < 0.001). However, they were only willing to pay less than the amount that patients from Shenzhen were prepared to pay for most of the attributes.

Sensitivity analyses (S5–S11 Files)

First, we removed patients who did not initially select the dominated alternative in the testing scenario to assess data quality. Preferences and mWTP for each attribute were consistent with the primary analysis (S5 File).

With consideration that the linear assumption might be violated in the ordinal variable frequency of blood monitoring since the intervals between each level were unequal, we also treated this as a categorical variable (S6 File). Compared to nonnecessity for monitoring, patients preferred less frequent 6 monthly hospital visits (β: 0.43; 95% CI: 0.18, 0.69; p < 0.001). The extent of dislike deteriorated as the frequency increased (every 3 months β: −0.30; 95% CI: −0.54, −0.05; p = 0.02 and every 1 month β: −0.85; 95% CI: −1.09, −0.61; p < 0.001).

The consistent results of the sensitivity analyses, which adjusted blocks of scenarios, stratified by outpatient and inpatient settings, and excluded patients who did not have any medical insurance, from main analysis, indicated that there would not be obvious influence from those aforementioned factors on patients’ preferences (S7–S9 Files).

By adopting a nested logit model to assess the effects of opt-out on decision-making, the preferences for all attributes were principally similar to that in the main analysis. However, due to the inability to deal with the correlated alternative selection, the results from multinomial logistic regression were not consistent with the main analysis (S10 and S11 Files).

Ethical approval and informed consent

This study was approved by Medical Ethics Committees of the Chinese People’s Liberation Army General Hospital (Reference number: S2018-043-01) and HKU-Shenzhen Hospital (Reference number: [2018]78).