Attribute nonattendance in COVID‐19 vaccine choice: A discrete choice experiment based on Chinese public preference

Abstract Objectives The global coronavirus disease 2019 (COVID‐19) pandemic has not been well controlled, and vaccination could be an effective way to prevent this pandemic. By accommodating attribute nonattendance (ANA) in a discrete choice experiment (DCE), this paper aimed to examine Chinese public preferences and willingness to pay (WTP) for COVID‐19 vaccine attributes, especially the influence of ANA on the estimated results. Methods A DCE was designed with four attributes: effectiveness, protection period, adverse reactions and price. A random parameter logit model with an error component (RPL‐EC) was used to analyse the heterogeneity of respondents' preferences for COVID‐19 vaccine attributes. Two equality constraint latent class (ECLC) models were used to consider the influence of ANA on the estimated results in which the ECLC‐homogeneity model considered only ANA and the ECLC‐heterogeneity model considered both ANA and preference heterogeneity. Results Data from 1,576 samples were included in the analyses. Effectiveness had the highest relative importance, followed by adverse reactions and protection period, which were determined by the attributes and levels presented in this study. The ECLC‐heterogeneity model improved the goodness of fit of the model and obtained a lower probability of ANA. In the ECLC‐heterogeneity model, only a small number of respondents (29.09%) considered all attributes, and price was the most easily ignored attribute (64.23%). Compared with the RPL‐EC model, the ECLC‐homogeneity model obtained lower WTPs for COVID‐19 vaccine attributes, and the ECLC‐heterogeneity model obtained mixed WTP results. In the ECLC‐heterogeneity model, preference group 1 obtained higher WTPs, and preference groups 2 and 3 obtained lower WTPs. Conclusions The RPL‐EC, ECLC‐homogeneity and ECLC‐heterogeneity models obtained inconsistent WTPs for COVID‐19 vaccine attributes. The study found that the results of the ECLC‐heterogeneity model considering both ANA and preference heterogeneity may be more plausible because ANA and low preference may be confused in the ECLC‐homogeneity model and the RPL‐EC model. The results showed that the probability of ANA was still high in the ECLC‐heterogeneity model, although it was lower than that in the ECLC‐homogeneity model. Therefore, in future research on DCE (such as the field of vaccines), ANA should be considered as an essential issue. Public Contribution Chinese adults from 31 provinces in mainland China participated in the study. All participants completed the COVID‐19 vaccine choice questions generated through the DCE design.


| INTRODUCTION
The coronavirus disease 2019 (COVID-19) pandemic was caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), a new virus that was not known before 2019. The pandemic has had a serious impact on the health system, and no individual was immune to the virus at the beginning of this pandemic, because of which the pandemic has still not been well controlled, with rising numbers of infections and deaths in many countries. There are two ways to control the pandemic: Multiple lockdowns that happened in the beginning or effective vaccination to achieve global herd immunity. China has achieved great success in vaccine research and development since the outbreak of the COVID-19 pandemic. 'But bringing a vaccine to market is only half the challenge; also critical is ensuring a high enough vaccination rate to achieve herd immunity'. 1 Although the public acceptance rate of the COVID-19 vaccine in China is extremely high compared with that in other countries, 2 studies show that the Chinese public acceptance rate of the vaccine has a downward trend. 3 Wang et al. 3  vaccine. 7 Therefore, the achievement of a vaccination rate of 75%-90% to achieve herd immunity 8 in China may be a considerable challenge. 9 It is crucial to understand the factors influencing vaccination and trade-offs among these factors, especially the preferences for COVID-19 vaccine attributes. To better understand the preferences, we include price attribute in this study, which is used to understand the relative value of the nonprice attributes.
The discrete choice experiment (DCE) is a widely used method in nonmarket valuation; its popularity is increasing because it can evaluate multiple attributes of a product simultaneously. 10 The theoretical basis for DCE comes from random utility theory and Lancaster's 11 consumer theory. According to Lancaster's consumer theory, the utility of the COVID-19 vaccine chosen by respondents does not come from the COVID-19 vaccine itself but from the multiple attributes of the COVID-19 vaccine, which are selected to be comprehensive and complete. DCE has been used to assess respondents' preferences and willingness to pay (WTP) for the attributes of vaccines such as the meningococcal vaccine, 12 40,41 In these cases, the influence of ANA on the research results should be considered.
The COVID-19 vaccine studied in this paper was relatively unfamiliar to the respondents. Therefore, this case study is actually an application of ANA to the COVID-19 vaccine.
There may be two main reasons for ignoring attributes. First, respondents ignore some attributes by adopting simplified heuristics to reduce the cognitive burden due to the complexity of choice tasks, which reflects the real ANA. 26,28,42 Second, respondents ignore some attributes because these attributes are not important to them, or are of low importance, which reflects the preference heterogeneity of respondents. 26,27,43 Traditional processing models for ANA, such as the equality constraint latent class (ECLC) model considering only ANA, would lead to confusion between ANA and preference heterogeneity and may incorrectly identify respondents with low preference as nonattenders. 24,44 Therefore, this paper adopts an ECLC model that takes both ANA and preference heterogeneity into account to obtain more reliable results.
The first objective of this paper was to determine the Chinese public preferences and WTP for COVID-19 vaccine attributes using DCE. The second objective was to take the COVID-19 vaccine as a case to study the impact of ANA on DCE estimation results. Specifically, we focus mainly on the following three points: Whether considering ANA will improve the goodness of fit of the model; whether considering preference heterogeneity will reduce the probability of ANA; and whether the model considering ANA and preference heterogeneity will yield inconsistent WTP results with the model without considering ANA and improve the reliability of the modelling results.

| Selection of the attributes and levels for DCE
The design of the DCE follows International Society for Pharmacoeconomics and Outcomes Research guidelines. 45 First, the research team reviewed the related literature, including papers on respondents' preferences for the COVID-19 vaccine and other vaccines using DCE, as well as literature on the influencing factors of COVID-19 vaccination. Based on the information obtained from the literature review, the research team held several discussions, and the four most frequently mentioned attributes that impact vaccine preference were selected: effectiveness, 46-48 protection period, [16][17][18] adverse reactions 16,49 and price. 16 The selection of attribute levels was based on existing studies on the COVID-19 vaccine and official reports of the COVID-19 vaccines announced in China. Specifically, the effectiveness of the Sinopharm COVID-19 vaccine reached 79.34% when it conditionally entered the market. 50 Clinical trials of the Sinovac vaccine, CoronaVac, in Turkey and Indonesia yielded 91.25% and 65.3% effectiveness, respectively. 51 Therefore, the effectiveness levels of this study were 65%, 80% and 95%. The protection period is still uncertain due to the short development time of COVID-19 vaccines. However, Sinopharm predicted that the protection period of its vaccines may reach 1-3 years. 52 Based on this information, we selected protection periods of 1, 2 and 3 years.
There were mild adverse reactions and no adverse reactions, 16,53,54 and the mild adverse reactions manifested mainly as local pain, redness and swelling at the injection site, transient low-grade fever, fever, and so forth. 55 Then, we invited two medical experts to participate in a discussion on the selection of attributes and levels. After the discussion, we used the above attributes and levels to build choice sets, and the price attribute levels selected were based on the pilot survey. Otherwise, other related information of the vaccine was specified in the questionnaire; that is, the hypothetical scenario positioned the COVID-19 vaccine as a Chinese vaccine that requires two injections with an interval of 14-28 days between them. 56 The attributes and levels can be found in Appendix S1.
Taking the difficulty of sample collection for 5,778 hypothetical scenarios into account, this article used an orthogonal experimental design to produce eight choice sets. To simulate the real market and reduce respondents' protest bias, we added the option of neither choice in each choice set, which means an opt-out option. 12,19 Additionally, the eight choice sets were randomly divided into two blocks in the questionnaire to mitigate the cognitive burden on respondents 57 ; hence, each respondent faced four choice sets. An example of a choice set is shown in Figure 1.
The first section of the questionnaire was the DCE, which first introduced the hypothetical scenarios in detail and then asked respondents to make choices among four choice sets. 2019, and whether they were employed in a medical-related industry) and health status (whether they had chronic diseases).

| Statistical analyses
In this paper, a random parameter logit model with an error component (RPL-EC) was used to study the heterogeneity of the respondents' preferences for COVID-19 vaccine attributes. The RPL-EC model relaxes the limitation on the independence of irrelative alternatives and assumes that the respondents' preferences for the vaccine are heterogeneous by assuming a random distribution of coefficients. 59,60 This paper adopts the distribution form commonly used in previous studies; the nonprice attributes are assumed to be normally distributed, the price attribute is assumed to be constant and 1,000 Halton draws are used. 40,[61][62][63] This paper specifies a common random error component in the random utility of the hypothetical options of vaccine 1 and vaccine 2 to capture any additional variance between the two hypothetical vaccine options. 33,38,64 In the RPL-EC model, the utility of respondent n choosing alternative COVID-19 vaccine j from choice set t is where X njt represents the vector of observable attribute levels related to the COVID-19 vaccine j; β′ n represents the vector of estimated individual-specific coefficients; ε njt represents the unobserved random term that follows an independently and identically distributed extreme value distribution; e nj is the error component, which follows the zeromean standard normal distribution; μ n is the coefficient of the estimated error component; and alternative specific constant (ASC) is an alternative specific constant representing the opt-out option, which adopts dummy coding. The opt-out option is coded as 1; otherwise, it is 0. ASC is separately added to the utility of the opt-out option to capture the potential current situation deviation. 65 All nonprice attributes adopt effect coding, while the price attribute was coded as a continuous variable. We also tested other forms of the price attribute. The model results can be found in the Appendices S10-S17.
The latent class (LC) model assumes that the choice behaviour of respondents depends on observed variables and latent heterogeneity that the analyst cannot observe. Therefore, the specification divides the population into several classes, each of which has the same preference, and the number of each class is endogenous. Compared with the RPL-EC model, the LC model assumes that the distribution of coefficients is discrete rather than continuous. Assuming that the population is divided into Q classes, the utility of individual n in class q where β q is a class-specific parameter vector, and the other variables have the same meaning as in Equation (1) of all classes to be equal in the same group but allows the change of attendance attribute coefficients between different groups. In this study, the two ECLC models were used to infer the ANA categories of respondents and to factor the impact of ANA into the estimations.
The WTP for COVID-19 vaccine attributes is calculated as follows 67 : where β nonprice denotes the coefficient of nonprice attributes and β price denotes the coefficient of price attribute. The 95% confidence intervals for the mean WTP are derived using the Krinsky and Robb 68 method with 5,000 draws.

| Preferences and WTP for COVID-19 vaccine attributes
In this paper, RPL-EC, ECLC-homogeneity and ECLC-heterogeneity models were developed. The regression results are shown in Table 5.
The goodness of fit for all models was compared using the Bayesian information criteria (BIC) and the Akaike information criteria (AIC).
Compared with the RPL-EC model, the ECLC-homogeneity model did not significantly improve the goodness of fit; however, the ECLCheterogeneity model did improve the fit and obtained the optimal goodness of fit. The relative importance of attributes was estimated, which refers to the difference between the respondents' most preferred level and the least preferred level of each attribute. The greater the relative importance value, the higher the importance of the attribute compared with other attributes, which is determined by the attributes and levels described in this study. The relative importance of attributes was expressed as percentages, and the sum of the percentages of all attributes was equal to 1. The results are shown in Figure 2. Among all models, effectiveness is the most important attribute, followed by adverse reactions and protection period. The exception is preference group 3 in the ECLC-heterogeneity model, for which the protection period is the most important attribute, followed by adverse reactions and effectiveness.
This paper also predicted the change in the probability of vaccination when only one attribute level changed compared with the base vaccine. For ease of comparison, we defined the base vaccine with 65% T A B L E 2 Class memberships from the ECLC-homogeneity model  In addition, the χ 2 test was used to test differences in sociodemographic characteristics among the four groups. 70 The results can be found in Appendix S5. An RPL-EC model incorporating all sociodemographic characteristics was developed, and the results showed no significant impact on respondents' vaccine choice (Appendix S6).
We tested RPL-EC models with interaction terms and the LC model.  75 Finally, the results of the stated ANA can be used as a predictor of ANA; it has been proven in some studies that stated ANA is an effective indicator of the probability of ANA. 71 In the future, researchers could pay more attention to ANA under different circumstances and to investigate what factors may lead to a high probability of ANA, such as whether the levels of one attribute are two times or more times the levels of other attributes, whether the number of choice sets will have significant impacts on ANA results, and so forth.

| CONCLUSIONS
This study examined Chinese public preferences for COVID-19 vaccine attributes and their WTP using DCE by accounting for ANA. Effectiveness was the most important attribute, followed by adverse reactions and protection period. The RPL-EC model con-

AUTHOR CONTRIBUTIONS
Jianhong Xiao contributed to the overall research design and paper writing; Fei Wang analysed and interpreted the data, and contributed to paper writing; Min Wang performed data collection and analysed the data; and Zegang Ma contributed to questionnaire design and writing of the paper.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.