The DCE, which is based on a random utility theory, is a quantitative method that can be used to measure respondents’ preferences, and it has been widely used in economics, marketing and psychology [27]. In the DCE, respondents were asked to choose their preferred option from hypothetical alternatives that contained one or more combinations of attributes and levels [28, 29]. In this study, the DCE was designed and analysed following the user guide published by the WHO [30] as well as the checklist published by the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) [31].
Attributes and levels
Developing attributes and levels is a key step in DCE [32, 33]. Both a literature review and qualitative studies were conducted to ensure that the attributes and levels included are the most meaningful for respondents. Firstly, we identify 9 attributes from a literature review, including monthly income, work environment, work location, bianzhi (which refers to the established posts and can be loosely regarded as state administrative staffing), workload, training opportunity, career development opportunity, management style, and welfare [22-26]. Next, we consulted 3 experts and based on their opinion, the bianzhi, workload, welfare, and career development opportunity were removed; on the other hand, years to the promotion was suggested to be added and the monthly income level was suggested to be adjusted. Finally, one additional attribute the work unit (which refers to the nature of your workplace) was added and the work environment attribute was removed after 6 in-depth personal interviews and a focus group discussion (with 7 people) from Shandong University. The final 6 attributes and levels are shown in Table 1.
DCE design
Of the 6 attributes, 2 are two-level, 3 are three-level, and 1 is four-level, a full factorial design will produce 432 (=22×33×41) hypothetical scenarios and 93,096(=(432×431)/2) pairwise choice tasks. A D-efficient design was used to generate a manageable 24 choice sets using the Ngene DCE design software [34]. To reduce the response burden of respondents, the 24 choice sets were further divided into two blocks. An opt-out was included in the second-stage question after each DCE task to allow for unconditional choices [35]. In the first step, respondents were asked to choose the job they preferred from two hypothetical jobs, and in the second step, they were asked to answer whether they would take the job if the job appeared in the real life. An example of a DCE choice set is shown in Table 2. To check for internal consistency, one choice set was duplicated. Hence, in total every respondent answered 13 DCE questions. For those respondents who failed the consistency test, they were excluded from the main analyses following the previous literature[36, 37].
Sampling
The final year undergraduate pharmacy students were chosen as the targeting respondents in this study given it is highly relevant for these students that they would be on the job market soon. A multistage cluster sampling design was used. Firstly, according to the level of economic development and geographical location, we selected six provinces: eastern (Hebei, Shandong and Jiangsu), middle (Henan) and western (Shaanxi and Ningxia). Second, a representative university offering pharmaceutical courses was selected in each province. Next, 6 universities were surveyed, including 3 comprehensive universities and 3 medical universities, namely Shandong University, Henan University, Xi’an Jiaotong University, Hebei Medical University, China Pharmaceutical University and Ningxia Medical University. Finally, 1 to 2 graduation classes of students majoring in pharmacy were randomly selected from each school. The sampling map is shown in Fig. 1.
The sample size for DCE is not straightforward, which depends on many factors such as the number of attributes and levels, the number of choice tasks, and the accuracy of expected results [38]. A rule of thumb which based on the number of attributes and attribute levels is commonly used to estimate the sample size[39]. Accordingly, it was calculated that the sample size required for this study should be more than 83 respondents (500*4/2*12=83). Combined with de Bekker-Grob's sample size requirement[40] (i.e. how to calculate the sample size for healthcare-related DCE studies), we decided that the total sample size should be not less than 400 respondents; in addition, there should be more than 100 respondents in the eastern, middle and western China, respectively.
Data collection
Before the formal survey, we conducted a pilot survey among final year undergraduate pharmacy students at Shandong University. In the pilot survey, we tested whether the attributes and corresponding levels were reasonable and easy to understand. Minor adjustments were made based on the pilot. Finally, a face-to-face anonymous survey was conducted from April to July 2017.
The questionnaire consists of two parts: section one contains personal background and section two is the DCE. Prior to the survey, we obtained the oral informed consent from all respondents. This study was approved by the Ethics Review Board of the School of Preventive Medicine, Shandong University (Reference No. 20170301).
Data analysis
The DCE data were analysed using a conditional logit (CL) model (which assumes a homogeneous preference among respondents) and a mixed logit (MIXL) model (which allows for potential preference heterogeneity). Based on the random utility framework, the utility function can be expressed as:
Where Unjt refers to the utility obtained by the respondents n by choosing the alternatives j in the choice scenario t. Vnjt and the unobservable component εnjt. The observable component is equal to the attributes levels vector Xnjt multiply the coefficient vector β'n and the unobservable component is a random error term. Except for the monthly income which was treated as a continuous variable for the calculation of willingness to pay (WTP), all other attributes were coded as dummy variables [41]. WTP is calculated by , where βm is the monthly income coefficient and βq is the coefficient for attributes levels q [42]. In this context, the WTP shows the relative monetary value that pharmacy students place on different job characteristics and it will facilitate our understanding of the relative importance of non-monetary attributes in the DCE.
In MIXL, coefficients of attributes levels are usually assumed to follow a normal distribution (described based on a mean coefficient and a standard deviation). The mean coefficients reflect the relative preference weights and standard deviation reflects the extent of preference heterogeneity[43, 44]. The choice between CL and MIXL were guided by model fit statistics, including log-likelihood ratio tests, Akaike information criterion (AIC) and Bayesian information criterion (BIC). The distance between the best and worst preference weights within each attribute can be used to compare the relative importance of different attributes to the respondents. Sub-group analyses were also conducted. After estimating the regression coefficients, a series of policy simulations were conducted to predict the probability of job choices given the changes in job properties, the results of which would be of interests to policymakers. Descriptive statistics were also presented. All statistical analyses were conducted using Stata software version 15.1.