Job preferences for social medicine and health care management 1 PhD students in China: A discrete choice experiment 2

Background : The shortage of public health workforce and its uneven distribution between 2 developed and underdeveloped areas still exist in China’s health system. This study aimed to 3 investigate PhD students majoring in social medicine and health care management (SMHCM)’s 4 stated preferences when choosing a job, so as to inform policy-makers regarding alternative 5 interventions to address recruitment and retention problems at underdeveloped areas. 6 Methods: The relative importance of 7 job attributes was assessed by conducting a web-based 7 discrete choice experiment (DCE) survey on a sample of 193 SMHCM PhD students at different 8 grades during October 20 and November 12, 2020. The 7 attributes were monthly income, location, 9 housing, children’s education, working environment, career promotion and Bianzhi (which refers to 10 the authorized number of personnel (the number of established posts) in a party or government 11 administrative organ). 12 Results: All 7 attributes were statistically significant with the expected sign and demonstrated the 13 existence of preference heterogeneity. Monthly income, location and Bianzhi were of most concern 14 for SMHCM PhD students when deciding their future jobs. Among the presented attributes, working 15 environment was of least concern. For the sub-group analysis, the job in first-tier city was more 16 likely to lead to a higher utility value for PhD students who were women, married, coming from 17 urban area and has high annual family income. Compared with female PhD students, male PhD 18 students were willing to pay more for a shorter time to get promoted. 19 Conclusions: Both monetary and non-monetary attributes were found to be significantly 20 influential in affecting PhD students’ preferences for choosing a job. Wider use of choice 21 experiments can help improve the recruitment and retention of health workforce at disease control system, especially in third-tier cities where resource constraints preclude the use of all strategies.


Introduction
should appear at here 11

12
Once the attributes and attribute levels were defined, they were then combined into a set of carefully 13 selected scenarios (choice sets). We followed standard approaches for the design of the choice 14 experiment in order to achieve unbiased, statistical response efficiency [17] . The DCE was based on 15 7 attributes. 3 of the 7 attributes were described in choice tasks by 3 response levels, three attributes 16 by 2 levels and one attribute by 4 levels, yielding a total of (e.g.,3×3×3×2×2×2×4) 864 potential 17 combinations. Because the total number of comparisons ([864×863]/2) cannot be feasibly evaluated, 18 we developed an orthogonal and balanced designs to identify similar and efficient sets of profiles 19 among these profiles [18,19] . Applying the methods introduced by Huber and Zwerina [20] , DCE 20 macros for SAS (version 9.4) were used to select combinations for an orthogonal main effect design, 21 and the selected profiles were organized into the most D-efficient choice designs, given the design 22 parameters (Relative D-Eff: 77.9%). The design was evaluated based on the variance of parameters, 1 confirming that the design was orthogonal. Finally, 36 choice sets were identified and were further 2 divided into three blocks so as to minimize participants' cognitive burden. Within each version, a 3 single choice set was duplicated to examine the internal consistency of participants. All participants 4 were randomized to receive one of the 3 versions of the DCE questionnaire according to their month 5 of birth. (Block 1: January to April; Block 2: May to August; Block 3: September to December). 6 An example of the DCE choice set is provided in Table 2. 7 Table 2 should appear at here 8 Data collection 9 In addition to the DCE questions, the online questionnaire also contains questions related to students' 10 sociodemographic characteristics, career planning, and family income. A ranking exercise was 11 conducted prior to the DCE tasks to further examine the internal predictive validity of the DCE 12 estimates, in which participants were asked to rank three attributes (within 7 attributes) from most 13 important to least important with respect to their job preferences. At the end of the questionnaire, 14 the respondents were given a task to indicate on a 5-point scale the level of difficulty of completing 15 the 13 DCE choice tasks. The full questionnaire was piloted among SMHCM PhD students at Fudan 16 University and Shandong university before data collection conducted between July and October of 17 year 2020, aiming to examine the comprehensibility, acceptability, and validity of the questionnaire. 18 A brief explanatory statement was given to respondents, which described the study and highlighted 19 that their participation was voluntary and that no identifiable personal data would be collected. It is 20 an anonymous survey so we did not require a written consent. A return questionnaire also indicates 21 the implied consent which is commonly adopted in the anonymous online survey. Ethical approval 22 (Reference No.2020-10-0853) was obtained both for the consent procedure and for the study as a 1 whole from the ethics review board of the school of public health, Fudan University, and the 2 research adhered to the tenets of the Declaration of Helsinki. 3 Data analysis 4 We used STATA version 15.1 to clean and analyze the data. Descriptive statistics were reported for 5 participants' socio-demographic characteristics, the ranking results and the 5-point scale score. The 6 data from the DCE were analyzed within a random utility theory framework. The utility (U) 7 associated with a particular job is made up of two components: the deterministic component 8 (where is a function of observable characteristics) and the unobservable component . The model 9 of utility for an individual n associated with job i can be estimated as: 10 Second tier city + 2 First tier city + 12 3 ℎ Allowance + 4 ℎ Provided + 13 5 ℎ ′ Good + 6 3 year + 14 7 1 year + 8 1 year + 15 9 ℎ Provided + 10 ℎ + n 16 Two econometric approaches were used to estimate this utility function, including the classical 17 conditional logit model and a mixed logit model that could be used to capture potential unobservable 18 preference heterogeneity [21] . Conditional and mixed logit regression models were compared using 19 the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), which are 20 commonly used for model selection in random utility framework [22,23] . 21 The DCE data were binary, where '1' indicates that the alternative plan was chosen and '0' means 22 that the other alternative plan was chosen. Although most previous studies specify the coefficient 23 for monetary attribute in choice models to be fixed, it is often unrealistic to assume that all 1 participants have the same preferences regarding the monthly income of a job position [24] . In our 2 study, all attributes were dummy coded and specified as having a random component, except for 3 monthly income which was specified as a continuous variable in the models to facilitate the 4 calculation of willingness to pay (WTP), that is, the relative monetary value that students place on 5 various aspect of the job options. Through calculating the ratios of the coefficients between each 6 attribute level and the salary attribute, the marginal rate of substitution or WTP was calculated 7 where 10 is the salary coefficient and (1,2···9) is the coefficient for attribute level 1, 8 2···9). The positive and negative results indicate theoretically to what extent the participants would 9 be willing to pay/to be compensated for an attribute level. Finally, we also conducted a simulation 10 study to understand to what extent the probability of choosing a given post changes as the levels of 11 the attributes are changed. 12

13
A total of 193 individuals participated in the online survey, among which 26 were excluded because 14 their university does not have a SMHCM major, so we take them as invalid data. For internal 15 consistency, a choice test based on duplicated choice tasks among the remaining 167 participants 16 resulted in 14 (8.4%) participants failing the test. The detailed results reported below were based on 17 153 participants who passed the internal consistency tests. Meanwhile, a sensitivity analysis 18 including all 167 participants (Additional file 1: Table S1) was conducted and the results are 19 comparable to the results reported below. 20

21
The sociodemographic characteristics are shown in Table 3. There were no significant differences 22 between those who failed versus who passed the internal consistency tests. The analysis sample 1 (n=153) had a mean age of 28.8 years (SD=4.5). Most of them were female (62.1%), coming from 2 urban areas (65.4%) and single (69.9%). Around 79.1% of the PhD students would do major-related 3 job after graduation while 18.9% hasn't make up their mind. For the '5-point scale' question, 61 4 respondents (39.9%) thought it was easy or very easy to complete the 13 DCE questions, 66 5 respondents (43.1%) thought it was normal, only 26 respondents (17.0%) thought it was difficult or 6 very difficult. 7 Table 3 should appear at here 8

9
The AIC and BIC values suggested that the mixed logit estimates were preferable to the conditional 10 logit estimates for the analysis sample and the results from mixed logit model were not substantially 11 different from the conditional logit model. As such, we only report the preferred mixed logit 12 estimates in Table 4. The conditional logit estimates are presented in Additional file 2: Table S2. 13 A total of 153 respondents completed the DCE for a total of 3672 observations. The statistical 14 significance of all levels of each attribute indicates that all key characteristics identified in the DCE 15 design stage played a significant role in job choice. Results from the mixed logit model show that 16 respondents held strong preferences for first-tier city as opposed to third-tier city (β or relative utility 17 = 1.687; p < 0.001). SMHCM PhD students also exhibited strong preferences for provide housing 18 compared with no housing provide (β = 1.129; p < 0.001), as well as has Bianzhi compared with no 19 Bianzhi (β = 1.045; p < 0.001). SMHCM PhD students expressed a preference for 1 year to get 20 promoted (β = 0.719; p < 0.001), as well as good children education (β =0.555; p < 0.001). Good 21 working environment only signs of a mild preference (β = 0.401; p < 0.001).  The willingness to pay analysis, carried out in order to calibrate the strength of SMHCM PhD 5 student's preferences to a single standard, quantified how much salary they were willing to sacrifice 6 in order to obtain a desired level of an attribute. This analysis revealed that students were willing to 7 pay 11767.8 CNY to attend a job in the first-tier city rather than the third-tier city. Students were 8 willing to pay 7873.5 CNY for the house rather than no house provide. In terms of Bianzhi, students 9 were willing to pay 7287.1 CNY to get a job with Bianzhi. The results of selective sub-group 10 analyses were presented in Tables 5-8 and Figure 2. For the subgroups, all seven attributes remained 11 statistically significant in influencing preferences. Focusing on the WTP estimates, it can be seen 12 that the job in first-tier city was more likely to lead to a higher utility value for respondents who 13 were women, married, coming from urban area and has high annual family income. In addition, 14 compared with female students, the male students were willing to pay more for a job with 1 year to 15 get promoted. The simulation results are shown in Figure 3. The initial (baseline: 10000 CNY monthly income; 20 no housing; normal children's education; get career promotion after 5 year; no Bianzhi) probability 21 of taking a job in the third-tier city is 15.6%, hence the probability of taking job in the first-tier city 1 is 84.4%, the job in the first-tier city is thus preferred. For the single incentives, only increasing 2 monthly income from 10000 to 25000 CNY can make the probability of choosing the third-tier job 3 (61.4%) exceed the job in first-tier city (38.6%). For the given multiple incentives, the policy "②+ 4 ③+⑤+⑥" was the most attractive one, as it can increase the probability of taking a third-tier city 5 job to 79.4%. We have used DCE data to quantify the preference of the final year undergraduate healthcare 9 administration students and to model the likely impact of different human resource strategies on 10 rural recruitment in China [13] . In the absence of data from rigorous evaluation studies, such analyses 11 provide useful insights into the potential effectiveness of different human resource policy 12 interventions. We hope the findings of this study will help inform policy-makers' actions to 13 encourage the recruitment and retention of SMHCM PhD students at health system (for example, 14 CDC), especially in third-tier cities of China. 15 Our study shows that most of the PhD students prefer to find a job in the university or research 16 institution, only few students plan to work at CDC. Turnover is a common phenomenon both in the 17 national CDC and local CDC in China because of the low salary [4] . Our study confirms that financial 18 incentives are very important in persuading PhD students to choose a position in third-tier cities, 19 but only if they are fairly large. In our study, a 5000 CNY salary increase was relatively ineffective 20 ( Fig. 3). Although higher salary has a large effect on preference for third-tier city positions, it is 21 often infeasible in the real world settings, for it may require significant financial investments upfront 1 which may deter policy-makers from implementing this intervention [25] . 2 Among non-monetary attributes, working in the first-tier cities is the most important factor, 3 especially for students from urban areas. The results of ranking job posting attributes according to 4 their importance in Figure 1 confirmed the importance of working location again. The simulation 5 results in our study shows that for the single incentives, only increasing monthly income from 10000 6 to 25000 CNY can make the probability of choosing the third-tier city job exceed the job in first-7 tier city. In China, living conditions in most of the third-tier city are still poor compared with first-8 tier city in terms of infrastructure (e.g. telecommunications and transportation), schools for children 9 and employment opportunities for spouse [26] . Studies from other countries have reported that the 10 more centrally located the job, the more it will be preferred by health workers or graduate students 11 [27,28] . In addition, our study also reveals that compared to students from rural areas, those from 12 urban areas shown a much stronger preference to work in the first-tier city rather than in third-tier 13 city. Therefore, attracting and retaining SMHCM PhD students with a rural background for third-14 tier city would be a more feasible strategy. 15 In teams of housing, providing housing allowance is moderately effective, but providing housing 16 is a very powerful non-financial strategy. This shows the importance of providing housing for 17 SMHCM PhD students when choosing a job. Other study also showed similar results [29] . Influenced 18 by the traditional concept, most Chinese people think that 'one does not have a home if one does 19 not own a house'. In addition, a series of welfare benefits attached to the house, such as children 20 education, pension, social services, etc., have exacerbated the current situation that 'it is difficult to 21 own a room' [30] . In recent years, although Chinese government has always adhered to the policy 22 that 'houses are used for living, not for speculation', and local governments have also implemented 1 a series of measures, such as restricting the purchase and loan of houses and increasing the supply 2 of affordable housing, the housing prices still exceed the affordability of ordinary office workers 3 [30] . Limited by the financial capacity, for the local governments of third-tier cities, the challenge is 4 not only to identify which single intervention is more likely to promote recruitment and retention of 5 SMHCM PhD students, but to identify the best combination of interventions which can be truly 6 implemented. It is often not possible in the real world to provide housing for the PhD students, 7 alternatively, a combination non-monetary combined with monetary incentives, such as housing 8 allowance combined with 15000 CNY and get career promotion after 1 year can achieve lager 9 impact than provide housing incentive only. 10 Contrary to our previous studies with heath administration [13] , nurse [14] or medical students [31] 11 which found that Bianzhi has the lowest utility in job preferences, Bianzhi is another important non-12 monetary factor that influence the SMHCM PhD students job choice in this study. In China, Bianzhi 13 refers to the authorized number of personnel (the number of established posts) in a Party or 14 government administrative organ, a service organization or a working unit, a job with Bianzhi means 15 more stable [32] . This is perhaps because the respondents in this study were older, with an average 16 of 28.8 (while the average age of undergraduate healthcare administration students in our previous 17 study was 22.2), and some of them had started a family, so a job with Bianzhi may be more important 18 for them. In addition, the PhD students in this study are more likely to work in the university or 19 scientific research institution and governmental, for those departments, there is a big difference in 20 salary, welfare, health insurance, children' education, and etc. between the job has Bianzhi or not. 21 Career promotion is another important nonmonetary factor, especially for male SMHCM PhD 22 students. Similar results have been obtained in other human resource DCE-based studies in low-1 and middle-income countries [29,33] . Snow et al [34] study indicates that the absence of senior posts 2 in underdeveloped areas is an important factor associated with the feeling of "professional 3 imprisonment" identified by those working in rural and remote posts. Developing clear career 4 paths for rural and remote areas posts and adopting strategies to increase public recognition are 5 strongly recommended strategies [35] . This is contrary to our previous study with undergraduate 6 healthcare administration students which found that the career development opportunities did not 7 appear to be as important as the other non-monetary attributes [12] . This is perhaps because the 8 SMHCM PhD student's expectation of career achievement is generally higher than the 9 undergraduates. A survey of career expectations for PhD graduates carried by China association 10 for science and technology showed that development prospects are the main considerations of a 11 PhD graduate in obtaining employments [36] . 12 The children's education attribute was found to have slightly smaller effect than the work location, 13 housing or career promotion. It seems contrary to the study conducted in Nepal [37] in which children' 14 education was found to be much stronger predictors of choice. It could be that most of the SMHCM 15 PhD students we studied had not started a family, so perhaps their future children's education was 16 not among their main concern. The subgroup analysis in our study also strengthened the above 17 assumption which the married PhD students have a stronger preference for the children's education 18 compared with the unmarried PhD students. 19 An unexpected finding from our study is the relatively lower utility of working environment in 20 job preferences. It was contrary to our previous studies which was strongly suggestive of a 21 preference for improved working environment [13] . In this study, working environment refers to 22 management support, the relationship between superior and subordinate, high-risk work 1 environments, and availability of equipment. We assumed that this may be related to the specificity 2 of our study population, because PhD students have a high expectation for their career achievement 3 and don't expect to find a job with poor working environment [36] . This finding is consistent with 4 the results of an earlier quantitative study in which working environment was not think as the major 5 contributing factors towards job choice for the PhD students in China [36] . Despite all this, we do not 6 recommend the government or policy-makers ignore the importance of working environment when 7 recruiting the SMHCM PhD students, for the reason that due to the high expectation for their future 8 career, a position with poor working environment may incur larger negative impact for their job 9

preferences. 10
The pandemic of COVID-19 highlights the importance of strengthening public health systems. 11 In the future, the demand for public health workforce in disease control system will increase. In 12 addition to our study, other studies also found that many public health graduates were unwilling to 13 devote themselves to CDC [3,38] . To address potential challenge of human resources shortage in the 14 disease control system of China, further qualitative research such as in-depth interviews and focus 15 group discussions involving SMHCM PhD students is required to determine the specific reasons 16 why they are unwilling to work at CDC. 17 There are several limitations in this study. The first limitation stems from the fact that a choice 18 experiment does not offer the universe of attributes because the choice task becomes difficult and 19 respondents are less willing to critically appraise each attribute as the list grows. Not all potentially 20 important attributes, such as workload, were assessed. Second, DCEs are based on stated 21 preferences and not on actual behaviors. The existing literature suggests that response in choice 22 experiments predicts behavior, but this association is far from perfect. As a result, though choice 1 experiments can efficiently narrow the field of candidate interventions, evidence of SMHCM PhD 2 students' behavior may still be needed in many circumstances. Finally, unlike our previous studies, 3 the respondents in this study were not limited to the final year SMHCM PhD students. Though job 4 preferences may vary between PhD students at different grades, given the limited sample size, we 5 were unable to examine this difference. 6

Conclusions 7
Although China has conducted a series of DCE-based studies on graduates, the respondents are 8 mainly undergraduate graduates [13,14,31,39] . To the best of our knowledge, this is the first study using 9 DCE methodology to investigate job preferences of PhD students internationally. The SMHCM 10 PhD students are in their early stages of career preparation, so the results of this study will be more 11 effective to inform policy-makers regarding the design of recruitment and retention policies. Both 12 monetary and non-monetary attributes were found to be significantly influential in affecting students' 13 job preferences. Wider use of choice experiments can help improve the recruitment and retention of 14 health workforce at disease control system, especially in the third-tier cities where resource 15 constraints preclude the use of all strategies. The authors thank all the participants for their time and effort. Responsibility for any remaining 6 errors lies solely with the authors. 7

Funding 8
All meetings, time invested, travel, etc. was funded by the individuals. The project had no financial 9 support. 10

Availability of data and materials 11
The data used and/or analyzed during the study are available from the corresponding author on 12 reasonable request.

Ethics approval and consent to participate 18
Ethical approval (Reference No.2020-10-0853) was obtained both for the consent procedure and for 19 the study as a whole from the ethics review board of the school of public health, Fudan University, 20 and the research adhered to the tenets of the Declaration of Helsinki. 21

Consent for publication 22
Not applicable. 23