Menu
Menu
Menu
Menu
Attracting, training and retaining high-quality health workers are critical for a health system to function well, and it is important to know what health workers value in their roles. Many studies eliciting the labour market preferences of health workers have interviewed doctors or medical students, and there has been little research on the job preferences of lower-skilled cadres such as community health workers, mid-skilled clinical care staff such as nurses and midwives, or non-patient facing staff who manage health facilities. This study estimated the job preferences of public health sector community health extension workers (HEWs), care providers including nurses and midwives, and non-patient-facing administrative and managerial staff in Ethiopia. We used a discrete choice experiment to estimate which aspects of a job are most influential to health worker choices. A multinomial logistic regression model estimated the importance of six attributes to respondents: salary, training, workload, facility quality, management and opportunities to improve patient outcomes. We found that non-financial factors were important to respondents from all three cadres: e.g., supportive management [odds ratio (OR) = 2.96, P-value = 0.001] was the only attribute that influenced the job choices of non-patient-facing administrative and managerial staff. Training opportunities (OR = 3.45, P-value < 0.001), supportive management (OR = 3.26, P-value < 0.001) and good facility quality (OR = 2.42, P-value < 0.001) were valued the most amongst HEWs. Similarly, supportive management (OR = 3.22, P-value < 0.001), good facility quality (OR = 2.69, P-value < 0.001) and training opportunities (OR = 2.67, P-value < 0.001) influenced the job choices of care providers the most. Earning an average salary also influenced the jobs choices of HEWs (OR = 1.43, P-value = 0.02) and care providers (OR = 2.00, P-value < 0.001), which shows that a combination of financial and non-financial incentives should be considered to motivate health workers in Ethiopia.
The study was conducted in Ethiopia, which is divided into nine regions and two city administrations. In each region, woredas (districts) are administrative units, managed by decentralized councils of elected members (Workie and Ramana, 2013). The Ethiopian healthcare delivery system, referred to as the three-tiered system, provides healthcare services to people at primary, secondary and tertiary levels (Alebachew and Waddington, 2015). The primary level, where this study operated, consists of three service delivery points: health post, health centre and primary hospital (Figure 1). Ethiopia health system (reproduced with permission from Alebachew and Waddington, 2015) The primary healthcare workforce includes facility- and community-based health workers supported by non-patient-facing management and administrative staff (World Health Organization, 2020). HEWs are assigned to health posts as salaried government employees following a 12-month training programme (Assefa et al., 2019). They are usually hired as Level 3 health workers and have the opportunity to upskill and be redeployed to higher positions in the health system after taking a competitive exam. The average attrition rate of HEWs is around 3% per year with some evidence suggesting a continuing rise since the start of the Health Extension Programme (HEP) (Arora et al., 2020). Evidence suggests that around 70% of HEWs have a desire to upgrade as a nurse, although to what extent that is possible is not clear (Teklehaimanot et al., 2007). Yet, factors affecting the retention of HEWs are largely due to non-material factors, such as community acceptance or acknowledgement from supervisors and senior managers (Arora et al., 2020). In contrast, factors affecting the retention of public sector nurses and midwives are a mix of financial and material incentives (e.g. better remuneration and improved infrastructure), whereas one recent study in Ethiopia found that around 50% of nurses and midwives intended to leave their current job in the following year (Ayalew et al., 2015; Muluneh et al., 2021). Some evidence also suggests that access to a large labour market with competing salaries and infrastructure quality (e.g. in non-governmental organizations, the private sector and international labour market) was also another reason for the high turnover of government-employed nurses and midwives in Ethiopia (Mariam, 2013; Ayalew et al., 2015; Muluneh et al., 2021). To our knowledge, there is no published evidence of retention among non-patient-facing staff or the factors influencing retention of non-patient-facing staff in public sector health facilities, despite their essential role in overseeing the functioning of the healthcare delivery system. The DCE was embedded within a baseline data collection of a survey conducted as part of the process evaluation of a quality improvement (QI) programme implemented by the Institute of Healthcare Improvement and the Ethiopian Federal Ministry of Health. At the time of data collection, no participants had been exposed to the QI programme. Data were sampled from four out of the nine Ethiopian regions for this study. Using a random number generator, we randomly selected one QI programme woreda per region from Oromia, Amhara, Southern Nations, Nationalities, and Peoples’ Region (SNNPR) and Tigray. We added one additional randomly selected woreda in Amhara since the first randomly selected woreda in Amhara had too few health facilities to reach the sample size. We further purposively sampled two additional woredas from Oromia and SNNPR (Bunno Bedelle and Chencha, respectively) where other evaluative work was also taking place. For each of the seven QI programme woreda chosen for data collection, we chose one matched woreda from the same region which was not subject to QI activities, resulting in 14 woredas in total. The woredas were matched using service utilization data from the last three Ethiopia Demographic and Health Surveys (2005; 2011; 2016). In each woreda, we sought to interview 30 participants across a range of health worker and management cadres, where the latter included facility heads alongside woreda and regional health office managers. Senior non-patient-facing staffs in each woreda were not randomly sampled due to their small number, but staff at primary hospitals, health centres and health posts were randomly sampled. The heads or clinical directors of each woreda (one), primary hospital (one) and health centre (three) were interviewed. Four maternal and child health clinical care providers and two from each health centre were interviewed in the hospital. One HEW was interviewed from each health post under each health centre. The baseline survey was conducted from April 15 to May 10, 2018. We obtained a stratified random sample of 401 workers in the Ethiopian health system across three cadres: 202 (50.4%) HEWs, 155 (38.7%) care providers (including 100 midwives) and 43 (10.7%) non-patient-facing administrative and managerial staff. A team of seven trained research assistants from the authors’ institute implemented a face-to-face survey administered in English, Amharic and Oromifa languages using Open Data Kit (https://opendatakit.org) software on tablet computers. Informed consent was obtained from all participants before data were collected. The attributes and findings of published DCEs conducted in east Africa were analysed to inform the development of our DCE (Mangham and Hanson, 2008; Blaauw et al., 2010; Kolstad, 2011; Rockers et al., 2012; Mandeville et al., 2014; 2016). A shortlist of potential attributes was generated and reduced to six using the findings of a qualitative study conducted 1 year previously, assessing the motivation of HEWs in Ethiopia (Tesfaye, 2017). As there is some debate on the use of text or images to represent attributes and levels in DCEs, we opted to display choice tasks as text since pictures can convey their own meanings, sometimes different from the text, which can cause confusion (Veldwijk et al., 2015). We displayed two job profiles in each choice task using an unlabelled design where each alternative represents a generic health worker’s job, within which all selected characteristics change as prescribed by the statistical design. Participants were asked the following question: ‘Here are two jobs described by some of their characteristics. Compared to your current job, please choose which you would prefer’. To increase realism and allow for the estimation of unconditional demand, a generic opt-out alternative was included to represent their current job. The final six attributes of the DCE and their levels are shown in Table 1, and Figure 2 shows an example of how choice tasks are presented to respondents. The final design incorporated seven choice tasks. Choice experiment attributes and levels Example of choice task presented to study participants The DCE was piloted among 19 woreda health office staff in December 2017. No changes were made to the DCE between piloting and the final survey as it was understood well by participants. Priors from analysis of pilot data (n = 19) were used in NGENE software (http://www.choice-metrics.com/) to generate a single D-error minimizing design with 10 tasks, which avoided dominant or duplicated alternatives with the aim of improving precision in the final model estimates. Choices were modelled based on McFadden’s random utility theory (McFadden, 1973). This assumes that respondent will choose alternative in choice set if that alternative provides the most satisfaction out of all other alternatives. This is shown in the following Equation (1): where is the utility function of individual from choosing alternative in choice set ; signifies the observable element for choosing alternative and represents the random, unobservable element for choosing alternative . Equation (2) represents the ‘indirect utility function’ of Equation (1). where represents a linear specification of the DCE attributes, as shown in Equation (3). The probability of choosing alternative is captured by a set of observable attributes, , which takes the following form: where represents the constant, and salary, impact, management, facility, training and workload were the attributes used in the DCE. This is underpinned by Lancaster’s consumer behaviour theory, which assumes that utility is derived from the characteristics of a certain good (Cascetta, 2009; Lagarde and Blaauw, 2009; Mandeville et al., 2014; Lancsar et al., 2017). Using specifications from Equation (3), Equation (1) was estimated using a multinomial logit (MNL) model, which generally assumes that the stochastic term, , is independently and identically distributed (IID). The IID assumption assumes that unobserved effects are not related in any systematic way with the observed effects and in practice assumes preference homogeneity across individuals (Hensher et al., 2005). Standard errors were clustered at the facility level, relaxing the IID assumption by allowing for intra-group correlation. This meant that observations from the same facility were not independent, but observations remained independent across groups (Lancsar et al., 2017; StataCorp, 2019). Preference heterogeneity was explored by cadre. We conducted a subgroup analysis of the main effects by running separate regressions on three sub-groups of health workers to reveal any variation in preferences. This included HEWs, care providers such as nurses and midwives, and non-patient-facing administrative and managerial staff. The same attributes were used across the three cadres to allow us to make comparisons between each cadre’s trade-offs. Individual characteristics were not adjusted for in the model due to the small sample size of some of the sub-groups and the presence of multicollinearity. Stata 15 was used to estimate the MNL models, and odds ratios (ORs) used to estimate the relative importance of each attribute; the attributes were dummy coded and standard errors were clustered at the facility level. Utility was estimated as a measure of choice. ORs that are larger than one indicate positive impact on utility whereas those below one indicate disutility attached to the attribute level. Face validity was assessed by checking if ORs were of the expected sign (de Bekker‐Grob et al., 2012). Additional robustness checks were carried out to check if the results changed by adjusting the standard errors at the individual level or removing the cluster adjustment altogether. A goodness-of-fit model was estimated using log pseudolikelihood and pseudo R-squared (Hauber et al., 2016). Only ORs that are statistically significant at either the 5% or 1% level are reported.
N/A
