Factors influencing job preferences of health workers providing obstetric care: Results from discrete choice experiments in Malawi, Mozambique and Tanzania

Background: Task shifting from established health professionals to mid-level providers (MLPs) (professionals who undergo shorter training in specific procedures) is one key strategy for reducing maternal and neonatal deaths. This has resulted in a growth in cadre types providing obstetric care in low and middle-income countries. Little is known about the relative importance of the different factors in determining motivation and retention amongst these cadres. Methods: This paper presents findings from large sample (1972 respondents) discrete choice experiments to examine the employment preferences of obstetric care workers across three east African countries. Results: The strongest predictors of job choice were access to continuing professional development and the presence of functioning human resources management (transparent, accountable and consistent systems for staff support, supervision and appraisal). Consistent with similar works we find pay and allowances significantly positively related to utility, but financial rewards are not as fundamental a factor underlying employment preferences as many may have previously believed. Location (urban vs rural) had the smallest average effect on utility for job choice in all three countries. Conclusions: These findings are important in the context where efforts to address the human resources crisis have focused primarily on increasing salaries and incentives, as well as providing allowances to work in rural areas.

The study was approved by the Institutional Review Board of Columbia University, New York; Global Health Ethics Committee Trinity College, Dublin; and the Institutional review boards of College of Medicine, Malawi, Eduardo Mondlane University, Mozambique and Ifakara Health Institute, Tanzania. This paper explores health workers’ preferences for job attributes using a discrete choice experiment (DCE). The DCE method has its foundations in probabilistic choice and random utility theory [12]. It enables observation of individuals’ preferences in situations where either the market does not exist (e.g. for new goods and services), is imperfect (e.g. public goods, such as parks); or when there is insufficient variation between attributes to permit accurate estimation of demand functions (as is the case for employment preferences). It is assumed that when faced with alternatives an individual will choose that which yields the greatest utility. The true utility an individual derives from an alternative is not directly observable, but is assumed to be composed of utility associated with constituent attributes that can be observed [21]. The individual is assumed to be rational and consistent in his/her choices. In human resources applications, DCEs are used to describe hypothetical job alternatives (or choice scenarios) presented to respondents who are requested to choose one. Each respondent evaluates a series of choice scenarios carefully designed in order to have some desirable statistical properties [24]. The multiple choices made by each respondent permit measurement of the relative importance of the job attributes upon which health workers make their choices. DCEs, therefore, provide valuable evidence to inform policies to attract and retain human resources for health since they enable observation of what influences health workers’ employment decisions. The design of DCEs involves different stages, from the selection of attributes and attribute levels to the construction of choice scenarios [43]. The aim is to construct hypothetical scenarios that are meaningful and important to the respondents, without resulting in heavy cognitive burden, whilst at the same time being statistically efficient [4, 25]. The first step in the process is the selection of a valid and comprehensive set of attributes and attribute levels related to the choices being analysed. All possible combinations of attribute and levels are enabled through a factorial design, and a fraction of possible combinations are selected to be included in the choice surveys (this is known as a fractional factorial design). The literature on experimental design for DCE is large and continuously evolving; with contributions coming from diverse fields such as environmental economics, marketing, and transportation economics [24, 47]. The selection of attributes for this study was based on previous research (qualitative interviews) conducted with mid-level cadres in Malawi [30] which showed that how people were treated by their managers, their involvement in decision making and opportunities for development and advancement (all elements of human resource management and professional development) were amongst the strongest predictors of job satisfaction. A strong correlation between management support and intention to leave the job [29] suggested that good human resources management might be an important consideration in job choice. Housing, pay, urban location and availability of resources and equipment required for the job were the most commonly reported attributes of importance in previous studies conducted with similar populations (e.g. [13, 28]). Table 2 below presents the set of attributes and their respective levels; a detailed description of each attribute is presented in Appendix 1. Also presented is the variable coding scheme used for statistical analysis, discussed below in the section on model fitting. Attributes and attribute levels for job alternatives – three countries All possible combinations of attributes and attribute levels (i.e. a full factorial design) would result in 144 possible scenarios or job descriptions (24 × 32, i.e. four attributes with two levels and two attributes with three levels). In order to have a manageable number of scenarios, a fractional rather than full factorial design was used. A set of choices was selected to allow the main effects (the effect of each independent variable on the dependant variable) to be explored. A constant comparator method was used i.e. holding one job specification constant while changing the levels of the attributes in the second job specification. In total 15 choice sets were presented. In DCE applications in the health research arena there has been a move towards the use of optimal designs and the use of SPEED software to generate orthogonal designs. A recent review of DCE designs [3] identified fractional factorial designs as the most commonly used for DCE applications. In addition they found the mean number of attributes to be 5 and the mean number of choice sets to be 14. We did not include an opt-out option in the design. The rationale for employing a forced choice is that although an opt-out option can reduce biases in parameter estimates, it cannot provide sufficient information on respondents’ preferences for the attributes if too many respondents choose the opt-out option [40]. Field staff received a one-week training on all steps in the data collection process. This included a trip to the field to pilot test the instrument on a small sample of health workers. Although the survey was designed to be self-administrated, field staff were required to remain in the facility during the data collection period to explain the contents of the survey and answer any questions that staff might have. Ensuring a common understanding of the attributes and levels and the provision of standard explanations across all sites was emphasised to fieldworkers during training. The descriptions of attributes and attribute levels are contained in Appendix 1. This was included in the survey instrument and respondents were instructed to read and make sure they understood these before completing the questionnaire. The primary target for the DCE was health care workers who had performed at least one of the EmOC signal functions in the previous three months; thus the focus was on maternity staff, as well as health care workers who provide surgical services, such as caesarean section. Since it was not possible to randomly sample healthcare workers themselves, guided by existing staffing levels, the project randomly sampled hospitals and health centres to be visited to approach the minimum target of 500 health care workers per country for the provider survey. Hospitals were intentionally oversampled because the majority of EmOC is provided in hospitals rather than health centres. In Malawi, a near-national sample of facilities (N = 84) intended to provide EmOC services was identified and included central, district, rural and CHAM (faith-based organisations) -operated hospitals and a randomly sampled urban and recently upgraded health centre designated to provide EmOC. A few districts/facilities were excluded in Malawi due to their recent participation in another human resources study in which similar data had been collected from health workers. In Tanzania, due to the size of the country, cluster sampling was employed. One region was randomly selected in each of the eight geographic zones and all districts within those eight regions were then included in the sampling frame. The primary hospital serving the district was identified for inclusion; either the government-run district hospital or voluntary agency-run (VA) designated district hospital (DDH). In some districts that also contain the regional headquarters, the regional hospital was included in the sample when there was no district hospital serving the community. One health centre (HC) was randomly selected in each district, thus there were two facilities from each district in the study (N = 90). In Mozambique, a near national sample of general, district and rural hospitals was included to maximise the potential participation of the NPC cadre tecnico de cirurgia. In addition, two to three health centres (type 1 and type 2) providing maternity care, and therefore at least some basic EmOC functions, were randomly selected in each district for inclusion in the study (N = 138). Data collection was conducted in the three countries during 2008–09. In each of the selected facilities staff were deemed eligible to participate if they were present during the study visit and reported having provided at least 1 of the 9 EmOC signal functions within the previous 3 months, and had granted informed consent. There are 7 signal functions for basic EmOC (parenteral antibiotics, parenteral utertonics, parenteral anti-hypertensives, removal of retained products, manual removal of placenta, assisted vaginal delivery, neonatal resuscitation), and 9 signal functions for comprehensive EmOC (the basic 7, plus caesarean delivery and blood transfusion). The questionnaire was self-adminstered in the English language. Details of the data collection procedure are provided in Appendix 2. Each respondent was asked to evaluate 15 choice sets and chose one job description; each choice set containing two job descriptions (see Fig. 1 contains an example of choice set). Besides the choice experiment the questionnaire also included demographic data. Example of a discrete choice experiment question (choice set) Discrete choice models are Random Utility Models (RUMs) that are widely used for the analysis of discrete choice experiments. Three underlying assumptions of discrete choice models are that (i) choice is discrete (individuals either choose a particular alternative or not), (ii) the utility for an alternative is a random variable that varies over individuals and (iii) in a choice situation, individuals choose the alternative for which their utility is maximized. The aim of discrete choice models is to estimate the probability of an individual choosing one alternative over the other alternatives presented in the choice scenario [15, 25]. Individuals choose goods and services that yield the highest utility (or satisfaction). Therefore, the choice between alternatives in a choice experiment is based on the combination of attributes and attribute levels that results in an increase in utility for the respondents ([27, 38]). The task is then to estimate parameters that determine the relative importance of different attributes affecting the choice process. Conditional (or multinomial) logit models are discrete choice models that have been utilized in many fields of research, from marketing to medicine. In recent times however, these models have been superseded somewhat by the more flexible mixed logit model. The mixed logit model has become popular following the development of simulation methods that enable it to be estimated more readily, and following the integration of these methods into popular software tools [14]. The mixed logit is a highly flexible discrete choice model that can approximate any random utility model [31]. Hensher and Greene [14] and Train [46] describe this model in detail. A more detailed description of the model and its parameter estimation is contained in Appendix 3. Mixed logit models were fitted to the discrete choice data from each country to estimate job preferences. All choice scenarios presented to individuals contained two unlabelled alternatives (two job descriptions). Each job was described by six attributes, four of which had two levels (location, equipment, professional development and HRM) and two of which had three levels (pay and housing). All job attributes were represented as categorical measures (Table 2) and therefore were coded as dummy variables for statistical analysis. Attributes were coded for analysis as binary (dummy) variables. Each two-level attribute was represented by a single binary variable, while each three-level attribute (pay and housing) was represented by two binary variables. Table 2 shows the attribute coding system used in the analysis. Pay was included as a categorical, rather than a continuous predictor, to allow for the possibility of a non-linear effect of pay on utility. It was considered likely that the added utility of 1.5 × base over base pay, was not the same as the added utility of 2 × base over 1.5 × base. Binary mixed logit models were fitted to estimate the probability of an individual choosing a given alternative (job 2) over the other (job 1). Normally distributed random coefficients were specified for each of the eight attribute variables. It is possible that an individual’s utility for particular job attributes may differ depending on observed characteristics of that individual. For example, it could be possible that older individuals place a higher value on superior quality of housing, or that females have a stronger preference for jobs with improved availability of continuing professional development. To allow effects such as these to be captured, we tested for fixed effect interactions between each alternative-specific attribute (Table 2) and each of the individual-specific characteristics listed in Table 3. Sample demographics for each country *baseline category Note that the first four individual-specific characteristics in Table 3 are categorical, while the fifth is a numeric variable. The baseline category for each categorical variable is marked in the table (*) and the variable is therefore represented by the inclusion of dummy variables for the other categories. Health workers were grouped into basic, mid and high level cadres within each country, as defined in Table 4. Note that health workers in Malawi were grouped into mid and high level cadres only and therefore only a single dummy variable was required for cadre (the baseline category is mid-level cadre, while a dummy was included for high level cadre). Grouping of cadres for statistical analysis Fitting a mixed logit model with eight random coefficients is highly computationally intensive. It was therefore infeasible here to perform variable selection on all fixed effect interaction terms under the specified mixed logit model. Instead, bootstrap variable selection was carried out using conditional logit models (assuming that all coefficients were fixed). For each of the three country datasets, 200 bootstrap samples were drawn from the data and a forward greedy search algorithm was carried out to select the fixed effect interaction terms that should be included. The Bayesian Information Criterion (BIC) proposed by Schwarz [41] was used to decide whether covariates should be added or removed from the model. For each bootstrap sample, the greedy search algorithm proceeded as follows: A similar variable selection strategy to the above was used in Raftery and Dean [34] and in Galligan et al. [10] to select variables for inclusion in clustering and classification models respectively. Results across the 200 bootstrap samples were compiled. Fixed effect interaction terms that were chosen in 50% or more of the bootstrap samples were considered to be important, and hence were selected for inclusion in the mixed logit model for that country. Mixed logit models were fitted with varying numbers of Halton draws [45], starting at 500 draws and increasing the number of draws by 500 until convergence of the parameter estimates was reached. A large number of draws was required for each dataset, attributable to the eight random coefficients for which distributional parameters must be estimated. Likelihood ratio tests were carried out to test for the inclusion of correlated (vs. independent) random effects. In all cases, likelihood ratio tests provided evidence that correlated random coefficients improved model fit (Table 5) and therefore correlated random coefficients have been included in all mixed logit models presented below. Likelihood ratio tests comparing models fitted with uncorrelated, and correlated, random coefficients Conditional logit models were fitted in the mlogit package in R (R: A language and environment for statistical computing). Mixed logit models were fitted here using the mixlogit command [16] in Stata Version 12.1.

N/A

• Dima Health