Motivation and incentive preferences of community health officers in Ghana: An economic behavioral experiment approach

Background: Health worker shortage in rural areas is one of the biggest problems of the health sector in Ghana and many developing countries. This may be due to fewer incentives and support systems available to attract and retain health workers at the rural level. This study explored the willingness of community health officers (CHOs) to accept and hold rural and community job postings in Ghana. Methods: A discrete choice experiment was used to estimate the motivation and incentive preferences of CHOs in Ghana. All CHOs working in three Health and Demographic Surveillance System sites in Ghana, 200 in total, were interviewed between December 2012 and January 2013. Respondents were asked to choose from choice sets of job preferences. Four mixed logit models were used for the estimation. The first model considered (a) only the main effect. The other models included interaction terms for (b) gender, (c) number of children under 5 in the household, and (d) years worked at the same community. Moreover, a choice probability simulation was performed. Results: Mixed logit analyses of the data project a shorter time frame before study leave as the most important motivation for most CHOs (β 2.03; 95 % CI 1.69 to 2.36). This is also confirmed by the largest simulated choice probability (29.1 %). The interaction effect of the number of children was significant for education allowance for children (β 0.58; 95 % CI 0.24 to 0.93), salary increase (β 0.35; 95 % CI 0.03 to 0.67), and housing provision (β 0.16; 95 % CI -0.02 to 0.60). Male CHOs had a high affinity for early opportunity to go on study leave (β 0.78; 95 % CI -0.06 to 1.62). CHOs who had worked at the same place for a long time greatly valued salary increase (β 0.28; 95 % CI 0.09 to 0.47). Conclusions: To reduce health worker shortage in rural settings, policymakers could provide “needs-specific” motivational packages. They should include career development opportunities such as shorter period of work before study leave and financial policy in the form of salary increase to recruit and retain them.

This study is part of a formative research undertaken prior to the design of an intervention study for the Ensure Mothers and Babies Regular Access to Care (EMBRACE) Implementation Research [18].2 EMBRACE is a Maternal Newborn and Child Health (MNCH) initiative with the aim of increasing the uptake of maternal health care as well as reducing maternal and neonatal morbidity and mortality, employing the concept of “continuum of care (CoC)” as a key element. The implementation study was conducted in three Health and Demographic Surveillance System (HDSS) sites: Dodowa, Kintampo, and Navrongo. These sites are respectively located within the Greater Accra (southern), Brong-Ahafo (middle), and Upper-East (northern) belts of Ghana. Most communities (about 73 %) in these HDSS sites are described as rural, and 27 % of them are considered non-rural. According to the 2014 annual report of the Ghana Health Service (GHS), there are about 7210 CHOs in Ghana. Out of which, 1369 (19 %), 560 (7.8), and 484 (6.7 %) had been allocated to the Greater Accra, Brong-Ahafo, and Upper-East regions, respectively [19]. Meanwhile, according to the MOH (2007), about 14,291 CHO were targeted to be deployed into the Ghana Health System by October 2011; thus, considering population growth rate of 2.4 %, 8134 potential CHOs could have been deployed as at 2014 [20]. All the CHOs working in these three HDSS sites were contacted and subsequently recruited for face-to-face interviews and a DCE. The face-to-face interviews used a structured questionnaire that aimed to gather information on the CHOs’ demographic background, professional experience, health service, working conditions, and work attitude. The DCE used a self-administered questionnaire that measured the CHOs’ preference for different job packages. The survey instruments were jointly designed by the Japanese research team and Ghanaian research team including Ghana Health Service (GHS) and its three Health Research Centers (HRCs). Series of meetings were held involving all partners to develop the survey instruments. We followed the experimental design process’s steps summarized in the literature [21] to ensure the validity. First, we refined the problem. Then we identified the influential attributes as well as their levels to be included through literature review and discussion. Experimental design was generated along with the attributes and their levels considering statistical efficiency. A pretest was conducted in CHPS compounds outside the study area in November 2012. Sixteen CHOs were involved in the pretest: three in Navrongo, five in Kintampo, and eight in Dodowa. Some minor inconsistencies detected during training and pretesting were corrected, and the instruments were finalized before starting data collection. Data were collected between December 2012 and January 2013. The interviews and experiment were conducted successively. The instruments were not translated into the local languages because all CHOs could speak and understand English. A field supervisor, a research assistant, and a filing clerk checked the data collected by field workers before data entry. The data were then double-entered into EpiData and transferred into STATA ver. 12.1 for processing and analysis. The data from all three sites were merged into one dataset and cleaned. The DCE has become a commonly used instrument in health economics [22]. It is a useful tool to investigate the relative importance to the health worker of different attributes of employment options and to predict their hypothetical choice [23]. In a DCE, respondents are asked to choose between two or more alternatives. Table 1 illustrates an example of the job preference question that the DCE respondents are asked. The DCE determines which incentives would motivate health workers by analyzing their job preference based on the attributes presented in each hypothetical scenario. The DCE results can also be used to calculate the probability that health workers will take a job given certain conditions [7]. Example of choice set for discrete choice experiment Question: Which of these two job postings do you prefer? Select one by checking the box under the job posting you prefer for each choice set A choice experiment is a combination of the characteristics theory of demand [24] and random utility theory [25], implemented through experimental design theory and econometric analysis [22, 26, 27]. The characteristics theory of demand assumes that goods or services can be valued in terms of their constituent characteristics. The random utility model allows us to analyze choice data obtained from respondents’ stated preferences using econometric methods as follows. The preference of an individual is not embedded in just one factor, but in a combination of factors that may not be readily observable [28]. A utility level Uij is assigned to each alternative j = 1,…, J for each CHO i = 1,…, I. CHOs are assumed to choose the alternative that provides them the highest utility. The utilities are determined by the attributes of both the individuals making the choice and the alternatives available. Not all of those determinants are observed, yet one can separate overall utility into two independent additive parts: a deterministic part (systematic component), Vij, and a stochastic part (random component), εij. Then the CHO’s utility becomes It is then assumed that CHO i will choose alternative j if and only if that alternative maximizes his or her utility among all J alternatives included in the choice set Cj. In this study’s discrete choice experiment, J = 2. The probability Pi1 that CHO i chooses alternative 1 is equal to the probability that the utility Ui1 is larger than Ui2. The probability that CHO i chooses alternative 1 is Given the deterministic parts of the utility functions Vij, this probability depends on the assumptions on the distribution of the stochastic error terms εij. Although a DCE is a quantitative method to model preferences, a qualitative method could be useful to define the attributes and levels of choices when designing the choice set [29]. In order to properly understand CHOs’ motivational preferences, both literature review and pretesting were used to identify sets of motivational factors that are relatively more likely to increase acceptance of community job posting and retention. From several previous studies on job preferences in rural settings, salary, better working conditions, effective support systems, career promotion opportunity, financial incentives, better living conditions, and family support systems have been recognized as major determinants [3, 30–38]. The instrument in this study has eight choice sets with two alternatives. Each alternative has seven attributes: salary, children’s education, equipment, management style, study opportunity, housing, and transportation (Table 2). While some attributes/motivational factors are described quantitatively, others are presented in qualitative terms. For instance, salary is described as “basic salary” and “basic salary plus 50 % of the base salary.” On the other hand, management style is described as “supportive workplace” and “unsupportive workplace.” Attributes and levels For practical reasons, a fractional-factorial design, which has fewer runs than a full-factorial design, is used to develop an experimental deign in most cases [38]. This study employed an orthogonal fractional-factorial design to ensure efficiency in the design of choice sets. Orthogonal arrays are perfectly efficient because of their both balanced (each attribute level appears equally often) and orthogonal (every pair of levels appears equally often) nature [38]. In this way, eight choice sets are constructed as an orthogonal array, based on the design in Sloane [39]. Table 3 illustrates choice sets as showing the level of attributes assigned to job B for each choice set. The profile of paired job A is a foldover of that of job B, i.e., the mirror image of the design (0 = 1 and 1 = 0). Table 3 also shows the percentage of CHOs who chose job B over job A. In choice set 1, seven CHOs (3.5 %) chose job A over job B, though job B was assumed to be superior to job A. Lancsar and Louviere [40] argued that preferences that may appear to be “irrational” may in reality be compatible with some form of rationality. Since deleting such responses may be inappropriate, all respondents are included in this analysis. Attribute levels assigned for choice sets of job B and rate of choice The multinomial logit (MNL) might have been the most commonly used model in a DCE [22, 26, 41]. However, the MNL models require three assumptions: independence of irrelevant alternatives (IIA), independent and identically distributed (IID) error terms across observations, and no taste heterogeneity. Because of these assumptions, the MNL is usually criticized as not being an exact representation of choice making [28, 42]. Recent literature shows a clear shift toward more flexible econometric models such as mixed methods [22, 43]. Mixed logit relaxes the assumption of taste homogeneity. According to some systematic reviews of the literature, many studies have found evidence of preference heterogeneity and reported an improved goodness of fit using mixed logit [22, 43]. However, the mixed logit also requires assumptions about the parameters to randomize and the distributions of parameters [29]. Specification tests were performed to determine whether a model allowing a random parameter was appropriate. The selection of random parameters is usually based on either the Lagrange multiplier (LM) test proposed by McFadden and Train (2000) or the t-statistic for the deviation of the random parameter [41, 44, 45]. Either a Wald or likelihood ratio test statistic can be used for the LM test. The t-statistic for standard deviation is commonly used in the literature to determine the random parameters, considering its straightforward and simple application [44]. The LM test by a Wald statistics showed that the chi-square was 0.13, which suggested that we could not reject the hypothesis of no random coefficients. The t-statistic for the standard deviation was also checked with 50 Halton draws (Table 4). The small p value in the likelihood ratio test for the joint significance of the standard deviations implies that the null hypothesis that all the standard deviations are equal to zero is rejected [46]. The result in Table 4 shows significant preference heterogeneity for salary, children’s education, equipment, and study leave. Thus, while the LM test does not reject the assumption of no random coefficients, the t test results based on the mixed logit model show that these four attributes have preference heterogeneity. Result of the specification test (t-stat for standard deviation) Although the main-effect designs remain dominant, there has been an increase in the proportion of analyses catering for interactions [43]. De Bekker et al. [22] suggest that future work should explore the inclusion of interaction terms in DCE analyses. Interaction terms between the attributes of alternatives and the choice-invariant socioeconomic characteristics of health workers have been introduced in several previous studies [28, 47, 48]. Regarding the outcome, the use of simulation is useful with DCE data [43]. Such simulation results are potentially useful to estimate the response to job openings. The choice probability is calculated through simulation to approximate the integral. The logit formula Li(η) is calculated with draws of η. This process is repeated for a certain number of draws. The mean of the resulting Li(η)s is taken as the approximate choice probability. R is the number of draws of η, ηr is the rth draw, and SPi is the simulated probability that an individual chooses alternative i. For these reasons, we conducted mixed logit with and without interaction, using STATA’s mixlogit command, and performed probability simulation. The interaction terms are (a) number of children under 5 in the household, (b) gender dummy (male = 1, female = 0), and (c) years worked at this CHPS as a dummy variable (less than one year = 0, between 1 year and 2 years = 1, and more than 2 years = 2).

N/A

• Dima Health