Toward improving respectful maternity care: A discrete choice experiment with rural women in northeast Nigeria

Introduction There is a limited understanding of the importance of respectful maternity care on utilisation of maternal and newborn health services. This study aimed to determine how specific hypothetical facility birth experience of care attributes influenced rural Nigerian women’s stated preferences for hypothetical place of delivery. Methods Attributes were identified through a comprehensive review of the literature. These attributes and their respective levels were further investigated in a qualitative study. We then developed and implemented a cross-sectional discrete choice experiment with a random sample of 426 women who had facility-based childbirth to elicit their stated preferences for facility birth experience of care attributes. Women were asked to choose between two hypothetical health facilities or home birth for future delivery. Choice data were analysed using multinomial logit and mixed multinomial logit models. Results Complete data for the discrete choice experiment were available for 425 of 426 women. The majority belonged to Fulani ethnic group (60%) and were married (95%). Almost half (45%) had no formal education. Parameter estimates were all of expected signs suggesting internal validity. The most important influence on choice of place of delivery was good health system condition, followed by absence of sexual abuse, then absence of physical and verbal abuse. Poor facility culture, including an unclean birth environment with no privacy and unclear user fee, was associated with the most disutility and had the most negative impact on preferences for facility-based childbirth. Conclusion The likelihood of poor facility birth experiences had a significant impact on stated preferences for place of delivery among rural women in northeast Nigeria. The study findings further underline the important relationship between facility birth experience and utilisation. Achieving universal health coverage would require efforts toward addressing poor facility birth experiences and promoting respectful maternity care, to ensure women want to access the services available.

This study was conducted in Gombe state in northeast Nigeria. MNH indices are suboptimal for Nigeria as a whole but there is also considerable regional variation, with the northwest and northeast regions having lower utilisation of healthcare compared with southern regions of the country.16 17 This is true of Gombe State which, relative to national estimates, has lower coverage of at least four antenatal care visits (44% vs 57%), lower coverage of facility delivery (28% vs 39%) and higher infant mortality rate (90/1000 live births vs 70/1000 live births).17 Almost 98% of formal health services in Gombe State are provided through government at three levels—primary, secondary and tertiary levels.18 19 Primary health services are delivered through dispensaries, health posts, health centres and primary health centres, and secondary services are provided through the state specialist hospital and general hospitals, which also serve as referral centres. Tertiary services are provided through the Federal medical centre.18 19 The majority of healthcare workers in Gombe State are lower cadre for example, Community Health Extension Workers (CHEWS), Junior CHEWS and Health officers.20 21 Skilled healthcare providers including medical doctors and nurses/midwives constitute only 4% and 27% of the health workforce, respectively.20 21 The majority of women deliver in primary healthcare facilities, attended to by the lower cadre healthcare workers.20 In healthcare, patients or clients often have strong preferences for treatment and health service options, and these can affect service utilisation.22 23 In this study, conducted in March 2018, we used DCE to elicit women’s preferences for a facility-based birth based on different attribute levels presented in table 1. DCEs are health economic tools, widely used to understand user preferences in healthcare.24 DCEs ask respondents to choose their preferred service from a set of hypothetical alternatives over several choice tasks. Studying how respondents choose across repeated scenarios allows researchers to quantitatively elicit the key drivers of decision making.25–28 Attributes used in discrete choice experiment on respectful maternity care attributes influencing women’s stated preferences for facility-based childbirth Attributes and attribute levels derived from literature review,3 and revised based on qualitative findings. The first step in designing a DCE is to select the key service attributes—or characteristics—which may be important to users. To do this, we conducted a comprehensive review of the literature to identify attributes of respectful maternity care. Data bases searched included PubMed, Google Scholar, EMBASE, CINAHL and EBSCO. We further searched the reference list of the identified articles and reached out to experts in MNH to identify additional literature. From the results of the review, we selected the revised typology of mistreatment by Bohren et al,3 based on a systematic review of 65 studies from 34 countries. The typology builds on earlier work of Bowser and Hill29 and revised the dimensions of mistreatment to include seven domains: physical abuse, sexual abuse, verbal abuse, stigma and discrimination, failure to meet professional standards of care, poor rapport between women and health providers and health system conditions and constraints. We used the typology to design a qualitative study to investigate the relevance of the dimensions in the study setting, and to derive attribute levels by investigating how the different dimensions manifest for each attribute. The qualitative study included in-depth interviews with 31 women and four focus groups with 32 women (eight women per focus group). The qualitative study participants were women who had recently delivered in a health facility, purposively sampled from the communities. The qualitative study was conducted in December 2017. The qualitative data were analysed using thematic content analysis, with a manifest approach,30 in which the data analysis focused on what women said about their experience during labour and delivery. In-depth interviews and focus groups, alongside the literature review findings, informed our selection of six attributes and a total of 18 attribute levels relevant to the context (three attributes of four levels each and three attributes of two levels), these are highlighted in table 1. Further, the qualitative study provided us with locally appropriate expressions and language translation, enhancing respondents ease of comprehension.31 From the number of attributes and attribute levels decided, a full factorial design would have consisted of 729 (34×32) possible alternatives—too many for a survey, and tedious for the respondents to handle.27 32 Therefore, we developed 16 choice sets based on a fractional factorial orthogonal main effects design from a design catalogue, that ensured the inclusion of levels proportionally (level balance) with no correlation between levels of different attributes (orthogonal).32 We constructed an unlabelled choice experiment of three choice alternatives, with each set consisting of two unlabelled facility alternatives and home delivery, an example is highlighted in table 2. We decided to include a home delivery option to avoid bias in estimating parameters in a forced choice design.33 Example of discrete choice experiment choice task as shown to women who had a facility-based childbirth The choice sets were reviewed and validated for content in collaboration with a group of health workers (doctors, nurses and midwives) working in Gombe, followed by a pilot test with 40 women with similar characteristics as the target sample (women with recent facility-based childbirth). None of the women that took part in the pilot participated in the main study. The pilot exercise involved completing the DCE and answering questions afterwards regarding the exercise, including clarity of instructions, understanding and relevance of the choice sets and ease in answering. Following the pilot, minor modifications were made which included small changes to the wording used to describe the attribute levels, and to introduce the choice sets. For example, participants from the pilot suggested we use ‘hospital’ rather than ‘health facility’ when describing the choice sets. The final DCE tool was incorporated into a larger study instrument consisting of questions on sociodemographic information, and experience of care during institutional delivery. The final questionnaire was programmed in CSPro.34 The final DCE was nested within an ongoing measurement, learning and evaluation project in Gombe State.35 As part of that project, a total of 1889 birth observations were carried out in ten primary health facilities across three time points: June 2016, March 2017 and August 2017.36 The health facilities were spread across six of the 11 local government areas in Gombe State. Six of the health facilities were located in urban settings and four were located in rural settings. Subsequently, in August 2018, a simple random sample of 450 observed women was taken and these women were followed up at home to ask what they recalled about their facility birth experience. Of the 450 eligible and selected women, 426 (95%) women were successfully interviewed in March 2018 at their homes, while 24 (5%) of the eligible women selected could not be reached or were unable to participate. According to commonly used rules of thumb for DCE sample size calculation, the sample size of 426 women was enough to guarantee precision in the estimation of all model parameters.37 Orme posit that the sample size N required for main effects depends on the number of choice task (t), the number of alternatives (a) and the number of cells (c), according to the following equation N>500 c/(t×a).37 38 Based on this equation, an approximate sample of 56 participants would have been sufficient to model our preference data. Lancsar and Louviere suggested that a sample of 20 respondents per questionnaire version as adequate to estimate reliable DCE models.39 However, we recruited much larger sample to allow for more variability between respondents and to allow for other post hoc analysis. Data were collected using personal digital assistants, and interviews took about an hour to complete. Before data collection, data collectors and supervisors received 5 days of training on data collection and the study tools. The discrete choice data were analysed based on the random utility model.25 We specified our analytical model around a utility maximising framework. This assumes, given alternatives to choose from, a respondent i (i=1, …, N) will choose the one alternative that yields the maximum utility among the choice bundle (j=1, 2, 3, …) at the moment of choice. The utility of the respondent is defined by a deterministic or observable component and a random error component: Where Uij represents the utility of respondent,vij the observable component and εij the random error term with standard statistical properties. Following from equation (1) the probability of a respondent selecting a specified place of delivery is modelled. The probability of a choosing a place of delivery is determined by the indirect utility function for the respondent i from choice j in choice set s, assuming this is linear and additive and of the form: Where Vijs represent the utility derived from a choice, and Xijsβ the utility component and εijs as the random component. The vector Xijs is specified below, where β1-6 represent the design attributes of the choice experiment and β0 the constant. We analysed the DCE data using STATA V.15. We conducted a dominance test of internal validity, presenting an additional choice set to all the respondents where one of the hypothetical health facilities is more favourable, but do not exclude participants if they fail this test.40 We first estimated standard multinomial logit model (MNL) (online supplementary table S1) to provide a benchmark for more detailed analysis, and which were used as starting values for estimating a mixed multinomial logit model (MMNL) (online supplementary table S2). Further, we fitted MMNLs to the DCE data from rural women investigated, using 500 Halton draws. We used the MMNL to avoid restrictions due to the independence of irrelevant alternatives assumption required to interpret findings from the MNL.41 42 bmjgh-2019-002135supp001.pdf When estimating MMNLs all parameter estimates may be treated as random, and in a model where more than one parameter or all parameters are estimated as random, there is no requirement that the distribution be the same.43 McFadden and Train have shown that mixed multinomial logit does not embody any theoretical restrictions on the distribution of preferences or the choice model.44 Also, with mixed multinomial logit, we can accurately approximate any choice model, with any distribution of preferences.44 Therefore, all attributes levels were effects coded, specified as random components and a multivariate normal distribution—a generalisation of the one-dimensional normal distribution to more than one variable was assumed. Assuming a multivariate normal distribution to represent the distribution of a multivariate random variable that is made up of multiple random variables that can be correlated with each other was a rational choice because this allows for correlation which introduces error dependence across the alternatives in each choice situation.45 46 Furthermore, assuming a random effect to be respondent specific induces correlation across choice situations, thus accounting for the dependence structure in unobserved utility among the repeated choices of a respondent due to the panel structure of the data.45 Findings are reported below in line with Strengthening the Reporting of Observational Studies in Epidemiology statement.47 An important output from the main effects MMNL I estimation is the SD associated with each parameter estimate, which indicate the distribution about the corresponding mean preference weight, also the preference variability among women around the mean. Some of the differences between the parameter estimates in the multinomial logit and MMNLs further indicate that preferences vary among women in Gombe. Respondents characteristics are constant across alternatives for example, a woman’s ethnicity does not change because she is considering delivering in a health facility as opposed to home delivery.43 Respondents characteristics are likely to influence their choice decisions, but they are not part of the attribute description of alternatives and not a direct source of utility.43 One way to predict how respondents’ characteristics influenced their choices is to extend equation (2) to allow attribute weights to vary with respondent characteristics, through the inclusion of interaction terms between attribute and individual characteristics.48 Therefore, to understand the preference (taste) variability among women the MMNL (model I) was extended with interaction terms between attribute levels with significant SD and women sociodemographic characteristics likely to influence women’s behaviours or predisposition to mistreatment, as highlighted in equation (4) below (and MMNL model II). This approach often leads to models fit improvement. It is also easy to interpret the parameters related to covariates in a relative sense both within and between alternatives—holding all else equal.43 48 Revelt and Train have shown that entering demographics into the MMNL itself is a more direct and accessible way to hypothesis testing, as such, this approach has been widely used in DCE studies.49 50 Where β1-6 represent the design attributes of the choice experiment and β7–26 the parameters for the interaction terms that were introduced for the women sociodemographic characteristics variables. The sociodemographic characteristics used included ethnicity coded (0=others, 1=Fulani), age coded (0= ≤29 years, 1= ≥30 years), education coded (0=low-level education, 1=high-level education) and socioeconomic status (SES) coded (0=low SES (lower 60%), 1=upper SES (upper 40%).

N/A

• umar, nasir
• quaife, matthew
• exley, josephine
• shuaibu, abdulrahman
• hill, zelee
• marchant, tanya