Objective Maternal and newborn mortality rates are high in peri-urban areas in cities in Kenya, yet little is known about what drives women’s decisions on where to deliver. This study aimed at understanding women’s preferences on place of childbirth and how sociodemographic factors shape these preferences. Methods This study used a Discrete Choice Experiment (DCE) to quantify the relative importance of attributes on women’s choice of place of childbirth within a peri-urban setting in Nairobi, Kenya. Participants were women aged 18–49 years, who had delivered at six health facilities. The DCE consisted of six attributes: cleanliness, availability of medical equipment and drug supplies, attitude of healthcare worker, cost of delivery services, the quality of clinical services, distance and an opt-out alternative. Each woman received eight questions. A conditional logit model established the relative strength of preferences. A mixed logit model was used to assess how women’s preferences for selected attributes changed based on their sociodemographic characteristics. Results 411 women participated in the Discrete Choice Experiment, a response rate of 97.6% and completed 20,080 choice tasks. Health facility cleanliness was found to have the strongest association with choice of health facility (β = 1.488 p<0.001) followed respectively by medical equipment and supplies availability (β = 1.435 p<0.001). The opt-out alternative (β = 1.424 p<0.001) came third. The attitude of the health care workers (β = 1.347, p<0.001), quality of clinical services (β = 0.385, p<0.001), distance (β = 0.339, p<0.001) and cost (β = 0.0002 p<0.001) were ranked 4th to 7th respectively. Women who were younger and were the main income earners having a stronger preference for clean health facilities. Older married women had stronger preference for availability of medical equipment and kind healthcare workers. Conclusions Women preferred both technical and process indicators of quality of care. DCE’s can lead to the development of person-centered strategies that take into account the preferences of women to improve maternal and newborn health outcomes.
The study was conducted at Embakasi-North, a sub-county in Nairobi County with a population of 181,388 people and is located about 10 km to the East of Nairobi City. Embakasi-North is home to Dandora, an area that houses the largest municipal dumpsite in Nairobi, and is characterized by low-income residential housing estates. The area is served by a mix of public, private and faith-based facilities of different levels. Mama Lucy Maternity Hospital, a secondary referral hospital, is located in the neighboring sub-county. Maternity facilities utilized by women in these informal settlements vary widely in terms of quality of care that they provide. The facility-based delivery rate in Nairobi is high with approximately 88.7% of women delivering within a health facility [6]. However, within peri-urban settings and informal settlements in Nairobi have been known to have lower rates of facility-based delivery [6]. The study entailed conducting a literature review and doing a qualitative study to determine attributes and attribute levels that were important to women. The qualitative study sought to explore the perceptions and experiences of women visiting health facilities in the area. The results of the qualitative study can be found here [30]. After obtaining informed consent from the women, trained facilitators led the focus group discussions (FGDs). Women were asked to explain how they made the choices and identify which facility features drove their child birth choices. Women were purposively selected and each FGD had 6–8 women. The characteristics of the 40 women interviewed are contained in (S1 Appendix). Qualitative Interview data were entered into Nvivo 11 and coding done. Thematic analysis was done following the six key steps, namely, familiarization with the data, coding, grouping codes, identifying themes, additional coding and refining of themes, and writing up the results. Four broad themes were identified: perceived quality of delivery services, financial access, physical amenities at the facility, and health worker’s strike. (See S2 Appendix). The themes helped in deriving attributes and attribute levels. The selected attributes were piloted on 30 women residing outside of the study setting in a neighboring sub-county to test for suitability and the cognitive response of the women in understanding the selected attributes. The pilot showed that the attributes could be easily understood and traded-off by the women. Some attributes such as the costs of delivery attributes were revised and were chosen based on what the women reported they had paid when going to deliver. The costs ranged from 3000 to 8000 Ksh for normal uncomplicated deliveries in both the public and private health facilities. The costs were inclusive of out-of-pocket costs that the women were charged during delivery. These costs were present even at facilities that had the “free delivery” policy. For a complete list of the attributes and attribute levels selected for the DCE, See Table 1. *Note. Costs are in Ksh (1 USD = 100Ksh) Costs are not zero even with free delivery policy due to incidental fees charges at government facilities. The study was designed as an unlabeled DCE with sixteen choice set presented under three alternatives: alternative of health facility A, alternative of health facility B, and an opt-out alternative where the woman would choose none of the two facilities, explained as a preference for home delivery. S3 Appendix shows a sample choice-card with a scenario showing the final attributes and attribute levels included. The attributes of the health facility were explained to the women using a choice-card that contained a brief description of the definition of the attributes. For example. Cleanliness meant a health facility that had a clean ward with clean beds, bathrooms and toilets (See S4 Appendix). All attributes in the choice experiment were dichotomous, except cost, which had three levels. This resulted in a design of (25 x 13) = 96. The number of alternatives of attribute levels in the full fractional design was calculated to (96*95)/2 = 4560. A fractional factorial design helped to reduce the choice-sets to 16, making it simpler for the respondents. We used JMP software for a D-efficient experimental design and resulted in a D-error of 0.3 (JMP Pro). (See S5 Appendix). The D-efficient design also allowed for favorable design such as orthogonality, level balance, minimum balance and overlap [31].The 16 choice-set questions were generated from the design. The choice-sets were grouped into two through a process called blocking using ODK software and each woman answered eight questions in a single block. Following administration of informed consent, a random sample of women of reproductive age (18–49 years) were recruited from a larger household survey in the area. The inclusion criteria were women who had delivered in the past five years. The main household questionnaire was a composite tool carrying questions from the Kenya Demographic Health Survey and the African Population and Health Research Survey [5, 6]. The survey contained questions on women’s sociodemographic characteristics and maternal health services utilization variables. For The questionnaire (See S6 Appendix), and details of the sampling process from the larger household survey provided in (See S7 Appendix). The sample size for the DCE was calculated using the Johnson and Orme methodology [32]. The household survey was conducted between August and September 2017 by trained research assistants using Open Data Kit (ODK) platform. This was followed by the DCE survey, which asked women to imagine a hypothetical scenario where they were expecting a baby and had to choose between facilities A and B for delivery (or none). The women were told that the opt-out option implied home delivery. They were also told that there were no wrong or right answers, and that they were free to stop the experiment at any time (See S8 Appendix). Ethics approval for the study was provided by the African Medical Research Foundation (AMREF) research Committee, the National Commission for Science and Technology (NACOSTI) as well as the Country Directors of health in charge of the sub-county. The DCE data was analyzed using the random utility model, a model that expresses the utility ‘U’ in of an alternative i in a choice set Cn (perceived by individual n) as two parts: 1) An explainable component specified as a function of the attributes of the alternatives V (Xin, β); and 2) an unexplainable component (random variation) ε in. [33]. The individual n will choose alternative i over other alternatives in a choice set C if and only if this alternative gives the maximized utility. The relationship between the utility function and the observed k attributes of the alternatives can be assumed under a linear-in-parameter function [34]. Therefore, the utility the respondents attach is related to the attribute and attribute levels within the choice-sets, meaning that if alternative i is chosen within a choice set, i will yield the maximum utility compared to j alternatives. Α is the alternative specific constant, x are the attributes in the DCE and β are the coefficients describing the marginal utility of the attribute. The standard conditional logit model is below: The data were imported and analyzed in Stata 15 (StataCorp LP, College Station, USA). Descriptive statistics were calculated for the non-DCE variables. The cost attribute was assumed to be linear while all other attributes were categorical variables, therefore non-linear. A base conditional model was used to estimate the mean change in utility, preference which respondent placed on attributes [34]. αi is a constant term that represents the general preference for place of delivery at a health facility compared to the alternative of opting out and having a home delivery. Dummy coding was used for the data, each attribute level was assigned a value of 1 whenever it was retained and 0 when omitted. The cost of delivery service was entered in the model as a continuous variable. All the other five variables were coded as categorical variables. The Utility Model makes the assumption that women will trade-off between the different attribute levels and choose the alternative that gives the greatest utility. The conditional model is suitable for estimating average preferences across respondents. The utility function was estimated for the following model: αi is the alternative specific constant (ASC) term that shows the preference for place of delivery (either a health facility or home), β’s 1–11 are the parameters for each of the attribute levels and ε is the error term. The dependent variable is the place of delivery represented by the unlabeled choices health facility A, health facility B and the opt-out (home delivery), while the independent variables are the respective attribute levels of the characteristics of the place of delivery. The base conditional logit model assumed homogeneous preferences across respondents [34]. The output of the conditional logit model contains the beta which shows the magnitude of the preferences for the attribute. Due to the assumption of irrelevant independent alternatives, the presence of heterogeneity in choices we estimated a generalized mixed logit model to assess for preference heterogeneity amongst the women [35].This was done by extending the generalized model and testing interactions between the sociodemographic and the women’s attributes in order to investigate how preferences may vary according to observed individual characteristics. The sociodemographic characteristics that were included as interaction terms include sociodemographic characteristics that have been known to influence place of delivery in Kenya were also included such as maternal age, marital status, education and income status [36–39]. The output of the mixed logit model includes both the mean and the standard deviations of the random parameter estimates with confidence levels. The mean parameter estimate represents the relative utility of each attribute while the standard deviations for a random parameter suggest the existence of heterogeneity in the parameter estimates over the sampled population around the mean parameter estimate i.e., different individuals possess individual-specific parameter estimates that may be different from the sample population mean parameter estimates [35]. The p-value of the interactions shows statistical significance for an interaction between sociodemographic variables and attributes hence signifying the influence of the woman’s characteristics. Insignificant parameter estimates for derived standard deviations indicate that the dispersion around the mean is statistically equal to zero, suggesting that all information in the distribution is captured within the mean. The theoretical validity of the design will be explored by examining the signs and significance of parameter estimates [35]. A correlation matrix analysis was also done to ensure that there is no inter attribute correlations between certain attributes that are close in semantic meaning (See S9 Appendix).