Moving toward patient-centered care in Africa: A discrete choice experiment of preferences for delivery care among 3,003 Tanzanian women

Objective: In order to develop patient-centered care we need to know what patients want and how changing socio-demographic factors shape their preferences. Methods: We fielded a structured questionnaire that included a discrete choice experiment to investigate women’s preferences for place of delivery care in four rural districts of Pwani Region, Tanzania. The discrete choice experiment consisted of six attributes: kind treatment by the health worker, health worker medical knowledge, modern equipment and medicines, facility privacy, facility cleanliness, and cost of visit. Each woman received eight choice questions. The influence of potential supply- and demand- side factors on patient preferences was evaluated using mixed logit models. Results: 3,003 women participated in the discrete choice experiment (93% response rate) completing 23,947 choice tasks. The greatest predictor of health facility preference was kind treatment by doctor (β = 1.13, p<0.001), followed by having a doctor with excellent medical knowledge (β = 0.89 p<0.001) and modern medical equipment and drugs (β = 0.66 p<0.001). Preferences for all attributes except kindness and cost were changed with changes to education, primiparity, media exposure and distance to nearest hospital. Conclusions: Care quality, both technical and interpersonal, was more important than clinic inputs such as equipment and cleanliness. These results suggest that while basic clinic infrastructure is necessary, it is not sufficient for provision of high quality, patient-centered care. There is an urgent need to build an adequate, competent, and kind health workforce to raise facility delivery and promote patient-centered care. Copyright:

This discrete choice experiment was conducted as part of baseline data collection for a maternal and newborn health quality improvement project currently underway in four rural districts of Pwani Region, Tanzania (ISRCTN 17107760). The study areas are primarily rural and most residents are employed in small-scale subsistence farming or unskilled manual labor [17]. The 2010 DHS reported higher healthcare facility utilization than the national average, with 74.9% of women utilizing a healthcare facility for their most recent birth.[17] The study includes 24 government-run primary healthcare facilities and their designated catchment villages.[12,18] For the population-based survey and discrete choice experiment (DCE) we conducted a full census of all households in the designated catchment areas. All women who were at least 15 years of age and delivered between six weeks and one year prior to the interview were invited to participate in a structured interview, including a DCE. If eligible women were not available on the day interviewers visited, at least two additional visits were made to reach the women. Participants were included in the current analysis if they completed at least part of the DCE. All eligible women were informed of the purpose of the study and their right to refuse to participate. All interviewed participants provided written consent, or in the case of minors, their written assent and guardian written consent. The study was approved by the ethics review boards at Columbia University and Harvard University in the U.S., as well as the Ifakara Health Institute and the Tanzanian National Institute for Medical Research in the United Republic of Tanzania. In order to identify potential delivery care attributes that were important to women we conducted a review of the literature and held meetings with policy-makers, healthcare workers, and health systems researchers in Tanzania. We then held five focus groups, each with six to eight women who had delivered a child in the past year and lived within the study districts, but not in catchment villages. Trained facilitators led the focus group discussions; participants provided informed consent. During these focus group discussions women were asked to identify important features of delivery care that they used to select a delivery facility and to distribute 20 token resources (stickers) among a list of care attributes. We focused on attributes that represented inputs and processes of care, as these elements can be directly addressed by implementers and are less likely to dominate other answer choices than an outcome attribute, such as life birth. We selected the six healthcare facility attributes that women ranked as most important when distributing their resources for inclusion in the DCE. We confirmed that women also discussed these attributes during the focus group discussions as characteristics they consider when choosing a delivery facility. A local artist created graphics for each attribute level in order to facilitate understanding in this low-literacy population. The full survey and the discrete choice experiment were then piloted with 40 women living outside the study catchment area who had delivered a child in the six weeks to one year prior to pilot. We used information from the pilot survey (e.g. how much the women had paid for their most recent delivery) as well as information from the DCE and conversations with the women after the pilot to refine the DCE, improving clarity, and ensuring locally relevant range levels for the price paid for the delivery, hereafter referred to as “cost.” The six selected DCE attributes were: kind treatment by the health worker, healthcare facility privacy, healthcare facility cleanliness, modern equipment and medicines, health worker medical knowledge, and cost. The health workers were referred to as “doctor” in the DCE as this was the locally appropriate label. However, most healthcare providers in the primary care clinics in this region are clinical officers, nurses, or medical attendants. All attributes were dichotomous except cost, which had five levels. This gives rise to 160 possible facility combinations in the full factorial design (51X25). In order to minimize the burden on respondents while maximizing efficiency, we selected five sets of eight choice scenarios (choice tasks). This was done using an experimental design that minimizes overlap of attribute levels within each task, maximizes level balance such that different attribute levels appear with approximately equal frequency across tasks, and achieves orthogonality among attributes by selecting levels for each attribute independently (Sawtooth Software Inc., Sequim, WA, USA). The average attribute efficiency was 0.95 (range: 0.86–0.98). This suggests a highly efficient design. The final DCE included eight experimental choice tasks plus a fixed choice task that was kept constant across all five sets and used to test the internal predictive validity of the model. Each respondent was thus presented with nine choices. The final set of attributes and their levels are outlined in Table 1 and a sample scenario is shown in Fig 1. The population-based survey collected information on women’s demographics, household characteristics, maternal and newborn health experiences and health system use and satisfaction. We used these data to assess how individual factors and health system experiences influenced women’s stated preferences. We assessed demographic and household characteristics including age, education (any secondary education versus less education as women with secondary education have been shown to be more selective users of health care), media exposure (index ranging from 1–9 constructed from frequency of exposure to newspapers, radio, and television), and primiparity (as primiparous women demonstrate different utilization patterns [19,20]). Using principal components analysis of a set of 18 questions on ownership of household assets we constructed a relative wealth index [21]. We compared the wealthiest 20% of women to all others as wealthier women may have increased access to facilities, which may shape their expectations and preferences. Because a woman’s experiences may affect her expectations for care and her preferences, we assessed her recent experiences, including place of most recent delivery (healthcare facility versus home), number of antenatal care (ANC) services received for most recent delivery (a standardized index of 8 items: weight, height, blood pressure, urine sample, blood sample, malaria prophylaxis, tetanus vaccine, and iron supplements), and number of services received during most recent delivery (a standardized index of nine recommended services: mother checked; baby checked; uterotonic received; and mother given advice on: immediate feeding, exclusive breastfeeding, umbilical cord care, washing hands, immunization, and how to avoid chilling the baby). Additional supply-side factors included distance from woman’s hamlet to the nearest hospital (spherical distance from the center of the hamlet to the hospital using the sphdist command in Stata 13.1) and the density of secondary health facilities within 25km of the woman’s hamlet. For distance calculations when the hamlet location was unknown the village center was substituted (23.5%). The survey was conducted face-to-face in Swahili by six teams of five local interviewers. All interviewers underwent 11 days of training, including one full day of DCE training and practice. Each choice task was presented to the respondent using a single sheet of paper (Fig 1). The interviewer read a standardized script that asked the respondent to imagine that she was pregnant and given the choice between delivering at the two facilities presented. She was asked to choose at which healthcare facility she would prefer to deliver, and was reminded that there was no right or wrong choice and that her answers would be confidential. Her responses were recorded using hand-held tablet computers. Data collection occurred between February 13 and April 28, 2012. Data were imported into Stata version 13.1 (StataCorp LP, College Station, USA) and were examined for outliers and missingness. Descriptive statistics were calculated for the non-DCE variables of interest. The attribute utilities were estimated with a mixed logit model using 500 Halton draws, with normally distributed parameters and an independent covariance structure. Conditional logit models allow respondents’ unselected alternative choices in each task to be taken into account when estimating preferences, as well as individual-specific characteristics. The mixed logit model is an extension of the standard conditional logit model that allows for attribute coefficients to be randomly distributed [22]. Mixed logit models are frequently used to analyze DCEs [23–25]. To assess the potential effects of both supply- and demand- side characteristics on the utility of each attribute, we analyzed a set of models that introduced individual characteristics as an interaction between that characteristic and the attributes of the facility. Cost was specified as a fixed effect in each model and all other clinic attributes were random [26]. We were particularly interested in characteristics that can be expected to change in the future. We assessed interactions with the demographic characteristics of wealth, secondary education, and primiparity. We also assessed an interaction with the women’s exposure to media, as this may influence her preference for delivery. Finally, because a woman’s experiences may affect her expectations for care and her preferences, and because facility utilization for delivery is increasing in Tanzania, we assessed an interaction with the place of most recent delivery (healthcare facility versus home).[17,27] We performed several validity tests, including evaluating the in-experiment predictive validity of the mixed logit model. The internal predictive validity of our model was assessed by comparing women’s predicted vs. actual choices on the fixed choice card (which is excluded from the sample used to estimate the mixed logit model). The fixed choice card offered women the option of selecting a healthcare facility with desirable nontechnical attributes (a kind doctor, privacy, and a clean and tidy facility) or a facility with desirable technical attributes (modern equipment and drugs, and a doctor with excellent medical knowledge). We report the actual choice and mean predictive values.[28] All analyses were conducted using Stata 13.1 (StataCorp LP, College Station, USA).

• larson, elysia
• vail, daniel
• mbaruku, godfrey
• kimweri, angela
• freedman, lynn p.
• kruk, margaret e.