Data sources
Several datasets from the Malawi Demographic and Health Survey (MDHS) program were used to create a master analysis dataset. First, woman’s questionnaire data from the 2015–16 MDHS were retrieved [11]. The 2015–16 MDHS was implemented using a two-stage cluster sampling design. In the first stage, all 28 administrative districts in Malawi were stratified into 56 urban and rural strata [11]. For each stratum, a sample of standard enumeration areas (SEA) was selected based on the complete list of SEAs derived from the 2008 sampling frame of the Malawi Population and Housing Census [11]. Selection of the SEAs also occurred in two stages [11]. One hundred seventy-three urban SEAs and 677 rural SEAs were independently selected using probability proportional to the size of the SEA [11]. Then, 30 households from urban clusters and 33 households from rural clusters were selected using an equal probability systematic selection from the complete list of households in selected SEAs [11]. Selected SEAs with more than 250 households were segmented due to their large size and only one segment of households with probability proportional to the segment size was used for household listing [11]. The woman’s questionnaire collected various health and demographic data from all women in the reproductive age range between 15 and 49 years living in the selected households or who were found as visitors in the selected households on the day of the survey [11].
Second, the GPS coordinates of the centroids of the study clusters from the 2015–16 MDHS were linked with the woman’s data through unique identifiers. The published GPS coordinates are not the exact locations of the study clusters because they have been systematically displaced using the “random direction, random distance” method [16]. This was done in order to protect the respondents from the threat of identity disclosure [16]. Each GPS coordinate was displaced a distance of up to two kilometers if it was an urban cluster. The majority of rural clusters (99%) were displaced a distance of up to five kilometers [17]. A randomly selected 1 % of the rural clusters were displaced a distance of up to ten kilometers [17].
The third data source was the 2013–14 Malawi Service Provision Assessment (MSPA). These data were collected from a census of public and private facilities in all 28 districts including facilities run by the government, Christian Health Association of Malawi, other faith-based organizations, non-governmental organizations, private for-profit organizations and others [18]. The 2013–14 MSPA includes a total of four different questionnaires – Facility Inventory, Health Provider Interview, Observation Protocols and Exit Interview questionnaires with select clients [18]. The information about whether health facilities provide PNC services was obtained from the health provider interview. At each facility, the goal was to interview an average of eight health providers. For facilities that had less than eight providers, every provider was interviewed. For larger facilities, providers who were deemed most knowledgeable about their facility were selected for interview. If any health provider mentioned that he or she provides PNC services, the corresponding health facility was labeled as one providing PNC services.
Then, the GPS coordinates of the health facilities in 2013–14 MSPA were spatially linked with the woman’s questionnaire data from the 2015–16 MDHS. Three distance bands around household clusters were considered for the spatial linkage. Health facilities located between 0 km and 5 km (≤ 5 km) were grouped as the closest distance band. Health facilities located between 5 km and 10 km (> 5 km and ≤ 10 km) were grouped as the mid-range distance band. Health facilities located between 10 km and 15 km (> 10 km and ≤ 15 km) were grouped as the farthest distance band. Unlike the GPS coordinates of the 2015–16 MDHS household clusters, the GPS coordinates of the facilities were not displaced and reflect the true location [17]. Skiles, Burgert, Curtis and Spencer compared three data scenarios where methodological considerations of geographically linking DHS household clusters with health facilities were explored [19]. The study used the 2007 Rwanda Service Provision Assessment and the 2007–2008 Rwanda Interim Demographic and Health Survey [19]. In the study, the most ideal data scenario was having a census of all health facilities and undisplaced geographic locations of household clusters [19]. Other less ideal scenarios were either having a census of all health facilities and displaced household cluster locations or having a sample of health facilities and displaced household cluster locations [19]. The current study fits in with the second scenario where data were collected from a census of all health facilities in Malawi but the household cluster locations were randomly displaced. Skiles et al. reported that in the second scenario, using a Euclidean buffer of 5 km resulted in 5.9 to 9.2% of hospitals being misclassified, 7.0 to 12.4% of health centers being misclassified and 4.9 to 7.6% of health posts being misclassified [19]. The degree of misclassification error due to random displacement of household clusters in Malawi is expected to be similar to that reported in Skiles et al. but the possibility that there could be greater error in Malawian context cannot be ruled out completely.
The Woman’s Questionnaire data from the 2015–16 MDHS, the GPS coordinates of the 2015–16 MDHS household clusters, the 2013–14 MSPA data and the GPS coordinates of the 2013–14 MSPA facilities are all publicly available on the Demographic and Health Surveys Program website [20] upon request.
Variables
All analyses were stratified by place of delivery. This is because types of health facilities and the proximity of these health facilities from household clusters are presumed to influence receipt of PNC differently based on where women delivered. Women who delivered at home may seek PNC at a health facility at the time of their choosing or receive a home visit by a health worker. In either circumstance, the proximity and the types of health facilities nearby women’s homes can potentially influence their receipt of PNC. However, women who delivered at health facilities face a slightly different set of options. After delivery at the facility, women may receive PNC on site before returning home for the first time, return home first then seek PNC at a later time at a facility or return home first then receive a postnatal home visit by a health worker. Due to these differences in care-seeking options based on place of delivery, there were several outcome variables used for analyses (in separate models). For women who delivered at home, the main outcomes were maternal PNC within 1 day of birth, newborn PNC within 1 day of birth, maternal PNC within 7 days of birth and newborn PNC within 7 days of birth. For women who delivered at health facilities, the main outcomes were maternal PNC between day 1 and day 7 and newborn PNC between day 1 and day 7. PNC between day 1 and day 7 was considered because women who received PNC right after delivery but before leaving the facility (to return home for the first time) will most likely do so in the first 24 h. Looking at this time interval can potentially capture the effects for women who were discharged and came back to a facility for a first or second postnatal check. As a supplementary analysis (see Additional file 1), PNC within the first day was still considered for women who delivered at health facilities to check for the assumption that some women receive PNC before discharge. In this case, the types of health facilities and their proximity should not have any significant positive influence on PNC seeking decisions because women are already at the facilities. All of the outcomes are binary with “1” indicating PNC in the specified time period and “0” otherwise. There were no women who responded “don’t know” for maternal PNC. For newborn PNC, less than 1% of the women responded “don’t know.” Among all rural women who delivered in the 5 years prior to the survey, less than 1% of the women had missing data for maternal and newborn PNC.
There were three main types of binary indicators for health facilities: clinic-level facilities providing PNC, health centers providing PNC and hospitals providing PNC. Clinic-level facilities included maternities, dispensaries, clinics and health posts. Health centers only included facilities designated as health centers. Hospitals included central hospitals, district hospitals, rural/community hospitals and other hospitals. Health centers were set apart from other lower-level facilities because they comprise the largest number among all health facilities in Malawi [18]. In addition, compared to other lower-level facilities, health centers are much more likely to offer basic client services and delivery-related services in Malawi [18]. Types of facilities are meant to serve as indicators of the level of quality that can be provided at the facilities while three separate rings of buffers (0–5 km, 5–10 km and 10–15 km) indicate different levels of proximity or distance from the household clusters. See Fig. 1 for a visual illustration.
Covariates in the models included season in which women gave birth, ownership of TV or a radio, whether cost of treatment is a perceived problem, women’s age at the time of the survey, women’s education, women’s employment, household wealth, number of total births, newborn size, newborn sex, religion and region. For women who delivered at health facilities, cesarean section, whether or not women were checked before facility discharge and whether or not the newborns were checked before facility discharge were also included in the models. Number of antenatal visits was not included in the models due to potential endogeneity. Types of facilities and their proximity to household clusters could influence decisions regarding antenatal visits. Antenatal visits could also mediate the effects of facilities on PNC use, which is a classic case of endogeneity [21] where it is correlated with the error term when left in the models. Hence, the reported effects of health facilities are total effects, rather than direct effects, which account for the omitted mediated pathways (in the model) through antenatal visits.
Season in which women gave birth was coded as “warm-wet season (November to April)”, “winter-dry season (May to August)” or “hot-dry season (September and October)”. It was meant to proxy varying road conditions due to seasonal rainfalls. Ownership of TV or a radio was a binary variable meant to proxy potential exposure to health messages in the media. Obtaining money for treatment of any sickness being a big perceived problem was a binary variable meant to proxy financial barriers to accessing care. Women’s age at the time of the survey was coded as “15 – 24”, “25 – 34” and “35 – 49”. Women’s education was coded as having “no education”, “primary education” and “secondary education or higher”. Employment was a binary variable with “1” indicating currently working in either formal or non-formal sectors (including but not limited to agricultural, fishery and sales) and “0” otherwise. Household wealth was a rural-specific quintile variable constructed by DHS using principal components analysis [11]. It was coded as “poorest”, “poorer”, “middle”, “richer” or “richest”. Number of total births was coded as “1”, “2 – 3” and “4 or more”. Newborn size was subjectively reported by the respondents as either “very large”, “larger than average”, “average”, “smaller than average” or “very small”. This variable was meant to proxy potential maternal and/or newborn complications. Newborn sex was coded as “male” or “female”. Religion was coded as “Catholic”, “Other Christian” or “Muslim, no religion or other unspecified religion”. Those with no religion or other religion were less than 1% and for this reason, they were grouped together with those of Muslim faith, the second smallest group. Region was coded as “Northern”, “Central” and “Southern” which are three administrative regions in Malawi. Cesarean section was a binary variable meant to proxy maternal complications. Whether or not mothers received a check before facility discharge and whether or not newborns received a check before facility discharge were both coded as binary variables. These two variables and cesarean section were only included for women who delivered at health facilities.
Analysis
A series of descriptive analyses were conducted. Then, generalized estimating equations (GEE) were used in STATA version 15.1 for all of the binary outcomes, with each in separate models. Clustering of households was accounted for by specifying the error correlation structure to be “exchangeable” which means that the variance-covariance matrix for each household cluster has an identical structure [22].
In equation form, the GEE models including the aforementioned outcomes, main predictors and covariates can be summed up below. For simplicity, the meaning of each unfamiliar notation is explained in the corresponding subscript.
$$ {Y}_{outcomes}={\alpha}_{intercept}+{\beta}_1{Clinics}_{0-5 km}+{\beta}_2{Health\ Centers}_{0-5 km}+{\beta}_3{Hospitals}_{0-5 km}+{\beta}_4{Clinics}_{5-10 km}+{\beta}_5{Health\ Centers}_{5-10 km}+{\beta}_6{Hospitals}_{5-10 km}+{\beta}_7{Clinics}_{10-15 km}+{\beta}_8{Health\ Centers}_{10-15 km}+{\beta}_9{Hospitals}_{10-15 km}+{\beta}_X{X}_{covariates}+{\varepsilon}_{error\ term} $$
The coefficients, denoted by ’s, were converted into differential effects in STATA using the “margins” command [23]. Differential effects were derived because they are more intuitive to understand than interpreting odds ratios. Differential effects can be obtained by first calculating the predicted probability of the referent category and the predicted probability of the alternative category and then taking the difference between the two. A general interpretation would be percentage point changes in the probability of the outcome given the alternative condition compared to being in the referent category. More specifically, a full interpretation for β1, for example, would be the percentage point changes in the average probability of receiving PNC associated with having a clinic-level facility providing PNC within 5 km compared to not having a clinic-level facility providing PNC within 5 km (averaged across all household clusters). This is controlling for the distribution of other health facilities providing PNC within 5 km, between 5 km and 10 km and between 10 km and 15 km as well as aforementioned covariates included in the models. In order to avoid repetition, however, a shortened version of the interpretation is presented in the results section.
Differential effects were only calculated and reported for the main predictors, as they are the focus of the analyses and covariates were carefully selected in order to obtain estimates that are as unbiased as possible for the main predictors. Lastly, all analyses, except those that are only focused on type of health facility and distance were weighted by individual women’s sampling probabilities.