Data
Women’s questionnaire data from the 2015–16 Malawi Demographic and Health Survey (MDHS) were used for the study. The 2015–16 MDHS collected a wide range of information including demographic indicators, fertility and mortality measures, family planning knowledge and use, immunization coverage, maternity care, infant and young child feeding, nutritional status, HIV and more [8]. The survey was conducted with a two-stage stratified cluster sampling design [8]. Each of the 28 administrative districts in Malawi was stratified into urban and rural strata [8]. From each strata, a sample of standard enumeration areas, which consist of about 235 households on average, were selected for household listing serving as a sampling frame [8]. In the second stage, 30 households were selected from each urban household listing or cluster and 33 households were selected from each rural household listing or cluster [8]. In the selected households, women’s questionnaires were administered to women between the ages of 15 and 49 who were either residents there or were visitors from the night before the survey [8]. In the current study, women who had a facility delivery in the last five years preceding the survey were considered for analysis and the focus is on the most recent singleton birth in a facility.
Variables
Three categorical endogenous (outcome) variables were examined in the study: whether or not mothers received a postnatal check between birth and discharge from the facility for their most recent birth in the past 5 years, whether or not their most recent newborns received a postnatal check between birth and discharge from the facility and whether or not women delivered by cesarean section. For all three endogenous variables, “1” indicated receipt of services and “0” indicated otherwise. There were no women who responded “don’t know” for maternal PNC between birth and discharge. For newborn PNC between birth and discharge, about 1% of the women responded “don’t know.” Information about how long after delivery the first check took place and the type of provider who checked on the health of the mother and newborn were not incorporated into creating the final outcome measures. This is because, as an exploratory research study, we wanted to understand and capture all postnatal checks that occur between birth and facility discharge and not just those that occur within the first 24 h. Among all women who had their most recent birth in the 5 years prior to the survey, about 7% of them had missing data for maternal and newborn PNC between birth and discharge.
Winship and Mare mention in their important work on structural equations for discrete data [19] that binary indicators can represent one of two ideas: an indicator that measures a discrete event or an indicator that serves as a proxy for an unobserved underlying continuous variable [19]. In this study, delivery by cesarean section is treated as a proxy variable for some underlying continuous phenomenon. The final decision to perform cesarean section is likely based on whether pregnant women, who are all on an unobserved continuum of complication risk, display characteristics that exceed providers’ threshold levels for risk of complication. Hence, this study acknowledges that the underlying continuous variable exists for cesarean section, albeit unobserved.
The main predictor of interest in the study was type of health facility where women delivered. It was coded to have four categories: [1] government hospital [2]; government health center, government health post and other public sector facilities (not specified in the dataset) [3]; private hospital and Christian Health Association of Malawi (CHAM) hospitals; and [4] CHAM health centers, Banja La Mtsogolo clinics (BLM) and other private sector facilities (not specified in the dataset). Health centers, health posts and other unspecified facilities were grouped together because only a small number of women delivered in health posts and other unspecified facilities (around 1.44%) and deliveries are typically done in health centers or hospitals in Malawi [20]. All of these facility categories represent the major health service providers in Malawi [20]. However, there are a couple of differences between health facilities based on their type and affiliation (public and private). Government-owned facilities provide services free of charge while privately-owned facilities charge a fee [20]. However, the possibility cannot be ruled out that even at government-owned facilities, health services may not be free at the point of use in some parts of the country. Health facilities of various types and affiliations also have very different resources for service readiness including basic amenities, equipment, infection control, diagnostic capacity, essential medicines, quality assurance, client feedback and provider training [20]. The main predictor variable was meant to capture these differences in general service readiness as well as potential differences in PNC practices by types and affiliations of health facilities.
Other exogenous variables included in the model were women’s age at most recent birth, number of antenatal visits, women’s education, household wealth, parity, newborn size, region of the country and residence. Women’s age at most recent birth was included as a continuous variable. Number of antenatal visits was coded as having had “less than 4 visits” or “4 or more visits”. Women’s education was coded to be either “no education”, “primary education” or “secondary education”. Household wealth was a quintile variable constructed by the Demographic and Health Surveys Program (DHS) using principal components analysis [8] and it was coded as either “poorest”, “poorer”, “middle”, “richer” or “richest”. Parity of the birth was coded as either “1”, “2 – 3” or “4 or more”. Newborn size based on women’s recall was coded as either “very large”, “larger than average”, “average”, “smaller than average” or “very small”. Region of the country was coded as either “northern”, “central” or “southern”. Residence was coded as either “urban” or “rural”.
In this study, receiving a postnatal check between birth and facility discharge was hypothesized to be more a function of the type of facility where women delivered and other “noticeable” maternal and newborn characteristics that could further attract attention by the providers in the respective facilities. However, sociodemographic variables such as women’s age at the time of birth and education were also included in the model to test for their effects.
Descriptive analyses of the study variables are presented in Tables 1, 2 and 3. Table 1 presents the coverage of maternal and newborn postnatal health checks between birth and facility discharge among women who had their most recent singleton birth in facilities in the 5 years prior to the survey. Table 2 presents descriptive summary information of the study variables for women who had their most recent singleton birth in facilities in the 5 years prior to the survey. Table 3 presents the percentages of women and newborns receiving postnatal health checks between birth and discharge by type of delivering health facility.
Table 1 Coverage of maternal and newborn postnatal health checks between birth and facility discharge, MDHS 2015–16
Table 2 Descriptive information about the study variables for women who had their most recent singleton birth in facilities in the past 5 years, MDHS 2015–16
Table 3 Type of health facility where women delivered stratified by maternal and newborn postnatal health check between birth and facility discharge in Malawi, MDHS 2015–16 Analysis
The model hypothesizes that: [1] cesarean section, age of the mother at most recent birth, education of the mother, parity, number of antenatal visits, type of health facility where women delivered, urban/rural residence and region of the country influence maternal postnatal health check between birth and discharge [2]; cesarean section, newborn size, education of the mother, number of antenatal visits, type of health facility where women delivered, urban/rural residence and region of the country influence newborn postnatal health check between birth and discharge; and finally, [3] all of the exogenous variables in the model, except education, influence delivery by cesarean section.
Because cesarean section is in the mediated pathway for most of the exogenous variables in the model, indirect effects of these variables on maternal and newborn postnatal health checks can be calculated. In the case where exogenous variables predict the latent continuous variable underlying the binary mediator and the same latent continuous variable is used to predict maternal and newborn PNC, the calculation of indirect effects just involves the product of two coefficients: the coefficient of an exogenous variable in predicting cesarean section and the coefficient of cesarean section in predicting maternal or newborn PNC [19]. This corresponds to the model 1 specification in the article by Winship and Mare [19]. The total effects are the sum of the direct effects and the indirect effects [19].
The indirect effects through cesarean section and the total effects were only calculated and interpreted for type of health facility where women delivered. This is because the primary interest of the study is to distinguish the direct effects of the delivering health facilities from their indirect effects through cesarean section. According to the Malawi Service Provision Assessment 2013–14 report, cesarean section is only done in hospitals [20]. However, the indirect effects and the total effects were also calculated for government and CHAM health centers because some women in the sample responded that they delivered by cesarean section in these lower-level facilities as well. The effects of other exogenous variables will indicate whether observable maternal and newborn characteristics as well as the socioeconomic background factors influence postnatal health check at the facilities outside of the mediated pathway through cesarean section.
The hypothesized model also shows that all of the exogenous variables are correlated in some way. It does not specify what the directions of the associations are among the exogenous variables. This is because those relationships are not of interest in the study. However, acknowledging that these variables are correlated is important for model specification and fit. Lastly, the residuals of maternal postnatal health check between birth and discharge and newborn postnatal health check between birth and discharge are hypothesized to correlate in the model. Due to this correlation, the model is not “fully recursive”. However, it is still “recursive” because there are no feedback relationships present in the model.
Checks for model identification and specification
There are many rules of identification reported in the literature to help researchers make sure that their models are identified. One of the rules of identification that applies to recursive models with correlated errors is that no variable influences another variable with the error terms of each being correlated [21]. According to Brito and Pearl, this is a sufficient condition of identification for recursive models with correlated errors [21]. The illustrated path diagram in Fig. 1 shows that this rule applies and is sufficient for identification.
Several model fit indices with the full information estimation method showed that the model specification fit the data well. Specifically, six different model fit indices were considered to assess if the model specification was appropriately fitting the data. These indices were the overall chi-square test, Tucker-Lewis index, incremental fit index, relative noncentrality index, root mean square error of approximation and the Bayesian information criterion [22, 23]. All of the six model fit indices suggested that the current model specification is consistent with the data and that no other modifications need to be made based solely on model fit. The summary of the model fit indices can be found in Table 4.
Table 4 Tests for model specification Table 5 presents probit regression coefficients from the simultaneous equation model that were estimated using the “lavaan” package in R studio. The probit regression coefficients were adjusted for cluster sampling. The signs of the probit regression coefficients can be interpreted as positive or negative influences on the binary outcomes themselves and on their underlying continuous variables. Lastly, because all endogenous variables in the model are categorical, diagonally-weighted least squares were used.
Table 5 The total, direct and indirect effects of key predictors on maternal and newborn postnatal health checks between birth and facility discharge in Malawi, MDHS 2015–16