Data
Administrative data on drug stock-outs were obtained from Zimbabwe’s MoHCC, whose Pharmacy Services Department collects the data. Facility authorities submit their stock information quarterly to the MoHCC, and the Ministry collates the dataset into a single database. The data contain information on SP availability at each facility in all the districts. The information is collected to monitor drug availability at the health facilities, which includes stock in hand, drugs used, and stock-outs. The data used in the current paper were collected quarterly per facility and district from 2011 to 2015.
The DHS dataset is collected from developing countries every five years and has been conducted in Zimbabwe since 1988. The dataset contains information on demographic and socioeconomic variables, healthcare services utilisation, maternal and child health outcomes, and maternal and child mortality. A total of 9,955 women were interviewed, yielding a 96.2% response rate. However, we only included information on women who gave birth between 2011 and 2015, for which stock-out data were available. Data on 2010 health outcomes were not included, as there were no data on stock-outs for the same year. After excluding individuals with missing observations, the data analysis was based on 3,432 observations. The survey data were prone to missing observations due to recall and non-response bias. Robustness checks were performed to examine whether the missing observations in the analysis resulted in biased estimates and the results are part of the supplementary information.
We first added facility GPS coordinates to the stock-out data by merging master facility data, which contained information on facility location and stock-out data to link facility and household survey data. To combine the drug stock-out data and the ZDHS data, we then used the facility and survey cluster GPS coordinates. A total of 296 facilities were merged into the clusters in the ZDHS in the analysis. The ZDHS cluster coordinates and the facility coordinates were merged based on the assumption that people in the same cluster use the same facility. The ZDHS GPS coordinates were displaced 2 km in urban areas and 5 km to 10 km in rural areas. The displacement of coordinates was done carefully, to ensure that clusters did not overlap with other administrative areas (see [22]). To correct the impact of coordinate displacement, we used a 10 km buffer zone by considering only the facilities within a 10 km radius. We combined the 2015 ZDHS and drug stock-out data in the analysis. We used 2015 cross-sectional data, but child’s year of birth provided a retrospective time dimension within the survey. Drug stock-out data, were, however are collected quarterly for each facility. The time dimension in the stock-out data was linked to the birth timing in the DHS cross-section.
Analysis
We used the Ordinary Least Square (OLS) model to examine the relationship between malaria prophylaxis stock-outs and birth- and maternal outcomes in Zimbabwe. We also included regional and birth-year fixed effects in the analysis to capture regional and birth-year variations. Clustered standard errors are used to account for common variations inside survey clusters. Therefore, we measured associations, and not causal relationships – specifically between stock-outs and birthweight for neonates and maternal anaemia. Given that haemoglobin levels are only monitored during data collection in 2015 and not earlier during pregnancy, when women are more vulnerable to malaria, the paper focused on mothers who had given birth or were pregnant in 2015. This is because these women were more likely to be affected by malaria infections than women who were not pregnant at the time of the survey. However, one major weakness of this analysis is that haemoglobin levels might have adapted to post-birth experiences in potentially biased ways. The results on the association between malaria prophylaxis stock-outs and maternal anaemia are reported in Tables B and C of the supplementary information and are presented as explanatory rather than definitive. The model for malaria prophylaxis stock-outs relationship with health outcomes was specified as follows:
$${y}_{ict}={alpha }_{1}{stockout}_{ct}+{alpha }_{2}{X}_{ict}+{{theta }_{i}+phi }_{t}+{varepsilon }_{ict}$$
(1)
where ({y}_{ict}) was the birthweight or haemoglobin level of individual i in cluster c at time t, ({stockout}_{ct}) was the SP stock-out at cluster c (where the cluster was equal to the facility) in period t. ({X}_{ict}) was the matrix of other control variables in the model, which were education, parity, preterm delivery, wealth index, BMI, currently pregnant, work status, birth interval, interacted ANC and stock-outs, interacted IPTp districts and stock-outs and HIV status, ({alpha }_{2}) was the vector of parameters, ({varepsilon }_{ict}) represented the error term, ({theta }_{i}) represented regional fixed effects, and ({phi }_{t}) represented birth year fixed effects.
After this analysis, we used the recentered influence functional (RIF) unconditional quantile regression model to estimate the association between malaria prophylaxis stock-outs and health outcomes along different quantiles of the birthweight distribution. The unconditional quantile regression model shows the marginal effects of explanatory variables on the unconditional quantile of the dependent variable [15, 35]. RIFs of the unconditional quantile provide a robust analysis of every unconditional quantile [35], useful for policy implementation [1]. According to Firpo et al. [15], the model is simple and easy and can be used for other distributional statistics like the Gini coefficient and conditional quantile. The unconditional quantile regression model focuses on unconditional quantiles when independent regressors are present, which differs from the conditional quantile used in the presence of endogenous regressors [15]. In this regard, conditional quantile regression shows heterogeneity in parameters that characterise the relationship between conditional quantiles of dependent variables and independent variables [1]. The effects of independent variables in unconditional quantile regression are a weighted average of conditional quantile regression. The RIF unconditional quantile regression was specified as follows:
$$RIFleft({y}_{i},{q}_{y}left(pright)right)=alpha +{beta }_{j}{X}_{i}+{varepsilon }_{i}$$
(2)
where y was the dependent variable, (alpha) represented the constant, and ({beta }_{j}) represented the unconditional quantile partial effect of changes in X in the model. ({X}_{i}) showed the independent variables used in the paper including interacted variables, (p) showed the quantiles, and ({varepsilon }_{i}) represented a normally distributed error term. We used bootstrapped standard errors. Although the RIF approach allows for estimation of the covariates across the outcome distribution in the population (rather than conditional on covariates), it may limit generalisability in contexts where the underlying distribution of covariates differs substantially from the study sample.
Description of variables
The independent variables used in this analysis are birth interval, parity, preterm delivery, stock-out index, wealth index, work status, education, geographical location, BMI, currently pregnant and HIV status. These variables were selected from the prior literature.
Work status was a dummy variable denoted by working (for women with a paying job or a business) and not working women. Working women were expected to have better health outcomes than non-working while the geographical location was a binary variable denoted by rural if the location is one and urban if the location is zero. The wealth index was another variable measured in ZDHS using principal component analysis from the household’s assets. Women from wealthy families tend to have enhanced health outcomes compared to those from less wealthy households.
Another variable was education, which depicted the number of years in which the highest qualification was acquired, and educated women were more likely to have improved health outcomes than their uneducated counterparts. Preterm delivery also affects birth and maternal outcomes. Preterm delivery was defined as birth before the 37th week of pregnancy, children born before 37 weeks are preterm represented by one in the analysis and zero if a child is born after 37 weeks of pregnancy. Preterm neonates tend to have lighter birth weights than the non-preterm neonates. Parity was another variable representing the number of children that the woman has ever given birth to, either stillborn or born alive. On the other hand, birth interval shows the spacing between pregnancies by the mother which affects the health of the children and it was represented by the number of children a woman gave birth to within a year in this paper. Women with short birth intervals were more likely to have compromised birth and maternal outcomes.
While there are many causes of anaemia in pregnant women, including iron deficiency and genetic factors, malaria is considered one of the major causes [46]. The WHO [43] defines anaemia as a condition where the number of red blood cells is below the recommended level. Haemoglobin level is used as a measure of maternal anaemia. Birthweight was used to measure the association of neonatal health outcomes with malaria prophylaxis stock-outs. According to the WHO (2019c), babies weighing more than 0.5 kg and less than 2.5 kg are considered to have low birthweight, which is undesirable for a child’s subsequent growth trajectory. Given that birthweight can either be affected directly via placental malaria or indirectly via maternal anaemia, we assumed that we detect the direct effects of malaria on birthweight, as maternal anaemia is mostly measured after pregnancy in the ZDHS. In addition, we created the stock-out index by calculating the proportion of stock-out days per quarter per facility, and then averaged these proportion of days drugs are stocked out at a facility for two quarters over time to get stock-out for the two final trimesters of pregnancy (see supplementary information A.2). Drug stock-outs increase the likelihood of receiving no or fewer SP doses than recommended, increasing the probability of malaria infections [14, 25]. Therefore, SP stock-outs were expected to be negatively associated with birthweight and maternal haemoglobin levels.
