Data sources & study population
The study used secondary data from Zimbabwe’s Demographic Health Surveys (ZDHS) of 2015. The 2015 ZDHS population sample was nationally representative, comprising of more than 11,000 households [30, 31]. The 2015 ZDHS was representative of each of Zimbabwe’s ten provinces: Manicaland, Mashonaland Central, Mashonaland East, Mashonaland West, Matabeleland North, Matabeleland South, Midlands, Masvingo, Harare, and Bulawayo. The 2015 ZDHS used the 2012 sampling frames [30, 31]. The 2015 Demographic Health Survey used a two-stage cluster sampling approach; in the first stage, the samples included 2015 400 Enumeration Areas (EAs), that is, 166 in urban areas and 234 in rural areas. The second stage of sampling included a complete listing of households conducted for each of the selected 400 Enumeration Areas (EAs) in March 2015, respectively [30, 31]. The study population was composed of women of child-bearing age (15–49 years) interviewed in 2015. The sample retained from the 2015 ZDHS before taking into account some observations with missing data on variables of interest was 9955 women. However, after including only observations that had full records on variables of interest the study sample reduced to 4595 women.
Statistical analysis
This study employed 3 statistical analyses namely; logistic regression, Erreygers Normalised concentration indices, and decomposition of the Erreygers Normalised concentration indices. The logistic regression models were used to estimate the likelihood of uptake of maternal health services (SBA,ANC & PNC) among women aged 15 to 49. When using logistic regression, the odds ratios were determined for all independent variables for each category of the independent variable with the exception of the reference category, which was used as a reference category in the analysis. After assessing the association of maternal health services uptake with the demographics variables, we estimated health inequalities in maternal health uptake as well as what was driving the health inequalities in Zimbabwe using the Erreygers Normalised concentration indices. We used the output of the logistic regression in developing and decomposing the Erreygers Normalised concentration indices. The Erreygers Normalised concentration indices are explained in detail under the concentration curves sub-heading.
Outcome variables
Maternal health in this study was measured using 3 outcome variables thus; skilled birth attendance, antenatal care, and postnatal care. Outcome variables were categorized into binary variables: Skilled birth attendance was assigned a value of 1 if a woman reported being attended by a doctor, nurse, or midwife during delivery, and 0 if otherwise. Antenatal care in this study was, defined as mothers who received pregnancy care from skilled health providers (doctors, nurses, and nurse midwives) [31], and represented by 1 if a woman had received at least four ANC visits and 0 for less than four ANC visits. Lastly, as safe motherhood programs recommend that women receive a postnatal health check within 2 days after delivery [31], for this study postnatal care was reported on mothers who had received a postnatal check in the first 2 days after delivery and coded as 1 for mothers who has received postnatal care, and 0 otherwise.
Selection of regressor variables
Socioeconomic factors such as women’s age, women’s education, partner’s education, residence status, household wealth, household head sex, employment status, place of delivery, antenatal care, postnatal care, birth order, distance to the health facility, and media access (radio/television) have been widely reported as key determinants of inequalities in maternal health care uptake [1, 7, 10,11,12,13,14,15,16,17,18,19,20, 23,24,25,26, 32, 33]. This study then used the aforementioned determinants as predictors in the regression models.
Analysis of the association of the predictors with the outcome variables
The study computed binary logistic regressions to predict the dependant variables: skilled birth attendance, antenatal care, and postnatal care. Binary logistic regression is known to be most useful when the dependent variable is a dichotomous [34]. Women’s and partners’ education were both categorized into four groups;0 no education, 1 primary, 2 secondary, and 3 tertiary education. Residence status was categorized into 2 groups and coded as; 0 urban and 1 rural. Birth order was grouped into 4 groups;1st,2nd, 3rd and 4+. Women’s age was grouped into 4 categories namely 15–24, 25–34, 35–44, and 45–49 years.
Socioeconomic status
The wealth index was retained as it was in the Demographic Health Survey [30, 31]. In the ZDHS survey, the household wealth index was calculated by constructing a linear index from asset ownership indicators using principal components analysis to derive weights [30, 31]. In the original survey, the wealth index was constructed by assigning household scores, then ranking each person in the household population by their score. Thereafter, the distribution was divided into five equal categories and each had 20% of the population with economic proxies, such as housing quality, household amenities, consumer durables, and size of landholding [30, 31]. This study then retained the wealth index as recorded in the original survey 5 groups (poorer, poor, middle, richer, richest). This study adopted the household wealth index as a proxy for a household’s economic status.
Concentration curves and indices
The concentration index approach is a standard measure of assessing health inequalities. The indices and curves investigate whether the health inequalities exist in one group or not. However, they do not estimate the magnitude of the health inequalities [35]. This paper used the Erreygers normalized concentration indices [36], to measure the degree of socioeconomic inequalities in utilization of antenatal care, postnatal care, and skilled birth attendance services in Zimbabwe. Among many of the indices that could have been used, we opted to adopt the Erreygers due to its ability to be decomposable.
The concentration index can be computed making use of the ‘covariance’ as shown below:
$$CI=\frac{2}{\hat{y}} COV\ \left({y}_i,{R}_i\right)$$
(1)
Where: yi is the health variable.
ŷ is the mean of yi.
Ri is the fractional rank of the ith individual.
COV denotes the covariance.
Concentration indices can be computed as twice the area between the concentration curve and the line of equality (the 45-degree line) [37]. No existence of health inequality is reflected by a concentration curve lying on the 45° line. The extent of the health inequality is shown by how far the concentration curve lies away from the line of equality (45° line). The further the concentration curve is from the line of equality, the greater the extent of health inequality [35]. Therefore, a true zero value of the Erreygers normalized concentration index indicates no existence of socioeconomic inequalities, while a negative value translates to the disproportionate concentration of socioeconomic inequalities among the poor and a positive value reflects the concentration of socioeconomic inequalities among the rich [9, 38].
Since skilled birth attendance, antenatal care, and postnatal care were cardinal variables, as the differences between health states were comparable, the study adopted the Erreygers normalized index (E(c)). The study opted to use the normalized formulae as, [36, 39] argued that normalization of the health concentration index formula ensured remedying the bounds issue for binary cardinal health variables. The Erreygers normalized index (E(c)) can be expressed as:
$${E}_c=\frac{4\hat{y}}{y^{max}-{y}^{min}} CI$$
(2)
Where ymax - ymin is the range of the health variable, which is ‘one’ in the case of binary variables. Given that both corrected CIs are commonly used in the health literature, the present study focused on the Erreygers normalised index.
Decomposing the Erreygers normalised concentration index
The Erreygers Normalised concentration index is decomposable, so as to compute the contributions of determinants of maternal health indicators [40, 41]. Health inequalities were decomposed into the contributions of various explanatory factors, with each contribution as the product of the elasticity of health. Assuming a linear relationship between individual health (yi) and a set of k explanatory variables yi will be:
$${y}_i=a+\sum_k{\beta}_k{X}_{ki}+{\varepsilon}_i$$
(3)
Wagstaff et al. showed that for any health variables exhibiting a linear relationship with a set of k exploratory variables, the concentration index for the health variable can be decomposed as follows:
$$CI=\sum_k\left(\frac{\beta_k{\dot{x}}_k}{\hat{y}}\right){CI}_k+\frac{GCI_{\varepsilon }}{\hat{y}}$$
(4)
Where: βk is the partial.
ŷ is the mean of the health variable (SBA or ANC or PNC).
ẋk is the mean of ẋk.
CIk denotes the concentration index of xk against Wealth index/Socioeconomic Status.
GCɛ is the generalized concentration for the error term.
Equation (4) can be modified as shown below to decompose the Erreygers concentration index [42]
$${E}_c=4\left[\sum_k\left({\beta}_k{\dot{x}}_k\right){CI}_k+{GCI}_{\varepsilon}\right]$$
(5)