Examining changes in maternal and child health inequalities in Ethiopia

Table 1 presents the six health status and five health service indicators analyzed in this study, chosen based on their relevance to the health MDGs and on the availability of data. For health status, there are three child undernutrition and three mortality indicators: stunting, wasting, underweight, neonatal mortality rate (NMR), infant mortality rate (IMR), and under-5 mortality rate (U5MR). The service indicators cover child immunization and maternal health services; they are measles vaccination, full immunization, prevalence of modern contraceptive use by married women, four or more antenatal care visits from a skilled professional (ANC4+), and delivery assistance from a skilled birth attendant (SBA). Additional file 1: Table 1a and b present details of the indicators and terms as used in the analysis.

The data are from four Ethiopia Demographic and Health Surveys (DHSs), in 2000, 2005, 2011, and 2014. These population-based surveys target mainly women of child-bearing age (15 to 49) but also collect some data about the household and some from men in the same age range. The main survey concerns are fertility, family planning, infant and child mortality, maternal and child health, and nutrition. Since 2000, they have been conducted about every 5 years. The first three surveys each sampled about 15,000 women from about the same number of households. The sample for the 2014 survey was about half that of the previous rounds. Much of the data used in this study comes from the women’s questionnaire, which compiles a comprehensive birth history for each woman, from antenatal care and delivery attendance through child survival and vaccination. The questionnaires used in the different surveys are standard and comparable. The data provide nationally representative information on the variables we selected for this study.

The primary qualifying variable is wealth ranking. We looked at socioeconomic inequalities in health by wealth ranking between the worse-off (bottom 40%) and the better-off (top 60%) and between the poorest (1st quintile) and the richest (5th quintile). These are computed from the household wealth index available with the data [10].

To evaluate child undernutrition, we computed anthropometric indicators based on the WHO 2006 growth standards: We calculated height-for-age, height-for-weight, and weight-for-age z-scores and then stunting, wasting, and underweight levels for children aged 0 to 59 months. Child mortality rates (IMR, NMR, and U5MR) are calculated using the standard DHS methodology, using data on all child deaths in the 5-year period preceding each survey [8]. We calculated the prevalence of modern contraception use by currently married women, antenatal care (most recent birth), and skilled birth attendance (all births in last 5 years) from the questionnaire administered to all women aged 15–49 years in the household.

For a more complete picture, we considered a combination of the approaches often used in inequality studies because each approach has some limitations that can lead to different conclusions [11, 12]. Our analysis started with how absolute and relative inequalities between the poor and the rich have widened or narrowed over time. Then, we looked at the concentration curve and concentration index, which capture inequality across a continuous spectrum of wealth, and what they reveal about the changing pattern of inequality over time. Finally, we looked at the decomposition of the concentration indexes to see the changing role of various demographic and socio-economic factors in the observed wealth-based inequality in health services and outcomes.

We computed absolute inequalities from rate differences between the poor and the rich, defined both as the bottom 40% versus the top 60% and as the poorest quintile versus the richest.

The difference-in-differences comparison is as follows: Let I_xt be the value of the indicator for group x (either r = rich or p = poor) in time period t (either t = 0 first survey or t = T latest survey). We perform an F-test of the hypothesis:

Second, instead of the difference between the value of the indicator for rich and poor, the ratio of the values was used. This emphasizes the difference between indicators where both groups have very low values. The hypothesis tested was:

The analysis of inequality based on absolute and relative gaps was limited to binary distinctions: rich vs. poor. It was thus somewhat sensitive to the definition of the binary distinction and also did not allow for analysis of inequality across the whole range of wealth outcomes. For example, the comparison of the bottom quintile to top quintile entirely ignored any changes in the health indicator for the middle three-fifths of the population. If we see a decrease in inequality between the bottom 40% and the top 60%, we do not know whether this was due to improvements for the poorest of the poor or for those closer to the middle of the income distribution. Therefore, we used concentration curves to illustrate the movement of wealth inequalities in health across the entire range of wealth between the earliest and latest surveys. We also used concentration indices (C), which quantify the degree of inequality in this analysis, and observed how they changed between the earliest and latest surveys [13].

We would expect the concentration curve for positive health indicators (immunization, maternal health services) to lie below the line of equity (poorer households account for a disproportionately low number of fully vaccinated children or attended births). This is shown in the measles examples above, where both points plotted lie below the 45-degree line. Conversely, we would expect the concentration curve for a negative health indicator (child mortality, malnutrition) to lie above the line of equality (poorer households account for a disproportionately high number of child deaths). We expect that the bottom 5% of children, ranked by wealth, account for more than 5% of underweight children and thus the point is above the 45-degree line. A concentration curve that moves closer to the line of equality over time indicates decreasing inequality.

The concentration curve does more than offer a nice visual summary of wealth-based inequality in an indicator; it is also useful to quantify the degree of inequality revealed. The concentration index (C) quantifies the degree of inequality—twice the area between the concentration curve and the line of equality—which is analogous to how the Gini coefficient quantifies the degree of inequality in a Lorenz curve. We calculated C, for each indicator as follows:where h is the health variable, μ is the mean, and r is the fractional rank of the individual in the wealth index.

When the outcome variable is binary, the concentration index has some questionable properties, especially comparing populations that have significantly different means. In particular, because it is mathematically bound between μ − 1 and 1- μ (where μ is the mean of the binary indicator), it tends to naturally fall (in absolute value) as the value of μ increases. If only the richest 10% have access to a health service in the base year, the concentration index would be 0.9. If the richest 90% have access to the health service in a subsequent year, the concentration index would be 0.1. It is therefore debatable whether that should be considered a large decrease in inequality [14, 15]. For binary indicators, we calculated as alternative indicators the Wagstaff concentration index [W = 2/ μ (1- μ) COV(h, r)] and the Erryegers concentration index, [E = 8 COV(h, r)], where h is a health outcome indicator and r is the fractional rank of the individual in the wealth index [14].

Our analysis also included decomposition of the concentration index measure of wealth-related inequalities in selected health outcomes and services. This allowed us to see how differences in, e.g., family size, women’s education, and access to safe water contribute to the observed wealth-related inequalities in health outcomes or health services, and how these patterns are changing over time. For example, the observed inequality in childhood vaccination rates might be explained entirely by differences in the mother’s education; decomposition of the concentration index would reveal that.

Decomposition of the concentration index is based on the algebraic transformation, which for a linear model of a health indicatorallows the concentration index to be written aswhere μ is the mean of y, \( {\overline{x}}_k \) is the mean of x _k , C _k is the concentration index for x _k , and CG _ε is the generalized concentration index for the error term [15].

The dependent variables of interest in this study are binary indicators and thus best modeled using a non-linear model such as probit. Following previous work [15, 16]), the linear approximation is given by the following specification,where h is the health variable of interest as defined earlier, x _j are the independent variables, α ^m is the constant term, \( {\beta}_j^m \) are the partial effects of each variable treated as fixed parameters and evaluated as sample means, and u is the error term. The decomposed concentration index (C) for a health outcome h _i is thereforewhere C _j are the concentration indexes for x _j , μ is the mean of the health variables h, \( {\overline{x}}_j \) is the mean of x _j , and GC _u /μ is the residual component that captures inequality that is not explained by systematic variation in the regressors by income.

Table 2 presents the profile and trends of selected MNCH outcomes and services. As the table makes clear, Ethiopia’s health service delivery is among the least developed in low-income countries—modern services reach only a small fraction of the population. For example, the results of the 2014 survey show that nationally modern contraceptive use is 45%. The situation is much worse for ANC visits, where services coverage was 25%, and SBAs, where coverage was 16%. Similarly, full immunization coverage in 2011 was 25%—among the lowest in similar countries in Sub-Saharan Africa. Over the last two decades, however, there has been considerable improvement in MNCH outcomes in Ethiopia, though child undernutrition and mortality rates are still high and coverage of maternal and health services is low. This holds true for all the health status and health service indicators analyzed in this study (Table 2): There is a consistent decline in ill health (undernutrition and mortality) and an increase in health services coverage (immunizations and maternal health services).

To look at the trends by wealth status, we disaggregated the progress made. For each indicator, each line represents the value of the indicator for one group over time, with the 95% confidence interval around each value. Figures 1 to 4 present the results for child nutrition, child mortality, immunization, and maternal health services. As expected, as a general pattern the lines for adverse outcomes slope downward for all groups and the lines for health services slope upward, showing MNCH is improving at all wealth levels; the improvements in average national figures, however, do not hide worsening results for the poor. The distances between the curves and the slope of each curve in each graph show differing initial and final inequality for the indicators.

In Fig. 1, the lines in all graphs move downward basically in parallel, with perhaps a slight widening of the gap between the different wealth groups. In the graph for stunting, we see that inequality is driven by the difference between the richest quintile and the rest, and that this difference is increasing slightly.

Figure 2 shows the trends for child mortality indicators. The biggest improvement has been in the U5MR, with modest improvements in the IMR and essentially no change in neonatal mortality. No clear pattern of wealth-related inequality is observable. The confidence intervals often overlap, showing no significant difference between different groups, and the lines crisscross each other, with poorer households sometimes seeming to have lower child mortality (which might be true and reflect differences in practices like breastfeeding). It is not clear whether it is actually true that there is no systematic difference in child mortality by household wealth level or whether the figures are obscured by the data collection strategy (which only counts children whose mother is alive) or cultural traditions (unwillingness to speak about an infant who died very young).

In Fig. 3, the trends show that immunization services have expanded for all wealth groups, with the lines moving roughly in parallel. Again, the biggest gap is between the richest quintile and all the others.

As Fig. 4 shows, maternal health services have also improved for all groups, but there is a widening of the gap between rich and poor, with a greater increase in the use of maternal health services by richer households. This trend is less pronounced for contraception use, and in fact there seems to be a slight closing of the gap in the last year of data. Again, we see that wealth-based inequality is primarily driven by the difference between the wealthiest quintile and the rest, especially for ANC and SBA.

The figures in the previous section give us a good indication of the general trends: fairly consistent improvements across all wealth groups, with some modest narrowing or widening of gaps depending on the indicator and the exact breakdown of households by wealth. To quantify and test the statistical significance of these observations, we report the results from eqs. 1 and 2 in Tables 3 and 4. The results in Table 3 compare the bottom 40% of households and the top 60%. Those in Table 4 compare the bottom quintile to the top quintile. In each table, the first section gives the value of the indicator for “poor” and “rich” households in both the first and the last year. The second shows the absolute inequality (the difference between the values of the indicator for the two groups) in the earliest and the latest survey and the p-value for whether absolute inequality changed significantly during that period. The third section shows the relative inequality (the ratio of the indicators) and the p-value for whether it has changed significantly. A positive difference or ratio greater than 1 for an ill-health outcome (child undernutrition or child mortality) shows a pro-rich inequality—child undernutrition and mortality rates were lower for children from better-off households. Likewise, a negative difference or ratio less than 1 in any of the immunization and maternal health service indicators (good health service utilization) implies pro-rich inequality. These socioeconomic differences are to be expected; our interest is in whether the differences are increasing or decreasing.

The results (Table 3) point to a widening of absolute pro-rich inequality in child nutritional outcomes between the poor (bottom 40%) and the rich (top 60%)—an inequality observed in all three child nutrition status indicators but more significant for stunting and underweight. For the poor, child stunting in the earliest survey was higher by about 4.2 percentage points and in the latest the difference rises to 10 percentage points. Similarly, the gap in underweight went up from 5 to 8 percentage points. In both cases, these changes are statistically significant, which implies that pro-rich inequality is widening. The same conclusions can be drawn when looking at relative inequalities: pro-rich inequality widened significantly during the period considered, at least for stunting and underweight.

When we restrict our analysis to the poorest versus the richest quintile (Table 4), the trend recurs: an ever-increasing gap between the rich and the poor in terms of child nutrition. These changes are less statistically significant, however, and only the relative inequality in stunting between the richest and the poorest changes significantly.

The results for infant and child mortality appear to show that mortality rates were in fact lower for poorer households in the first survey (the values for the difference between rich and poor are negative in all three indicators and in both ways of defining rich and poor). Then, by the final survey, the pro-poor gap appears to have disappeared for infant mortality and is reversed for under-5 mortality, with both rates higher for poorer households. While these figures may represent real changes in mortality patterns, it should be noted that the standard errors on the mortality rates for different subgroups are quite high, and that none of the changes over time are statistically significant.

For all services except measles vaccination, in the first survey, utilization by poorer households was very low; thus while the difference between the rates increased (significantly in the case of maternal health services), which suggests widening inequality, the ratio of the rates decreased, suggesting that inequality had narrowed. For measles vaccination, both the variances and the ratio suggest decreasing inequality, although only the changes in the ratio are significant.

The concentration curves in Fig. 5 show child malnutrition as fairly even distributed across income categories, with the curves lying close to the line of equality, but the curves moving away from the line indicate gradually increasing inequality. Overall, the results presented here are aligned with the findings already presented: the poor did not benefit as much from the improvements in child nutritional status.

Figure 6 shows the concentration curves for child mortality indicators. Here the curves overlap and some cross the line of equity. This reflects the fact that there are no obvious patterns or trends when mortality rates are broken down by household wealth level. Statistical tests confirm that for these indicators there is no significant change in wealth-based inequality.

Figs. 7 and 8 illustrate the movement of wealth-related inequalities in child immunizations and maternal health services. Here there is clearly significant inequality, with the concentration curves lying far below the line of equality. However, we also see substantial movement over time as the curves move closer to the line of equality, reflecting less wealth-based inequality in use of services. The figure for antenatal care shows clearly that improvements are concentrated among richer households— middle-ranked households are catching up to the wealthiest households, but the poorest households are not catching up to those in the middle.

For each indicator, Table 5 provides information on the concentration index and, for binary indicators, the alternative Wagstaff and Erryegers concentration indexes for the earliest and latest years. Note that, as expected, the index values are negative for malnutrition indicators—richer households have less malnutrition—and positive for MNCH services—richer households use health services more. The values of the concentration indexes for mortality can be either positive or negative; there is no clear trend. The absolute value of the concentration index can be taken as a measure of the inequality present. The tables also show the result of simple t-tests of whether the concentration index value changed significantly between the first and the last survey.

The concentration indexes for all three undernutrition indicators show pro-rich inequality heightening; in absolute value the concentration index more than doubled for all three, although from initially low levels of inequality. The changes are significant for stunting and underweight. Using the Wagstaff or the Erryegers concentration index yields the same result but with somewhat less significance. On the other hand, like the concentration curves (Fig. 5), there is no clear pattern in the indexes for mortality indicators (Table 5). Table 5 also shows mixed progress in decreasing inequality in the use of MNCH services; over time concentration index values decrease, but except for measles vaccination and contraceptive use the changes are not significant when the Wagstaff concentration index is used. However, the Erryegers concentration index leads to the opposite conclusions: it shows increasing inequality in four of the five indicators and is significant for maternal health services. This again reflects the pattern we see, particularly in maternal health services: having expanded rapidly from low levels tends to lead to a decrease in the standard concentration index that is not robust to using the Erryegers index.

Despite progress then, as both the concentration curves and indexes illustrate, inequalities are still substantial in the case of ANC4+ and SBA.

We used three different methods to assess changes over time in wealth-related inequality in indicators of health status and health service utilization: rate differences, rate ratios, and concentration indices. Table 6 summarizes the results.

Although child malnutrition has been reduced for all income groups, wealth-based inequality worsened over the period studied. This pattern is consistent for all three indicators and all methods of analysis.

The DHS data show no clear relationship between household wealth and child mortality. In earlier years, there may have been inequality in favor of poorer households, or possibly an inverse-U shaped pattern, with households in the middle of the wealth distribution having the highest child mortality. That pattern might be explained by cultural factors, or it might be an artifact of the data collection strategy. The concentration index suggests there is some evidence that inequality in child mortality has “improved,” moving from a pro-poor bias to a smaller pro-poor bias or a small pro-rich bias.

Inequality has narrowed slightly in child vaccination services, particularly for measles vaccinations. For maternal health services, the results seem contradictory: The absolute difference in maternal health service utilization by the rich and the poor has increased, but initially service use by poor households was extremely low and the gains they have made are substantial. Thus, the rate ratio, the concentration curves, and two of the three concentration indexes analyzed all rate the changes as a decrease in inequality.

The analyses so far discussed demonstrate that, despite improvements, there are still considerable inequalities in some MNCH outcomes and services. In this section, we examine what may be contributing to the wealth-related inequalities in certain indicators.

Tables 7 to 9 present the decomposition of concentration indexes for selected indicators. Various controls that may be related to both wealth status and the value of the indicator are included to see how much of the inequality can be attributed to factors like the mother’s education or adequate sanitation facilities.

Table 7 summarizes the results for stunting: in each year, even when controlling for other variables, the largest contribution comes from the wealth index. The next largest comes from mother’s education. Comparing the progression of stunting between 2000 and 2014, we see that the increased inequality comes only from the contribution of the wealth index itself, not from any other factors. The decomposition for 2014 has a large positive residual component, however, suggesting that other factors not captured in the decomposition are in fact offsetting the rise in wealth-related inequality in stunting.

Table 8 shows the decomposition for measles vaccination (Panel a) and full immunization (Panel b). Comparing results in 2000 and 2014, we see that for measles vaccination the contribution of the wealth index has gone from being the largest to being one of the smallest but when all immunizations are considered it was still very important. This suggests that the pure wealth component of inequality in measles vaccinations almost disappeared when coverage was expanded, as it was for measles vaccination but not full immunization. In both cases, the remaining inequalities can be explained almost entirely by differences in the education of the mother and the father.

In Table 9 Panels a through c present the decomposition results for maternal health services. A serious limitation is that in all specifications the residual component is large. With that in mind, however, it is still possible to discern some common trends: there is only a slight decline in inequalities over the period, the prevailing inequalities are still high, and the wealth index, education, and residence of the user all make important contributions to the differences in utilization of these services.