Factors contributing to household wealth inequality in under-five deaths in low- and middle-income countries: decomposition analysis

The routinely collected Demographic and Health Surveys (DHS) data was used for this analysis. Every five years, the DHS is conducted in each of the participating LMICs. The ICF Macro, the USA in conjunction with the designated organizations such as Statistics boards, Universities, Ministry of Health, etc. in the participating countries collect the data. The surveys nationally representative, population-based and cross-sectional. We merged the DHS conducted between 2010 and 2018 and was released on September 10th, 2020. This study comprised a total of 59 LMICs that matched the inclusion criteria. Data from 856,987 U5M from 66,495 neighborhoods in 59 LMICs was used in the analysis.

In each of the countries where the surveys were conducted, similar clustered multi-stage sampling methodologies were utilised. The sampling frames were mostly based on each country's most recent census figures. The number of levels and their stratification are determined by each country's current geographical and administrative structures. In some of these countries, the multi-stage sampling used regions/states/divisions as the first stage, districts as the second level, and clusters as the final stage. Clusters are defined in the same way in all countries and are also referred to as primary sample units (PSU) [15, 16]. At the last stage, the households were selected from the PSUs.

As a result of uneven population sizes among states/regions/districts within a country, DHS generated and provided sampling weights with the data for each country. The sampling weights were calculated using a multi-stage sampling technique to ensure that the estimates from the samples were representative of the general population. Each country employed a similar set of standardized questionnaires, research protocols, interviewer training procedures, supervision, and execution. The entire specifics of the sample procedures, as well as other data has been reported elsewhere [17].

The dependent variable is U5D, which is defined as deaths in the first five years of life among live births [18‐20]. It is defined as death occurring between the ages of 0 and 59 months after birth [21]. To ensure the study's accuracy, correctness, and completeness, women were questioned if they had given birth to a child within the previous five years. Those who said yes were then asked to recall the dates of their children's births, followed by a question about whether each of those children was living or deceased at the time of the interview. U5D was calculated using the dates of death or ages at death of the deceased children. As a result, U5D was a binary variable: Dead or Alive before 5th birthday.

In this study, descriptive and inferential statistics were used. The country, regions, U5Ds, and other significant features of the children by U5D was depicted using basic descriptive statistics such as maps, graphs, tables, and proportions. Table 1 shows the results of tests of equality in proportions of U5Ds among children from poor and non-poor homes in each country and region. The distribution of the background characteristics of the children by the prevalence of U5Ds among children from poor and non-poor households was reported in Table 2. The spatial distribution of under-five deaths per 1000 live births among children in poor and non-poor households are shown in Fig. 1(a) and (b) respectively. The maps were built in Microsoft Projects 2020. Also, to further examine household wealth inequality in U5Ds, absolute and relative measures of inequality recommended in the WHO Health Equity Assessment Toolkit Plus (HEAT Plus) were utilised [36]. These measures include Difference (D), Ratio (R), Population Attributable Fraction (PAF) and Population Attributable Risk (PAR). The R and D show the relative ratio and absolute difference between two categories within a dimension of inequality (highest and lowest wealth quintile). For D, a positive value indicates that there is pro-non-poor U5Ds and vice versa. The R statistic shows the relative inequality between poor and non-poor households. For an adverse indicator as U5Ds, R values equal to 1 indicate that there exists no inequality and values greater than one represent a pro-non-poor U5Ds. The higher this value is, the larger the gap between the poor and non-poor. The PAR is the difference between the most-advantaged subgroup (lowest wealth quintile) and the national average, while PAF is computed by ascertaining the ratio of the national average (μ) and the PAR, multiplied by 100, i.e. PAF = [PAR / μ] * 100. Unlike the R measure of inequality, the PAR and PAF take only negative values for adverse outcomes with higher values reflecting a wider gap between population subgroups. Comprehensive details regarding the computation of these measures have been reported [25]. The R, D. PAF and PAR estimate from household wealth inequality in U5Ds across LMIC using the WHO HEAT Plus are reported in Table 3. The graphical illustrations of the estimates are provided in Fig. 3.

We obtained the risk difference (RD) between the risk of U5Ds among children from poor and non-poor households for each country and showed the meta-analysis of these RDs in Fig. 2. We calculated the risk difference in U5D between poor and non-poor households and displayed the results in Fig. 2 as a country-level meta-analysis of U5D prevalence in each of the countries. A random-effects meta-analysis was used based on the assumption that each trial calculates a study-specific actual effect. Using the "metabin" tool in R, the meta-analysis was carried out by identifying the summary measure (SM) as risk difference (RD), the number of fatalities in poor and non-poor households, and the total number of participants for each country, stratified by regions [18]. Scatter and ordered balloon charts were used to show the distributions of the RDs viz-a-viz the prevalence of U5Ds in each country in Figs. 3 and 4. We defined pro-poor inequality as situations in which the RD in U5D is significantly lower among children from poor households than those from non-poor households and pro-non-poor inequality as situations in which the RD in U5D is significantly higher among children children from poor poor households than those from non-poor households [18, 19]. The countries formed 3 groups based on the RDs: countries with pro-poor, insignificant and pro-non-poor inequalities. The “pro-non-poor inequality” and “pro-poor inequality” countries are countries with higher U5D in poor households than in non-poor households and vice versa. Lastly, we fitted adjusted binary logistic regression to the risk of U5Ds among all the pro-poor countries and applied a Fairlie decomposition analysis (FDA) to the inequality in the U5Ds among children from poor and non-poor households and the results were presented in Fig. 5.

We applied sampling weights to all the analyses to control for different cluster sizes and stratifications, as well as to guarantee that our results accurately reflect the target population The "colin" tool in Stata version 16 was used to test for multicollinearity among the independent variables. The variable inflation factor was specified by the command (VIF). The VIF is around 1/(1-R2) and ranges from 1 to Regressing the jth independent variable on other independent variables yields the R2-value. All variables with a VIF greater than 2.5 were eliminated from the regression [37]. In several countries, insurance coverage, the employment status of father, access to media, cooking fuel type, and housing material were not reported and were excluded from the decomposition analysis. Prior to performing the decomposition analysis, we conducted a test of heterogeneity of U5D chances across all nations to confirm the presence of heterogeneity. We computed and presented the I-squared and the Mantel-Haenszel (MH) pooled estimate of the odds ratio (OR). We selected the pro-non-poor countries, conducted a homogeneity test among them, and provided the I-squared and MH pooled odds ratio (OR) estimates.

Several studies on the understanding of factors associated with inequalities in a wide range of health outcomes have adopted the technique of multivariable decomposition analysis [24, 26, 38‐40]. Multivariable decomposition analysis is ideal for the quantifications of the contributions of different factors to gaps in an outcome of interest between two groups [41]. It constrains the predicted probability of U5Ds to between 0 and 1. The difference between the predicted probability for one group (say, Group A – poor) using the regression coefficients of the other group (say, Group B – non-poor) and the expected probability for the non-poor group using its regression coefficients is measured in the decomposition analysis [42].

According to Fairlie et al., the decomposition of a nonlinear model Y=F(X) can be written as:

Where \({N}^{A}\) is the sample size for group \(J\). Other model details have been reported [18, 19, 31, 33, 43]. The independent contribution of \({X}_{1}\) and \({X}_{2}\) to the gap are expressed as follows:

and

respectively. Further numerical details have been documented in the literature [42, 44‐47]. In this study, the FDA was implemented in STATA version 16 (StataCorp, College Station, Texas, United States of America) using the “Fairlie” command. However, Fairlie’s sequential decomposition has issues with path dependence [42, 44‐47], whereby different ordering of variables in the decomposition analysis produces different results. To address this, we checked the robustness of the sensitivity analysis of variable re-ordering randomization. First, we conducted and assessed the performance of 10 different ordering of the variables and tested the sensitivity of decomposition estimates. Secondly, we invoked the “random” option with the “Fairlie” Stata command used in conducting the Fairlie decomposition. In this study, the FDA was implemented in STATA version 16 (StataCorp, College Station, Texas, United States of America).

The RDs, a measure of inequality in the risk of U5Ds among children from poor and non-poor homes across the 59 nations, are shown in Fig. 2. In all countries except Ethiopia, Tanzania, Zambia, Lesotho, Gambia, and Sierra Leone, and the Maldives, the prevalence of U5Ds was higher among children from poor households than among children from non-poor households. The RDs were considerably greater in 32 countries among children from poor families, but not in any country among children from non-poor households. The distribution of the fixed effects of poor-non-poor RD showed the widest gap in Nigeria (50.4 per 1000 children) followed by Guinea (39.9 per 1000 children). The random-effects irrespective of the child's country of residence was 11.4/1000 (95% confidence interval (CI): 8.2 – 14.5). This indicates that there is significant pro-non-poor inequality in U5Ds in LMICs. India (2.0%) contributed the most weight to the random effect, with 1.9% contributions each in Kenya, Egypt, Afghanistan, Indonesia, Jordan, Colombia, Peru, and Albania while the least weight contribution to the random effect was in Lesotho. The heterogeneity level among the RDs was 91.8% (p<0.01).

Household inequality in U5Ds in 59 countries using the measures recommended in the WHO HEAT Plus showed that there exist wide gaps in U5Ds among poor and rich households. The D values indicate that there was a pro-non-poor inequality in U5Ds in the majority of the countries, with the highest gap observed in Nigeria (D=7.4, CI: 6.4-8.4), Guinea (D=6.9, CI: 5.2-8.6), Mali (D=5.9, CI: 4.3-7.5) and South Africa (D=4.3, CI:2.4-6.2). There was a pro-poor inequality in U5Ds in only four countries, Comoros (D=-0.4, CI: -2.8-2), Maldives (D=-0.5, CI: -2.9-1.9), Sierra Leone (D=-1.4, CI: -3.3-0.5) and Tanzania (D=-1.5, CI: -2.9--0.1). The R values indicated that there were gaps in U5Ds but the largest gaps were seen in Cambodia (R = 5.9, CI: 3.3-10.3) and South Africa (R = 4.9, CI: 1.8-13.6). Similarly, the R values show that the burden of U5Ds was concentrated among richer households in Comoros (R = 0.9, CI: 0.6-1.5), Maldives (R = 0.7, CI: 0.2-2.7), Sierra Leone (R=0.9, CI: 0.8-1) and Tanzania (R=0.7, CI: 0.6-1). This pattern was also observed for the PAR and PAF measures of inequality as shown in Table 3. The visualization of the distribution of these measures in the LMIC is shown in Fig. 3.

Figures 4 and 5 shows the distribution of risk difference of U5Ds by the prevalence of U5Ds in each of the countries. In these charts, significant pro-non-poor inequalities are shown in red colour while insignificant inequities are shown in yellow. There was no significant pro-poor inequality in any country. As shown in the RDs, two of the nine nations in Eastern Africa, four of the six countries in Middle Africa, Egypt in Northern Africa, and six of the thirteen countries in West Africa exhibit significant pro-poor U5Ds inequality. There are two countries in each of Southeast Asia and Western Asia, three in the Caribbean, five in Southern Asia, and one in Southern Africa. Papua New Guinea in Oceania, South America, Central Asia, Central America, and Southern Europe all have high pro-non-poor U5Ds inequality, as seen in Figs. 1, 4, and 5.

We categorized the 59 countries into 4 distinct groups based on the prevalence of U5Ds in each country and based on the magnitude of the RDs which reflects the level of inequality: (i) High prevalence of U5Ds and high pro-non-poor inequality countries which were observed in countries like Nigeria, Guinea, Burkina Faso, Mali and Chad (ii) High prevalence of U5Ds and high pro-poor inequality countries such as Ethiopia, Lesotho and Sierra Leone (iii) Low prevalence of U5Ds and high pro-non-poor inequality countries such as Cambodia, South Africa, Guatemala and Myanmar (iv) Low prevalence of U5Ds and high pro-poor inequality countries such as Maldives, Gambia, and Zambia (Fig. 5).

The Mantel-Haenszel (MH) pooled estimate of the odds ratio (OR) of having U5Ds while controlling for the country of residence among children was 1.38 (95% CI: 1.35-1.41) and tested H_o: OR=1; we estimated z = 32.3 and p = 0.000 and (ii) Test of heterogeneity, we estimated X² = 431.8, degree of freedom (d.f.) = 58, and p = 0.000, I-squared (variation in odds ratio (OR) attributable to heterogeneity) = 86.6%. Thirty-four of the 59 countries showed a significant pro-non-poor odds ratio, no significant pro-poor inequality while other countries showed no significant inequality. The 34 countries are Afghanistan, Angola, Bangladesh, Benin, Burkina Faso, Burundi, Cambodia, Cameroon, Chad, Colombia, Cote D’Ivoire, Egypt, Ethiopia, Guatemala, Guinea, Haiti, India, Indonesia Malawi, Mali, Myanmar, Nepal, Niger, Nigeria, Pakistan, Papua New Guinea, Philippines, Rwanda, Senegal, South Africa, Tajikistan, Togo, Turkey and Yemen. A MH OR across these countries was 1.51 (95% CI: 1.47 – 1.54), test of homogeneity of odds ratio was significant with I² = 93.1%, X² = 155.16, d.f = 33, and p = 0.000.

Across the 34 countries, the largest contributors to pro-non-poor inequalities in U5Ds among the children are rural-urban differences in the location of residence, maternal education, neighbourhood SES, sex of the child, toilet types, birth weight and preceding birth intervals, sources of drinking water and household wealth. The countries with the largest contributions of these factors are Turkey, Cote D’Ivoire, and Niger. These countries were clustered together while the location of residence, birth order, maternal education, sex of the child, toilet type and maternal employment were clustered together as shown in Fig. 6. The largest contributors to pro-non-poor inequality in Turkey were residence location (473%), birth order (209%) and maternal education (319%). In Cote D’Ivoire, residence location (245%), birth order (213%) and maternal education (35%), contributed the largest to pro-non-poor inequality in U5Ds. Also, the contributions of household wealth to gaps in U5D in poor and non-poor households were shown in Fig. 5 with the highest influence in Turkey, Cote D’Ivoire, Colombia, Ethiopia and Senegal. In general, poor maternal education widens the wealth inequality in child death while better educational attainment closes the gap. Also, living in rural areas, social-economic disadvantaged communities, unimproved toilet type, low birth weight and high birth interval widens wealth inequality in under-five deaths.

This study has some limitations which are majorly data issues. The DHS survey was conducted at different times across the 59 LMICs and this may introduce bias in the comparison of U5Ds. Similarly, recall bias may have occurred as mothers were made to recall past events. There is also the possibility of under-reporting of U5Ds in some countries which is often related to cultural taboos forbidding parents from reporting the deaths of their children. Another limitation is that maternal weight or BMI would be useful as the control variables, however, this was not included because maternal height and weight were inadequately reported in the DHS. Only 35% of mothers have the two variables. In addition, FDA is useful for estimating the relative contributions of factors on the outcome of U5Ds. However, the technique does not account for the clustering and stratification elements of the DHS. It is therefore not unlikely that this may have had some impact on the results generated. Nonetheless, FDA is a reliable way for determining the contributions of various factors to gaps in a desired outcome between two groups, and this decomposition technique is an improvement over the Blinder Oaxaca decomposition method.