Sex inequality in under-five deaths and associated factors in low and middle-income countries: a Fairlie decomposition analysis

Background: There exist sex disparities in the burden of Under-five deaths (U5D) with a higher prevalence among male children. Factors explaining this inequality remain unexplored in Low-and Medium-Income Countries (LMIC). This study quantified the contributions of the individual- and neighborhood-level factors to sex inequalities in U5D in LMIC. Methods: Demographic and Health Survey datasets (2010-2018) of 856,987 under-five children nested in 66,495 neighborhoods across 59 LMIC were analyzed. The outcome variable was U5D. The main group variable was the sex of the child while individual-level and neighborhood-level factors were the explanatory variables. Fairlie decomposition analysis was used to quantify the contributions of explanatory factors to the male-female inequalities in U5D at p<0.05. Results: Overall weighted prevalence of U5D was 51/1000 children, 55 among males and 48 among females (p<0.001). Higher prevalence of U5D was recorded among male children in all countries except Liberia, Kyrgyz Republic, Bangladesh, Nepal, Armenia, Turkey and Papua New Guinea. Pro-female inequality was however not significant in any country. Of the 59 countries, 25 had statistically significant pro-male inequality. Different factors contributed to the sex inequality in U5D in different countries including birth order, birth weight, birth interval and multiple births. Conclusions: There were sex inequalities in the U5D in LMIC with prominent pro-male-inequality in many countries. Interventions targeted towards the improvement of the health system that will, in turn, prevent preterm delivery and improve management of prematurity and early childhood infection (which are selective threats to the male child survival) are urgently required to address this inequality. We obtained under-five children data from the Demographic and Health Surveys (DHS). The DHS holds approximately every five years across the participating LMIC. The ICF (USA) collects the data in conjunction with the designated organizations in the participating countries. Typically, the survey is cross-sectional, nationally representative and population-based. We pooled data from the most recent DHS conducted between 2010 and 2018 and in the public domain as of 10 September, 2020. A total of 59 LMIC met these inclusion criteria, and their data were included in this study. We pooled the data of 856,987 under-five children, from 66,495 neighborhoods across the 59 LMIC. The DHS utilized a similar clustered multi-stage sampling procedure in the participating countries based on countries’ sampling frames drawn mostly from the last census counts. Countries were stratified using the existing geographical and administrative structures. The multi-stage mechanism included the states/divisions/regions in the first stage, districts as the next stage in some countries, and finally, the clusters as the last stage. The clusters were the primary sampling units (PSU) [14, 15]. The households were then selected from the PSUs, from which women aged 15-49 years were interviewed. The surveys generated different datasets. We used the child recode data that captured diverse information on all births of the interviewed women five years before the survey. Sampling weights were computed and provided alongside the data by DHS. These computations were based on the multi-stage sampling method to ensure the representativeness of the sample concerning the general population. The DHS uses similar surveys and research protocols, standardized questionnaire, similar interviewer training, supervision, and implementation in all the countries. The full details of the sampling methodologies and other information are available at dhsprogram.com. The dependent variable was UD5 which was defined as death among live births within the first five years of life, that is deaths within 0 to 59 months of birth [14]. To ascertain the correctness of this outcome, mothers were first asked if they had given birth to any child five years preceding the date of the study. They were then asked to recount the date of birth and were assisted to estimate such dates when necessary. They were asked if each of those children were alive or dead. The dates of death or the ages at death for the dead children were then used to determine U5D. Therefore, U5D was binary: Alive or Dead before 5th birthday. The main group variable is the sex of the child: male or female. The variables identified to be associated with childhood deaths in the literature [16–20] were selected using Moseley’s systematic conceptual framework on the study of child survival in developing countries [17]. The variables were made up of individual-level and neighborhood-level factors. The individual-level factors consist of a child’s characteristics, mothers’ characteristics and the households’ characteristics. Child's characteristics were weight at birth (average+, small and very small), birth interval (firstborn, <36 months and >=36 months). birth order (1, 2, 3 and 4+) and whether a child is a twin (single, multiple (2+)). While mothers’ characteristics were maternal education (none, primary or secondary plus), maternal age (15 to 24, 25 to 34, 35 to 49 years), marital status (never married, currently and formerly married), maternal and paternal employment status (working or not working), and health insurance (yes or no). Households’ characteristics were the sex of the head of the household (male or female), access to media (at least one of radio, television or newspaper), sources of drinking water (improved or unimproved), toilet type (improved or unimproved), cooking fuel (clean fuel or biomass), housing materials (improved or unimproved), household wealth index (poorest, poorer, middle, richer and richest) and place of residence (rural or urban). The sources of clean fuel are electricity, liquefied natural gas/biogas and unclean fuel include wood, charcoal, kerosene, straw shrubs, animal dungs and grass. The improved sources of drinking water include a protected well, borehole, bottled water and spring rainwater, while spring water, tankers, unprotected well with drum, sachet water, surface water, and other sources constituted the unimproved sources. The housing material was based on a composite score according to the type of wall, floor and roof materials. If cement/carpet/rug/ceramic tiles/vinyl asphalt strips were used for the floor, the floor quality is coded 1, else it is coded 0. In the same vein, wall material quality is coded 1 if it is made of cement blocks/bricks, else 0. If roof material is made of calamine/cement roofing shingles/cement fibres/ceramic tiles/zinc, it is coded 1, else 0. If all the materials fall under “1” they are regarded as “improved”, else, they are “unimproved” [14, 15]. We defined “neighborhood” as the clustering of children. The DHS uses “clusters” as the PSU [14, 15], hence “neighborhood” in this context is the clustering of children within the same geographical environment and children were “neighbors” if they belonged to the same cluster. In this study, we considered neighborhood socioeconomic status (SES) as a neighborhood-level variable. It was computed using the principal component factor method from the scores that were aggregated from the proportion of respondents within the same clusters without education, belonging to a household in the two lowest wealth quintiles, no media access and unemployed. The “xtile” function in Stata version 16 was used to categorize the scores into five categories: Least disadvantaged, 2, 3, 4 and most disadvantaged [5]. We used both descriptive and inferential statistics in this study. Descriptive statistics including charts, tables, and percentages were used to show the distribution of the children by country, regions, U5D and other key variables. A bivariable analysis was conducted using the Z-test for equality of proportions of U5D among male and female children within each country and region (Table ​(Table1).1). We also determined if any association existed between the explanatory variables and U5D among the male and female children (Table ​(Table2).2). The risk difference (RD) in under-5 deaths among male and female children were computed. A RD greater than 0 suggests that U5D is higher among male children than among female children (pro-male inequality). While a RD = 0 signifies no difference and a negative RD indicates that U5D was higher among female children than among male children (pro-female inequality). The RDs were computed to identify the countries where significant differences existed in the U5MR by gender. Distribution of sample characteristics and prevalence of under-five deaths in LMIC by sex, 2010–2018 aSignificant at 0.05 in the z-test of equality of proportions Characteristics of the studied children and prevalence of under-five-deaths in LMIC by sex, 2010–2018 We estimated both random and fixed effects of the RD. The fixed effects are the weighted country-specific RD and the random effects are the overall RD irrespective of a child’s country of residence as shown in Fig. ​Fig.1.1. The purpose of the random effect was to estimate the overall prevalence and distributions of prevalence of U5MR among males and females irrespective of the countries the children are located. The fixed effects are to establish and identify the country-specific estimates. Charts were used to show the distributions of the RDs by the countries (Figs. ​(Figs.22 and ​and3).3). We categorized the countries into four distinct categories based on their prevalence of U5D and the size of their RD: (i) High U5D and high pro-male inequality (ii) High U5D and high pro-female inequality countries (iii) Low U5D and high pro-male inequality (iv) Low U5D and high pro-female inequality (Fig. ​(Fig.3).3). The Mantel-Haenszel (MH) Odds Ratio (OR) and tests of heterogeneity of ORs were used to ascertain that the countries were different with regards to the odds of U5D among the male and female children. A test of homogeneity of ORs among all the countries with a significant OR of U5D was also used to determine if the odds of having U5D in those countries were homogenous. Lastly, the multivariable-adjusted logistic regression method was applied to the pooled cross-sectional data from the U5D pro-male countries to carry out a decomposition analysis using the multivariable Fairlie decomposition analysis procedures. Sampling weights were applied in all our analyses to adjust for unequal cluster sizes, stratifications and to ensure that our findings adequately represent the target population. Risk difference in under-five deaths between male and female children by countries in LMIC Risk difference between children from houses with sex differentials in under-five-deaths by countries in LMIC Scatter plot of rate of under-five-deaths and risk difference between sex of children in LMIC Multicollinearity among the independent variables was tested using the “colin” command in Stata version 16. The command provided the Variance Inflation Factor (VIF). The VIF is approximate of 1/(1-R2) ranging from 1 to infinity. The R2-value is obtained by regressing the jth independent variable on other independent variables. All variables with VIF>2.5 were removed from the regression analysis. Literature has shown concerns about VIF >2.5 [18]. Health insurance cover, media access, paternal employment status, type of cooking fuel and housing material were not captured in some countries and were dropped in the decomposition analysis. The decomposition analysis was conducted by obtaining the logit estimates of the magnitude of contributions of the factors to gaps in U5D between males and females as the dependent variable among those countries that had significant RDs. We applied multivariable Fairlie Decomposition Analysis (FDA) based on the binary regression model. The FDA is one of the decomposition techniques used in the quantification of the contributions to differences in the prediction of an outcome of interest between two groups in multivariate models [19]. The method is an extension of the Blinder-Oaxaca Decomposition Analysis [20–22], which has been roundly criticized for inefficiency in handling the logit and probit model [22, 23]. The FDA was purposively developed for non-linear regression models including the logit and probit models [24]. The FDA was carried out by calculating the difference between the predicted probability for one group (say Group A – male children) using the other group’s (say Group B female children) regression coefficients and the predicted probability for male children using its regression coefficients [23]. The Fairlie decomposition technique works by constraining the predicted probability between 0 and 1. Fairlie et al. showed that the decomposition for a nonlinear equation Y = F(X), can be expressed as: Where NA is the sample size for group J (25). In Eq. (1), Y¯ is not necessarily the same as FX¯β^, unlike in BODA where F(Xiβ) = Xiβ. The 1st term is the part of the gap in the binary outcome variable that is due to group differences in distributions of X, and the 2nd term is the part due to differences in the group processes determining levels of Y. The 2nd term also captures the portion of the binary outcome variable gap due to group differences in unmeasurable or unobserved endowments. In other words, the explained factors are those factors attributable to gender differences in individual observable characteristics and life circumstances while the unexplained factors are related to gender differences in the unobservable characteristics and life circumstances. The estimation of the total contribution is the difference between the average values of the predicted probabilities. Using coefficient estimates from a logit regression model for a pooled sample, β^*, the independent contribution of X1 and X2 to the group, the gap can be written as and respectively. The contribution of each variable to the gap is thus equal to the change in the average predicted probability from replacing the group B distribution with the group A distribution of that variable while holding the distributions of the other variable constant. Further numerical details have been reported [23–27]. We implemented the FDA in STATA 16 (StataCorp, College Station, Texas, United States of America) using the “Fairlie” command.