Impact of Maternal and Child Health Interventions On Under-Five Mortality in Ghana: A Quasi-Experimental Study Design With Coarsened Exact Matching

Despite a 53% decline in under-ve mortality (U5M) worldwide during the period of the Millennium Development Goals (MDGs), U5M remains a challenge. Under-ve mortality decline in Ghana is slow and not parallel with the level of coverage of child health interventions. This study sought to assess the effectiveness of these interventions on U5M in Ghana. Methods A quasi-experimental study was conducted using secondary data of the 2008 and 2014 Ghana Demographic and Health Surveys. Coarsened Exact Matching and logistic regression were done. tetanus toxoid vaccination, clean postnatal care, hygienic disposal of stool, early initiation of breastfeeding, intermittent preventive treatment of malaria in pregnancy (IPT-p), iron intake, and skilled delivery. These are interventions expected to be implemented before, during or shortly after birth, and that are therefore likely to have preceded (temporal association between exposure and outcome) death for those children who received them(22). In addition, an overall or composite intervention group was created using children who received all of the eight (8) interventions. Denition of interventions can be found in the 2014 Ghana Demographic and Health Survey(11). For each child that received an intervention (treated), a control child who did not receive the intervention (untreated) was matched using coarsened exact matching(23-26). On household factors, children in households with 3 or more children under-ve years had their odds of death reduced by 81% (aOR=0.19, 95% CI: 0.07-0.51) compared to those in households with one or two children under ve years old. Also, children in polygamous homes were 2 times more likely to die compared to children in monogamous homes (aOR=2.16, 95% CI: 1.05-4.44). Odds of death was reduced by 46% among children whose mothers had health insurance coverage (aOR=0.54, 95% CI: 0.31-0.96).


Introduction
Under-ve mortality (U5M) remains a major public health problem despite a 53% decline globally between 1990 and 2015 (1). Decline of U5M in Ghana is considered slow (2)(3)(4). Thus, it could not achieve the Millennium Development Goal (MDG) 4 target (4, 5) despite several efforts at 90% for both surveys used in this analysis (3). Data sets of the 2008 and 2014 surveys were pooled for the analysis. The years 2008 and 2014 were chosen because they are the data sets closest to the transition from the MDGs to the SDGs. Information from the study will therefore provide understanding of the interventions that contributed to mortality reduction towards the end of the MDGs, and thus should be focused on during the SDGs period in order to achieve rapid mortality reduction.
The Household Recode (HR) and Children Recode (KR) les were used for this analysis. The KR le was the main data set and it contained information of children under-ve years old at the start of the data collection whose mothers where interviewed (3). The 2008 GDHS data set had 2,882 observations and 1,010 variables while the 2014 had 5,884 observations and 1,159 variables ( gure 1). The Stata versions of the datasets were downloaded from the DHS Program's website (https://dhsprogram.com). The HR datasets were merged with the KR data sets so that household insecticide treated nets/indoor residual spraying (ITN/IRS) information from the HR data set could be added to the KR data set. Household insecticide treated nets/indoor residual spraying is an intervention and a potential confounder and needed to be adjusted for in the analysis. The data sets of 2008 and 2014 were combined to obtain the nal data set for the analysis.
Forty-eight percent (48%) of observations had at least one missing value for the covariates to be adjusted for and therefore multiple imputation was done before the regression analysis were done. Imputation was done to preserve sample size and representativeness of the data. Variables with missing data were birth interval (16 observations), number of cowives (832 observations), mode of delivery of child (2 observations), child's weight (19 observations), employment status of mother (13 observations), NHIS status (1 observation), maternal BMI (2,601), anaemia (204), height (2,008) and child given other milk within three days after birth (58). Twenty (20) imputations were done using chain imputation and logistic regression. Imputation data format was marginal long and seed set was 200. Variables used in the multiple imputation were all variables without missing data that were used in the analysis. The design feature, weighting, was factored into the multiple imputation model. The stata syntax of the multiple imputation is presented in additional le 1.

Study variables
The variables used in this analysis were socio-economic and demographic variables, and interventions. Socio-economic factors were maternal and household factors including maternal education, marital status, religion, ethnicity, household wealth, household size, number of children under-ve years in a household, region, and place of residence. Demographic factors comprised child and maternal factors including maternal age, child's age and birth weight, birth interval, birth order and multiple birth. Interventions were antenatal visit, IPT-p, iron intake, tetanus toxoid vaccine, skilled delivery, clean postnatal care, early initiation of breastfeeding, hygienic disposal of stool, water connection in the home, improved water source, improved sanitation, ITN/IRS, antenatal care visits, and NHIS status. The year of the survey was also included. De nition of variables were based on the 2014 GDHS report (3).
Pre-processing of data Coarsened Exact Matching (23) was used to pre-process (match) the data to ensure balance in pre-treatment variables (covariates) between the treated and untreated groups. Covariate balance was assessed using linear 1 (L1) statistic. L1=0 means perfect covariate balance between treated and untreated groups while L1=1 means complete incompatible treatment and comparison groups with respect to the covariates. The closer the values of L1 to zero, the better the covariate balance (23). Information on the sample size and covariate balance before and after matching are presented in table 1.
CEM puts observations into groups based on factors that in uenced decision to received treatment. The groups were chosen by the researcher to re ect clinically important categories and to be consistent with that in the GDHS reports. Observations in similar groups were matched and treated and untreated individuals who did not have matches were deleted from the data set and the data analyzed as unmatched data. The covariates used in the matching process were variables that predicted intervention use from the literature review and were also associated with the intervention from chi-square test done in this data analysis (Table 1).

Analysis
Descriptive statistics were done and presented as numbers (percentage) by survival status (died or lived) and chi-squared test was used to assess the differences. Logistic regression was also tted to assess factors associated with U5M after the multiple imputation. The design features (strati cation, clustering and weighting and selection of primary sampling units) of the GDHS were factored into all the analysis. Also, the differences in women's population between the two time periods was also accounted for in the analysis. De-normalise weights were used to account for the differences in the population of women in reproduction age in the population at the time of each survey. The de-normalised weight was calculated as = ψ X ϕ15-49/ϕ S 15-49 (33). Where ψ is the DHS sampling weight for women. ϕ15-49 is the population of women in the country at the time of each survey, while ϕ S 15-49 is the total number of women 15-49 years interviewed during the data collection for each of the survey as reported in the GDHS reports (11,34).
Potential confounders adjusted for were those related to mortality and/or exposure from the conceptual framework and are not intermediates in the causal pathway of the exposure and outcome (35,36). Robust estimates of odds ratios (ORs) with 95% con dence interval were reported.
For the analysis of causal effect, under the null hypothesis that no intervention had a causal effect on U5M, a logistic regression was tted using the pre-processed CEM data. Each intervention (antenatal care visit, IPT-p, iron intake, tetanus toxoid vaccine, skilled delivery, clean postnatal care, early initiation of breastfeeding, and hygienic disposal of stool) was the exposure, accounting for design features, and controlling for potential confounders. Potential confounders adjusted for were those identi ed from the conceptual framework. Average treatment effects were reported as OR with 95% CI and at a 2-tailed α level of 0.05 (37). Analyses were performed using STATA version 13. The CEM syntax and algorithm are included in additional le 1.

Sensitivity analysis
1. Complete case analysis without maternal nutrition factors When logistic regression was tted with the completed cases, the total number of observation was 5,224 and early initiation of breastfeeding, child's age, multiple birth, birth interval, number of children under-ve years old in the household and polygamy were associated with U5M from the adjusted analysis.

Multiple imputation with maternal nutrition factors
After multiple imputation, when maternal anaemia and Body Mass Index (BMI) were added, the sample size reduced from 6,098 to 3,165 observations with very wide 95% con dence intervals and the variable, water connection in the home omitted from the regression model. BMI and anaemia were not statistically signi cant while household size, child's age, multiple birth, maternal education, household indoor residual spraying and bed net ownership, sex of household head and early initiation of breast feeding were statistically signi cant. National Health Insurance Scheme status, had borderline signi cance, p=0.05. Con dence interval for multiple births ranged from 9.83 to 417.8.

Coarsened exact matching with completed cases without maternal nutrition factors
For the CEM analysis with completed cases, sample size after matching for the regression analysis ranged from 4,914 to 6,043 observations. Only early initiation of breastfeeding had causal effect on U5M reduction from both the crude and adjusted analysis. Clean postnatal care and hygienic disposal of stool were signi cant from the crude analysis but not from the adjusted analysis.

Results
Descriptive statistics There were 6,098 children under-ve years and 93 (1.46%) died (Table 3). Among children who died, 47 (47%) were less than one month old. A greater proportion 27 (37%) of mothers had primary education as their highest level of education. Fifty-six, representing 65.7% of children were born to mothers below 35 years. Among the interventions, coverage of antenatal care visits was the highest coverage level 5,045 (84.0%) while water connection in the home was the lowest coverage level 469(8.1%) ( Table 4). Less than 1.5% (58) of children received all the eight (8) interventions and none of those who received all eight interventions died.

Socio-demographic factors and interventions associated with under-ve mortality
Results of the association of child, maternal and socio-demographic factors with U5M are presented in Table 6. The crude and adjusted ORs and 95 % CI for the associations are presented. On child factors, after controlling for potential confounders, compared to children less than 1 month old, odds of death was reduced by 98% (aOR=0.02, 95%CI: 0.01-0.04) among children 1-5 months. Also, a 98% reduction of odds of death was reported among those 6-11 months (aOR=0.02, 95% CI: 0.01-0.05) and 99% among those 12-59 months (aOR=0.01, 95%CI: 0.002-0.01) relative to those less than 1 month old. Children who were multiple births were 7 times more likely to die compared to singleton births (aOR=6.68, 95% CI: 1.56-28.56), while those with preceding birth interval of less than two years twice more likely to die relative to those with longer birth interval (aOR=2.25, 95%CI: 1.10-4.60). On household factors, children in households with 3 or more children under-ve years had their odds of death reduced by 81% (aOR=0.19, 95% CI: 0.07-0.51) compared to those in households with one or two children under ve years old. Also, children in polygamous homes were 2 times more likely to die compared to children in monogamous homes (aOR=2.16, 95% CI: 1.05-4.44). Odds of death was reduced by 46% among children whose mothers had health insurance coverage (aOR=0.54, 95% CI: 0.31-0.96).

Interventions with causal effect on under-ve mortality
The crude and adjusted ORs of the causal effect of the interventions are presented in tables 2 and 5. From the crude analysis, average odds of death was reduced by 68% among those with early initiation of breastfeeding (OR=0.32, 95% CI: 0.19-0.54) and 65% among children under-ve years who had clean postnatal care within 2 days (OR=0.35, 95%CI: 0.18, 0.70). After adjusting for potential confounders, early initiation of breastfeeding reduced odds of death by 61% (aOR=0.39, 95% CI: 0.22-0.71) while clean postnatal care caused a 64% reduction in the average odds of death (aOR=0.36, 95%CI: 0.15, 0.90) ( Tables 2 and 5).

Discussion
Interventions with causal effect on under-ve mortality Despite efforts at reducing U5M, mortality decline in Ghana is slow. Although interventions targeted at reducing under-ve mortality might show e cacy under experimental conditions, their effectiveness might be suboptimal resulting in their low impact on mortality. This study evaluated the effectiveness of the various child health interventions on U5M in Ghana. The results showed that the various child health interventions impact child mortality differently.
Children who had clean postnatal care had a 64% reduced average odds of death. Clean postnatal care is de ned as neonates receiving a preventive postnatal visit within 48 hours of delivery (3). The assumption is that neonates who receive clean postnatal care will subsequently receive adequate clean postnatal care in the home (3,38).
Sepsis is a major cause of neonatal mortality and cord care in uences the incidence of neonatal sepsis. Therefore clean birth and postnatal care practices are recommended for reducing infection and neonatal mortality (39) and thus the recommendation that women after delivery have postnatal care within 48 hours. This will ensure that infections are identi ed early for timely management. Clean postnatal care can signi cantly contribute to the elimination of neonatal tetanus which is a major contributor to neonatal mortality (40). From the 2014 Ghana Demographic and Health Survey, mothers received clean postnatal either from skilled attendants or from traditional birth attendants. Skilled health workers include doctors, nurses, midwives, community health nurses or community health o cers. Clean postnatal care is one of the interventions with the lowest coverage levels in Ghana. Coverage in 2008 was 6.5% while that in 2014 was 22.8% (11). Its low coverage is therefore an avenue to further reduce deaths if its coverage level is increased. Results of the protective effect of clean postnatal care on mortality in this study is similar to that documented (5,(41)(42)(43).
Breastfeeding have been shown to reduce the risk of infections and consequently death of children under-ve years old (41,(44)(45)(46)(47). Early breastfeeding has additional bene ts as it promotes warmth and bonding between infants and their mothers. In this study, early initiation of breastfeeding caused a 61% reduction in the average odds of death. Similar results of the association between early initiation of breastfeeding and under-ve mortality have been documented (44)(45)(46)(47)(48). According to the GDHS, coverage level of early initiation of breastfeeding was 25.5% in 1998, 46.3% in 2003, 52.3% in 2008 and 55.6% (38) in 2014(3). This trend of coverage increase is slow. Considering the 2014 coverage level of the intervention, it has the potential to further reduced mortality if its coverage level is increased. Since this intervention does not require much logistics or expenditure to implement compared to interventions like skilled delivery, it should be prioritized.

Other factors and under-ve mortality
In the midst of interventions, some socio-demographic factors remained associated with U5M. Children in households with three or more children under-ve years were 81% less likely to die compared to those in households with one or two children. Having more children under-ve years could mean the mother will have more experience with taking care of children including identifying signs and symptoms of diseases. Having more children under-ve years in a household has been associated with early care seeking in Niger (49). Children in households with higher number of members also had reduced odds of neonatal and under-ve mortality in Ghana (33).
Health insurance membership offers nancial access to healthcare and is association with increased and timely healthcare seeking (50)(51)(52). It could therefore increase the use of all the healthcare associated interventions, and therefore, its association with a 46% reduction in the odds of death of children under-ve years in this study. Similar result of the protective effect of NHIS on mortality has been reported(33).
The higher odds of death of younger children and multiple births have been documented (41). In this study, while multiple births were about 7 times more likely to die, compared to neonates, odds of deaths was reduced by about 99% among children 12 months and older. Addressing the increased odds of death with multiple births and younger children will require improved access to quality skilled delivery. Unfortunately, skilled delivery was not associated with mortality reduction. Younger children, especially neonates, are less developed and more susceptible to infections. However, specialized and advanced care needed at the neonatal period are usually not available especially in poverty ridden communities. Considering the causes of neonatal deaths such as sepsis, diarrhoea, pneumonia and asphyxia, quality skilled delivery, improved nutrition and improved hygiene will play important roles in its reduction (53). But, coverages of hygiene and sanitation interventions are among the interventions with the lowest coverage levels in Ghana (3,38). Also, coverage of exclusive breastfeeding and complementary feeding are on the decline (3,38).
Additionally, children from mothers in polygamous marriages had twice higher odds of death. This nding is consistent with results reported in other studies (54,55). Rivalry among wives was attributed to the higher odds of death. Limited resources and overcrowding have also been cited as reasons for the positive correlation of polygamy and child mortality (56).
Lastly, children of birth intervals of less than two years also had twice higher odds of death compared to those of longer intervals. With shorter birth interval, mothers might not have replenished nutrients used up during the previous pregnancy resulting in under-nutrition(57). After delivery, there will be competition for maternal resources including time to care for the children. This can compromise the quality of their care resulting in infections, improper nutrition and poor health. Shorter birth interval will also limit the duration of breastfeeding which can affect the development of the older child. Similar results on the higher risk of death among children with shorter birth intervals have been documented in Ghana(5, 58).
The lack of effect of the other interventions on under-ve mortality could be attributed to the context of the implementation of these interventions. The context could include the incidence of the disease(s) each intervention is targeted at, and the quality of the implementation of the interventions. There is evidence of poor quality of skilled delivery contributing to lack of effect of skilled delivery on child mortality(59). Disparities in the association of skilled delivery with neonatal mortality have been documented in different geographic areas. While skilled delivery improved neonatal mortality in Latin America and the Caribbean, it was associated with worse neonatal mortality in Africa(60). The quality and availability of logistics and health personnel for skilled delivery might explain the differences in effectiveness of skilled delivery in different places(60).
Low incidence of disease that the intervention is targeted at could also explain the lack of statistical signi cance of the association between interventions and under-ve mortality(61). With low incidence of disease, the intervention will have fewer deaths to prevent and therefore, low effectiveness. That could be the case for tetanus toxoid vaccine. Improvement in birthing practices could reduce infections due to Clostridium tetani, and therefore, low risk of neonatal tetanus, a resultant reduced risk of death from neonatal tetanus. Currently, Ghana is at the elimination state of neonatal tetanus which means lower risk of neonatal tetanus infection(61).
With multiple interventions which have direct and indirect effect on under-ve mortality been implemented in the midst of socio-economic factors that also have effect on under-ve mortality, Chowdhury (62) observed that proximate factors have a stronger effect on under-ve mortality than more distal factors. Iron intake, antenatal care visits and intermittent preventive treatment of malaria in pregnancy (IPT-p) have relatively indirect effect on U5M and therefore, could account for their lack of statistical signi cant effect on mortality.

Limitations and strength of study
On limitations, coarsened exact matching like other matching methods matches on measured potential confounders. However, the confounders adjusted for in this analysis comprised all potential confounders and therefore, the effect of unobserved variables is ignorable. The pooling of the data also increased the power of the study and thus the validity of the results. The matching method that was used reduced model dependence and thus, increases the robustness of the analysis. Lastly, the restriction of the analysis to only the last births reduced the in uence of recall bias on the study results. On the strengths of the study, nationally representative data were used which had high response rates (over 90%) and therefore, the study ndings are generalizable.

Conclusion
In conclusion, of the eight (8) interventions assessed for effectiveness at reducing under-ve mortality, two (2) interventions showed effectiveness on mortality reduction. This therefore suggests that the scope and content of the current package of interventions targeted at reducing U5M, will likely not achieve rapid decline in mortality. At best, the mortality rate will be stagnant. To achieve further decline in mortality, coverage of early initiation of breastfeeding and clean postnatal care should be increased. Early initiation of breastfeeding and clean postnatal care reduce the incidence of sepsis which is a major cause of neonatal deaths. Since the decline in neonatal mortality lags that of other age groups of children under-ve years, the Ministry of Health (MOH) and the Ghana Health Service (GHS) should put more emphasis on education of mothers on early initiation of breastfeeding and clean postnatal care at antenatal care and child welfare visits. In the midst of the various interventions, being a neonate, multiple birth or from a polygamous home put children at a higher risk of under-ve mortality. Addressing issues affecting the health of children in polygamous homes and multiple births could be bene cial. In contrast, belonging to a home with more number of children under-ve years and, a mother with health insurance coverage reduced odds of death. Therefore, measures to improve on the services of the National Health Insurance Scheme, including improving the availability of diagnostic and treatment logistics might contribute to improved child survival. Publically available data were used. The study did not involve human subjects as secondary data was used.

Consent for publication
Not applicable Availability of data and materials Data used for this work are available with the corresponding author and also on the website of the DHS Program (https://dhsprogram.com).