The child survival impact of the Ghana Essential Health Interventions Program: A health systems strengthening plausibility trial in Northern Ghana

Background The Ghana Health Service in collaboration with partner institutions implemented a five-year primary health systems strengthening program known as the Ghana Essential Health Intervention Program (GEHIP). GEHIP was a plausibility trial implemented in an impoverished region of northern Ghana around the World Health Organizations (WHO) six pillars combined with community engagement, leadership development and grassroots political support, the program organized a program of training and action focused on strategies for saving newborn lives and community-engaged emergency referral services. This paper analyzes the effect of the GEHIP program on child survival. Methods Birth history data assembled from baseline and endline surveys are used to assess the hazard of child mortality in GEHIP treatment and comparison areas prior to and after the start of treatment. Difference-in-differences (DiD) methods are used to compare mortality change over time among children exposed to GEHIP relative to children in the comparison area over the same time period. Models test the hypothesis that a package of systems strengthening activities improved childhood survival. Models adjusted for the potentially confounding effects of baseline differentials, secular mortality trends, household characteristics such as relative wealth and parental educational attainment, and geographic accessibility of clinical care. Results The GEHIP combination of health systems strengthening activities reduced neonatal mortality by approximately one half (HR = 0.52, 95% CI = 0.28,0.98, p = 0.045). There was a null incremental effect of GEHIP on mortality of post-neonate infants (from 1 to 12 months old) (HR = 0.72; 95% CI = 0.30,1.79; p = 0.480) and post-infants (from 1 year to 5 years old) -(HR = 1.02; 95% CI = 0.55–1.90; p = 0.940). Age-specific analyses show that impact was concentrated among neonates. However, effect ratios for post-infancy were inefficiently assessed owing to extensive survival history censoring for the later months of childhood. Children were observed only rarely for periods over 40 months of age. Conclusion GEHIP results show that a comprehensive approach to newborn care is feasible, if care is augmented by community-based nurses. It supports the assertion that if appropriate mechanisms are put in place to enable the various pillars of the health system as espoused by WHO in rural impoverished settings where childhood mortality is high, it could lead to accelerated reductions in mortality thereby increasing survival of children. Policy implications of the pronounced neonatal effect of GEHIP merit national review for possible scale-up. GEHIP was convened to test the hypothesis that health systems strengthening at the district level causes childhood mortality decline. Testing this hypothesis required longitudinal observation of organizational change and linked data on parental health seeking behavior and childhood mortality outcomes. Survey research was applied with cluster sampling to gauge changing access to health facilities over time, due to the GEHIP focus on expanding CHPS coverage. A baseline cluster survey was repeated at the endline, providing for the longitudinal documentation of expanding service operations by linking information on proximity to hospitals and clinics with monitoring data recording changes in the coverage of CHPS. Since impact of health care varies by age, the analysis took into consideration the age of the child, as well as ways in which the system at each level was changing relative to the exact age of each sample child as time progressed. GEHIP was initially implemented in three districts of the UER: Builsa, Bongo, and Garu-Tempane (Fig 1). Seven other UER districts served as a project comparison area. At the onset of the project, Builsa was split by an act of Parliament into two districts (Builsa North and South), making four treatment districts, in all.(see Fig 1). The 11 project districts rank among the poorest 5% of Ghana’s districts, each with economies that are dominated by subsistence agriculture. According to the Ghana Statistical Service (GSS), per capita income for these districts is about a quarter that of Ghana, ranking equivalently with the districts of the Upper West Region as the two most impoverished regions of Ghana [33]. Against this backdrop of profound economic adversity, the region is also health service deprived. Although Bongo and Builsa-North have hospitals, other districts in the region rely upon fragile and incomplete referral services or upgraded sub-district health centers for hospital care. There is a regional hospital in Bolgatanga, but apart from obstetrical care, specialized medical care of any kind is not available in the UER. Where the UER has registered progress, however, is with its implementation of community-based primary health care. Where coverage of the program has been lacking, interim facilities are often available, a strategy that has become more prominent in the GEHIP era. Thus, while tertiary health care is poorly developed, community-based primary health care has become more accessible in recent years, providing access to basic curative and preventive health services for children. Birth histories and corresponding information about deaths among children ever born were collected during the interviews of all women resident in sample households aged 15 to 49. Baseline survey interviews of 5511 women of reproductive age, out of an estimated sample of 6000, yielding an achieved sample of 91.8 include survival histories of 7410 children ever born who were ever 60 months of age or less during the five year period prior to each survey. Correspondingly, 5914 out of a targeted sample of 7588 women were interviewed in the endline, yielding a 76% achieved sample with survival histories of 7044 children ever born who were ever 60 months of age or less five years prior to the survey. Sampling was performed using a two-stage cluster design. In the first stage for the baseline, 66 clusters were apportioned among district census enumeration areas proportional to size using population projections based on the 2010 population and housing survey [34, 35]. In the second stage, random household selection proceeded within each cluster proportional to enumeration area size until the target sample total of 6000 women of reproductive age were selected. At the endline, the baseline surveys were reused to establish longitudinality of GEHIP exposure observation. However, since relisting and stage two resampling was pursued, GEHIP is a panel at the cluster level only. Interviews were conducted in the prevailing local language of sample households. For the purposes of the study, a live birth was defined as one in which the child cried or showed signs of life at birth such as pulsation of the umbilical cord or definite muscle movement. Crude annual estimates of under-5 mortality were calculated and compared to national estimates from the Ghana DHS over the same period [36–39]. Childhood survival was assessed for all children under 60 months of age. Observation of children was censored at 60 months of age or by the survey date. Analysis time was age of life in months. Since the mortality hazard followed a different pattern for neonates, post-neonate infants, and post-infants, the proportional hazard assumption was violated. This was addressed by introducing a categorical time interaction into our model to provide separate mortality hazard ratio estimates for (1) neonates in the first one month of life, (2) post-neonatal infants from age one month through 11 months of age, and (3) post-infant children from 12 months through 59 months of age. Although a separate hazard function was estimated for neonates, defined in days of age, results were identical to those produced by the age in month model. Covariates arrayed in Table 1 were incorporated in the full model [40–42]. Maternal variables included mother’s age at birth, religion, literacy, occupation, parity, wealth, marital status, and polygamy. Childhood characteristics included birth order, sex, and gestational age. Analyses incorporate an estimate of access to hospitals or sub-district health centers by measuring household distance to the nearest such facility via global positioning methods. Household wealth quintiles were constructed using principal components analysis (PCA) of discrete asset indicators [43] that defined access to sanitation and water, household possession of consumer durables (bicycle, radio, bicycle, motorbike, etc), and dwelling unit construction. Since these indicators were discrete variables, polychoric correlation matrix analysis was applied [44]. The principal component explained 40.9% of the common variance. To permit estimation of difference-in-differences (DiD) effects, longitudinal observation of clusters was combined with sampling within baseline clusters for the endline survey. Intra-cluster correlation between children of households in the same enumeration area was accounted for using robust standard errors via the sandwich estimator. With only four treatment districts and seven comparison districts, the number of districts was insufficient to provide a basis for randomization. However, for GEHIP to be relevant to policy makers, units of observation were required that conformed to units of programmatic decision-making represented by the district. In the absence of adequate statistical power at this organizational level, GEHIP embraced a quasi-experimental plausibility design. Owing to the policy relevance of this configuration, such designs have received growing attention in the implementation science literature, building upon the pioneering work of Campbell and Stanley (1966), and more recent advocacy of plausibility designs for implementation research [45–47]. Statistically rigorous responses to plausibility designs have been widely used with inference based on the Heckman difference in difference (DiD) concept [48] for the calculation of average treatment effects based on aggregate data [49]. A regression extension of the DiD concept is estimated for the present analysis that is based on individual observation [50]. In our mortality analysis, the DiD is a ratio of ratios comparing the ratio measuring mortality change in the treatment area over time with the corresponding ratio measuring mortality change in the comparison area over the same time period. Employing controls for pre- and post- treatment conditions, the GEHIP average treatment effect is estimated using a hazard model in which Gi is scored 1 if individual child i is resident in a GEHIP treatment area household and zero otherwise. Pit indicates period, where child i in month of life t is scored 1 if the month of life is July 2011 or after (the post-treatment time period) and zero otherwise. The DiD parameter is the interaction between P and G, δ ki, a parameter representing the net GEHIP incremental effect, relative to trends or areal differences that are unrelated to intervention, while also controlling for the kth maternal or household characteristic of child i. The overall GEHIP average treatment effect is given by the conditional hazard: β, γ, and δ are unknown parameters estimated by maximum likelihood. Background characteristics comprising the vector x comprised of C household indicators of distance to clinical care facilities and relative household economic status as well as maternal age, parity, educational attainment, and marital status, permitting estimation of K multivariate δ “nuisance” parameters that introduce control for imbalance (Table 1). Multiple imputation by chained equations was employed to account for missing values [51]. Greenland’s procedure for achieving model parsimony is used to estimate the final model, which is the reduced form of (1) [52–56]. This procedure introduces parsimony into the estimation of models that could otherwise acquire sparse data biases [54, 57–59]. The final reduced model was identified using Greenland’s recommended modeling strategy that combines a change-in-estimate approach with reduction of mean squared error. The final model excludes covariates included in the full model (1) that do not confound to the effect estimate and which, if removed, reduce the mean squared error of the effect estimate [54]. Ethical approval for the Ghana Essential Health Intervention Project (GEHIP) was granted by Ethical Review Committee of the Ghana Health Service, the Institutional Review Board (IRB) of the Navrongo Health Research Centre and the ethical review board of the Columbia University Medical Center, Mailman School of Public Health. A written inform consent was provided to study participants prior to their inclusion. Data collectors read a written informed consent form to participants in their preferred language and explained its content before participants who agreed to participate endorsed two copies of the form and a copy was given to the participant. This procedure was sanctioned by all three ethics committees that approved of the study to be conducted. All protocols were followed to ensure confidentiality during data collection, analysis and reporting of findings.