gives a summary of the selected covariates across the three countries. There are evident disparities by place of residence for all three countries, with rural children slightly disadvantaged in mortality. The same picture was observed by age, with children less than 1 year disadvantaged compared to older children, with the proportion dying diminishing with increasing age. Children without a living mother or father were likely to die in their first 5 years of life. However, there was no clear pattern in relation to the shelter or electronics indices, or with religion or sex of the child. Similar results were obtained in the bivariate analyses presented in , , .
Fixed effects for Rwanda child survival
Fixed effects for Senegal child survival
Fixed effects for Uganda child survival
In , model selection values are given for the discrete-time survival models with different specifications of the covariates for the three countries. For all the three datasets the models which combined fixed and random effects were better than those that did not combine effects, indicating the importance of both sets of factors at explaining child survival. For Rwanda data, the best model was model M3b, which combines fixed effects at individual and household levels and random effects at district and provincial levels. The DIC for model M3b was 7756.1 compared to the nearest model, M1b, with DIC=10240.3. Moving to Senegal data, again the model that combined fixed and random effects produced the best fit (model M3b). Model M3b had a DIC=14711.1 which is smaller compared to model M3a (DIC=15361.5; ). Similar results are obtained for the Uganda data with model M3b emerging as best fit, although model M2d was indistinguishable (see ).
Model comparison values based on Deviance Information Criterion (DIC) for the models
, , present estimates of fixed risk factors resulting from the models with the best fit. For Rwanda (), there was an overall decrease of risk of a child dying in the first 5 years of life (HR=0.04, 95% CI 0.02 to 0.09). Children in urban areas were less likely to die than those in rural areas (HR=0.79, 95% CI 0.73 to 0.83). The relationship of child dying and household electronic assets was nonlinear. At level 2 compared to level 1, the risk was higher with HR=1.29 (95% CI 1.16 to 1.51), while at level 3 we observed a lower risk with HR=0.64 (95% CI 0.46 to 0.89) and this is reversed at level 4 with HR=1.31, 95% CI: 1.16 to 1.51. For the shelter index, the risk was reduced at lower levels and increased at higher levels of the index, although this relationship was not significant at p<0.05. It is interesting to note that a child with a living mother and father had a reduced risk of dying (). Children up to 1 year of age were at increased hazard relative to those aged 5 years or older. At less than 1 year of age the log hazard was 0.77 (95% CI: 0.71 to 0.83), while at 1 year the log hazard was 0.48 (95%CI 0.42 to 0.55). As age increased, the hazard reduced. For example, those aged 2–4 years the log hazard was −0.02, −0.22 and −0.46, respectively.
The spatial variability of risk of dying is shown in , with log hazard ranging between −8.23 and 3.14. There were a number of areas that were associated with increased risk of death compared to the overall mean. These areas are identified by the right map, with a white colour and appear in the south, west and at the centre of the country. There are also areas of reduced risk shown by a black colour.
Figure 1 Left: structured spatial effects, at district level in Rwanda, of child surviva (model M3b). Shown are the posterior means. Right: corresponding posterior probabilities at 80% nominal level, white denotes region regions with strictly positive credible (more ...)
In we present results for Senegal. Overall the risk of death decreases with HR=0.024 (95% CI: 0.015 to 0.045). The risk significantly varied with ownership of dwelling unit, electronic assets, sex and age of the child. Ownership of a dwelling unit was associated with increased risk (HR=1.19, 95%CI 1.09 to 1.28) compared to those households without a dwelling unit. Male children were more likely to survive the first 5 years compared to female children (HR=0.88, 95% CI 0.85 to 0.91). The risk of dying was positively associated with all ages, however, this risk decreased with age, ranging from 2.48 at age less that 1 year to 1.13 at age of 4 compared to those aged 5 years or more. For ownership of electronic assets, the risk was higher for those at the lowest level (level 1) and decreased with increasing electronic assets, although the relationship was marginally significant at p<0.1 for levels 2, 3 and 4(results not shown). Nevertheless, the results were significant, at p<0.05 for the level 5 category when compared with those at level 1 (HR=0.93, 95% CI 0.87 to 0.99). Turning to the spatial distribution of risk in , there was substantial variation, with estimates of log hazard ranging from −3.44 to 5.87 (left map). The right map defined areas associated with significantly high risk (shaded white) as well as those of significantly low risk (black shading). We could not identify a clear pattern to the risk by region.
Figure 2 Left: unstructured spatial effects, at district level in Senegal, of child survival (model M3b). Shown are the posterior means. Right: corresponding posterior probabilities at 80% nominal level, white denotes regions with strictly negative credible intervals, (more ...)
Results for Ugandan data are given in . Again the overall risk of death decreases (HR=0.011, 95% CI 0.006 to 0.017). Risk factors associated with under-five mortality were identified to be number of under-five children in the household, marital status, education level of mother, ownership of electronic assets and shelter characteristics. Families with fewer under-five children predisposed children to a high mortality risk compared to those with 4 or more children (HR=4.18, 95% CI 2.92 to 5.81), while those with 1–3 children had a reduced risk (HR=0.54, 95% CI 0.39 to 0.73). Being married also appeared to confer increased risk of a child dying compared to those with children of single mothers (HR=2.10, 95% CI 1.78 to 2.42). Our results showed that education level of the mother matters when it comes to child survival. Children with mothers who had no formal education or only lower primary education were more likely to die than those with tertiary education (HR=1.48, 95% CI 1.28 to 1.70 and 1.21, 95% CI 1.05 to 1.45, respectively).For those with secondary education, the risk was lower relative to those with tertiary education (HR=0.57, 95% CI 0.36 to 0.81). In relation to electronic assets, the risk was non-linear with increasing risk at level 2, reduced risk at level 3 and increased risk again at levels 4 and 5, compared to level 1 (). We observed that the risk was lower at levels 2 and 3 of the shelter index and increased at level 4 and 5 relative to level 1. Nevertheless, the only significant difference was observed at levels 3 and 4 (1.32 and 1.12, respectively).
The geographical variation in risk is shown in . Estimates ranged from −0.61 (low risk) to 0.73 (high risk). See left plot. However, the significance map (right map) indicates that areas of high risk are in the south-west and north-west while those of low risk are in the north-east and centre-east. Notably Kampala district showed a significantly reduced risk.
Figure 3 Left: structured spatial effects, at district level in Uganda, of child survival (model M3b). Shown are the posterior means. Right: corresponding posterior probabilities at 80% nominal level, white denotes regions with strictly negative credible intervals, (more ...)
The unstructured spatial effects at provincial level were also fitted. shows caterpillar plots for the three countries at province and county level. No single province or county residual was significantly above or below zero indicating no difference in risk of death between provinces or counties in the three countries. However, there was clear variation in the risk of death, for example, in Rwanda there are four provinces with an estimated lower risk of death while six provinces have an estimated risk in the higher direction. For Senegal, there were four provinces with a reduced risk, and eight with estimated high risk. In Uganda, about a 100 counties were estimated to have a lower risk of child mortality, while another 70 had a high risk ().
Left: unstructured spatial effects, at province level in Rwanda and Senegal and at county level in Uganda, of child survival (model M3b). Shown are the posterior means and corresponding error bars at probabilities at 80% nominal level.