Page 53 - IJPS-3-2

Multilevel analysis of infant mortality and its risk factors in South Africa

inequality as measured by Gini index, are identified at municipal level, while HIV prevalence rate is at province level. All
these four variables were classified into two categories: lower and higher magnitude of the respective measures. The lower
and higher values dictate that the respective quantity in the area is less than and greater than the national estimate. For
instance, about 49% of the children live in municipalities where the level of income poverty is higher than the national
poverty head count ratio of 41%. Note also that among the child level variables, age of the child is an indicator variable
showing whether the child has age of less than one month (neonatal) or not.

2.3 Multilevel Models
Multilevel analysis is a suitable approach to take into account community level contexts at different levels, like at
municipal and province levels, as well as individual subjects. A three-level random intercept logistic regression model was
considered where the first level is children born 12 months before the census, whereas the municipalities and provinces
in which the children live are the second and third levels respectively. Let be the probability that child i living in
municipality j and province k died before reaching age one. Then, the three-level random intercept logistic regression
model in question with the predicator variables described above can, therefore, be expressed as

[Level 1]

[Level 2]

[Level 3]

where , , and the notations of the independent variables are as given in Table 2. The
coefficients , called fixed effects, measure the impact of the corresponding predicator variable on the log
of odds of death, whereas , the random intercept, measures the combination of municipal and provincial level effects
as defined in the second and third level of the model. Unlike ordinary logistic regression, there are two types of residual
terms, and , defined at level 2 and level 3 respectively and assumed to be normally distributed with mean zero and
constant variance. Bayesian approach with Markov Chain Monte Carlo (MCMC) was implemented to the parameters of
the above model. Further information regarding methods of parameter estimation is given in the appendix.
3. Results

3.1 Descriptive Statistics of Variables

The descriptive statistics of all individual, municipal and province level variables chosen for the analysis including the
bivariate odds of infant death are shown in Table 2. It shows that some of the variables, such as race and education of the
mother, living standard, birth order and HIV prevalence contribute to greater odds of death of the infant than others.

3.2 Multilevel Model Outputs
The final results of the regression are shown in Table 3. All parameter estimates were measured on the log-odds (logit)
scale. In order to make more specific and meaningful inference about the effect of the risk factors on the infant mortality,
the odds ratios (ORs) were given corresponding to each coefficient estimate in the same table. Note that among the
independent variables, proportion of poor people, income inequality and mean years of mother’s education were measured
at municipality level, whereas HIV prevalence rate was computed at province level. All these four variables were
dichotomised as higher and lower values of the respective quantities.
All coefficients of the living standard dummy variables are negative and their 95% confidence intervals exclude zero.
Compared to infants who were in the first quintile of living standard, those who were in the second to fifth quintiles
had 6%, 7%, 14 % and 24% lower odds to die, respectively. Likewise, the income poverty has a positive and significant
coefficient, entailing that children living in a household whose members earned a per capita income of less than the South
African poverty line were more likely to die than those who were above the poverty line.
Most of the municipal level indicator variables are significant, which implies that the level of poverty, women education
and inequality of the municipality affected the survival status of infants. An infant was more likely to die in a highly poor
and more unequal municipality compared to municipalities where the levels of poverty and inequality were lower after
controlling for other risk factors. Considering the magnitude of the impact, it seems that the income inequality mattered
more for infant mortality risk than the size of poverty in that more unequal municipalities were associated with 13%
higher odds ratio of infant death than less unequal municipalities, whereas municipalities where poverty was high were

46 International Journal of Population Studies | 2017, Volume 3, Issue 2