Analysis of complete birth histories
We extracted child-level data on date of birth, survival status, age at death (if applicable), and sampling weights from surveys with complete birth histories. The data for each child were expanded to encompass each month that the child started alive prior to reaching age 5 or the date of the survey, whichever occurred first. All child-months were subdivided by calendar year and into six age groups (month 0, months 1–11, year 1, year 2, year 3, and year 4) and the monthly probability of death in each age-year group was calculated as the weighted proportion of child months which ended in death, where the weights were the survey sample weights. We then derived U5MR for each calendar year as:
$$ U5{MR}_{j,t,s}^{(CBH)}=1-{\prod}_a{\left(1-{q}_{j,t,s,a}\right)}^{n_a} $$
where
qj, t, s, a is the monthly probability of death in area
j, year
t, age group
a, source
s; and
na is the number of months in age group
a.
We estimated \( {N}_{j,t,s}^{(CBH)} \), the number of births associated with a given \( U5{MR}_{j,t,s}^{(CBH)} \) estimate based on the number of child months contributing to that estimate. Specifically, the number of child months in area j, year t, and source s (summed across the six age groups) was divided by the mean number of months children in source s lived prior to death, reaching age 5, or the time of survey, whichever occurred first. We then estimated \( {Y}_{j,t,s}^{(CBH)} \), the number of deaths associated with each U5MR estimate, by multiplying the estimated number of births by the estimated U5MR:
$$ {Y}_{j,t,s}^{(CBH)}=U5M{R}_{j,t,s}^{(CBH)}\cdotp {N}_{j,t,s}^{(CBH)} $$
Analysis of summary birth histories
We extracted woman-level data on the number of children ever born and the number of children died from surveys and censuses with summary birth histories. We then applied the combined summary birth history method described by Rajaratnam et al. [
15] in order to generate preliminary estimates of U5MR by area and year from each survey. This method requires several inputs, indexed by mother’s age or reported time since first birth, including: number of women, total children ever born, total children died, and mean children born (per woman). Women-level sample weights were used to generate weighted estimates of all input parameters. The output of these summary birth history methods are annual estimates of U5MR for approximately 25 years preceding the date of the survey or census.
As an intermediate step in applying the summary birth history methods described by Rajaratnam et al. [
15], all reported births are distributed on an annual basis to the years preceding the survey or census using empirical distributions of time since birth indexed by mother’s age and the reported number of children at the time of survey. The resulting annual estimates of births (
\( {N}_{j,t,s}^{\ast (SBH)} \)) were taken as the starting point for estimating an effective sample size for each summary birth history estimate.
Summary birth history estimates are subject to both sampling error and model error, and we wanted to reflect this in our estimates of the effective number of births and deaths associated with each U5MR estimate derived from a summary birth history. To approximate the sampling variance of \( U5M{R}_{j,t,s}^{(SBH)} \), we assumed the number of children that die is approximately binomially distributed:
$$ {\sigma}_{j,t,s\ \left[ sampling\right]}^2=\frac{U5M{R}_{j,t,s}^{(SBH)}\cdotp \left(1-U5M{R}_{j,t,s}^{(SBH)}\right)}{N_{j,t,s}^{\ast (SBH)}} $$
To approximate the model variance, we utilized the results of the validation exercise reported by Rajaratnam et al. [
15] Five-fold cross-validation was used to assess the performance of the summary birth history methods in reproducing estimates derived from complete birth histories. We calculated the variance of the residuals from this comparison for each year prior to survey – i.e., the difference between the summary birth history estimate and corresponding complete birth history estimate, on a probability scale – as an approximation of the error introduced by using the summary birth history data and methods compared to the complete birth history data and methods. For each estimate of
\( U5M{R}_{j,t,s}^{(SBH)} \), we used this variance, matched for appropriate number of years prior to survey, as
\( {\sigma}_{j,t,s\ \left[ model\right]}^2 \). We assumed that the model error and sample error were independent, and calculated the total variance for each estimate of
\( U5M{R}_{j,t,s}^{(SBH)} \) as
$$ {\sigma}_{j,t,s\ \left[ total\right]}^2={\sigma}_{j,t,s\ \left[ sampling\right]}^2+{\sigma}_{j,t,s\ \left[ model\right]}^2, $$
and calculated the corresponding effective sample size \( {N}_{j,t,s}^{(SBH)} \) again assuming that the number of children who die is approximately binomially distributed:
$$ {N}_{j,t,s}^{(SBH)}=\frac{U5M{R}_{j,t,s}^{(SBH)}\cdotp \left(1-U5M{R}_{j,t,s}^{(SBH)}\right)}{\sigma_{j,t,s\ \left[ total\right]}^2} $$
This procedure was carried out both at the national level as well as each subnational area at the finest level available in given survey or census, and the resulting values of \( {N}_{j,t,s}^{(SBH)} \) for all subnational areas were scaled to sum to the estimated value of \( {N}_{j,t,s}^{(SBH)} \)for the country as a whole. Finally, we calculated the effective number of deaths by multiplying the estimated U5MR by the effective number of births:
$$ {Y}_{j,t,s}^{(SBH)}=U5M{R}_{j,t,s}^{(SBH)}\cdotp {N}_{j,t,s}^{(SBH)} $$
Small area models
We started with the following hierarchical generalized linear model defined for data stratified by area (department, district, or sub-district, depending on the country), year, and data source (a single survey or census in a particular area):
$$ {\displaystyle \begin{array}{c}{Y}_{j,t,s}\sim \mathrm{Binomial}\left({p}_{j,t,s},{N}_{j,t,s}\right)\\ {}\mathrm{logit}\left({p}_{j,t,s}\right)={\beta}_0+{u}_{0,j}+{\sum}_{i=1}^5\left({\beta}_i+{u}_{i,j}\right)\cdot {S}_i(t)+{\gamma}_s\end{array}} $$
where
Nj, t, s and Yj, t, s are the number of births and deaths respectively in area j, year t, and source s (equivalent to \( {N}_{j,t,s}^{(CBH)} \) and \( {Y}_{j,t,s}^{(CBH)} \) or \( {N}_{j,t,s}^{(SBH)} \) and \( {Y}_{j,t,s}^{(SBH)} \), depending on the birth history method used to analyze source s);
pj, t, s is the underlying U5MR in area j, year t, and source s;
β0 and u0, j are the country-level fixed intercept and the area-level random intercept, respectively;
Si(
t) is basis
i of a natural cubic spline [
16] with four equally-spaced interior knots evaluated at time
t;
βi and ui, j are the country-level fixed slopes and area-level random slopes on Si(t), respectively;
and γs is a source-level random intercept.
We then added a second component which allows us to incorporate data defined for other geographic levels (i.e., higher level or historical administrative units):
$$ {\displaystyle \begin{array}{c}{Y}_{k,t,s}\sim \mathrm{Binomial}\left({p}_{k,t,s},{N}_{k,t,s}\right)\\ {}\mathrm{logit}\left({p}_{k,t,s}\right)=\mathrm{logit}\left({\sum}_{j\in k}\frac{P_{j,t}}{P_{k,t}}\cdot {p}_{j,t}\right)+{\gamma}_s=\mathrm{logit}\left({p}_{k,t}\right)+{\gamma}_s\end{array}} $$
Where pk, t, s, Nk, t, s, Yk, t, s are defined analogously to pj, t, s, Nj, t, s, Yj, t, s but for some area k made up of multiple area j’s (e.g., a province, containing multiple districts). pj, t, the underlying U5MR in area j and year t is defined analogously to pj, t, s above, but with γs set to 0, and pk, t, the true prevalence in area k and year t, is given by the population (P) weighted average of pj, t across all areas j contained within area k.
The random effect terms
u0, j and
ui, j (
i = 1–5) were assigned conditional autoregressive priors as described by Leroux et al. [
17] These priors allow for spatial smoothing based on the neighborhood structure of the areas being modeled, specifically by assuming that the prior mean for a given area is a function of the values in neighboring areas. The full conditional distribution implied by this prior is:
$$ {u}_j\mid {u}_{-j},{\sigma}^2,\rho \sim \mathrm{Normal}\left(\frac{\rho {\sum}_{m\sim j}{u}_m}{n_j\rho +1-\rho },\frac{\sigma^2}{n_j\rho +1-\rho}\right) $$
where
m~
j indicates that area
m is a neighbor of (i.e., shares a boarder with) area
j and
nj is the number of neighbors of area
j. The two hyperparameters for each random effect were estimated as part of the model fitting process:
σ2 determines the overall amount of variation and
ρ, which varies between 0 and 1, determines the degree of spatial autocorrelation. Lower values of
ρ indicate less spatial relatedness, while higher values of
ρ indicate a high degree of spatial relatedness; at the extremes, this prior reduces to a Normal(0, 1) prior when
ρ is 0 and to an intrinsic conditional autoregressive prior [
18] where the prior mean is equal to the mean of all neighbors when
ρ is 1. Normal(0, 10) hyper-priors were specified for logit-transformed
ρ and half-Normal(0, 1) hyper-priors were specified for
σ for all random effects. Common
σ2 and
ρ parameters were estimated for all random effects on the spline bases,
ui, j (
i = 1–5). Posterior estimates of these hyperparameters are listed in Additional file
1.
Models were estimated separately for each country. All models were fit using the TMB package [
19] in R version 3.2.4 [
20]. We extracted point estimates and the variance-covariance matrix for logit(
pj, t) and used these to generate 1000 draws of
pj, t by drawing from a multivariate normal distribution and then inverse-logit transforming each draw. These draws were scaled to match existing national-level estimates of U5MR from the Global Burden of Disease (GBD) Study: [
3] for each year
t and each draw, we calculated the ratio of the national estimate from the GBD to the national estimate derived from population-weighting
pj, t and multiplied
pj, t by this ratio. Finally, we calculated point estimates of
pj, t from the mean of these draws and the lower and upper bounds of the 95% uncertainty interval from the 2.5th and 97.5th percentiles, respectively. Relative change in
pj, t over time was also calculated for each draw, and 95% uncertainty intervals were derived from the 2.5th and 97.5th percentiles.