This study used the IFLS, a longitudinal household survey, to estimate the contribution of mothers’ observed and unobserved characteristics to their risk of stillbirths and neonatal mortality. This study is innovative because it accounts for statistical dependence between births to the same mother using complete pregnancy histories from household survey data, allowing for evaluations of stillbirth and neonatal mortality in the same population.
IFLS is a population-representative, longitudinal survey of 13 (of 27) provinces consisting of five panels: 1993, 1998, 2000, 2007–08, and 2014–2015. In 1993, IFLS covered an initial sample of 7,224 households spread across six islands. This sample was selected to maximize the representation of Indonesia's cultural and socioeconomic diversity, representing approximately 83% of the Indonesian population at the time, while being cost-effective . Three hundred twenty-one enumeration areas (EAs) or clusters were chosen from an existing sampling frame of about 60,000 households. Urban clusters and those in smaller provinces were over-sampled to facilitate urban–rural and Javanese to non-Javanese comparisons. Field teams randomly selected 20 households in each urban cluster and 30 households in each rural cluster for inclusion. In the subsequent panels, the goal was to relocate and re-interview all households interviewed in the previous panels, including people that moved to another IFLS province . Individuals that split off from the original IFLS household but remained in the 13 provinces were also interviewed but given new household IDs.
IFLS included a household survey, with adult and children’s questionnaires, as well as a community and facility survey. The adult questionnaires included modules on education, marriage, migration, employment, health status, utilization of health services, individual and household assets, fertility and contraception, infant feeding practices, as well as proxy data on household members who were away. All ever-married women were asked questions about marriage, contraceptive use, pregnancies and outcomes, use of antenatal care, children ever born, infant feeding, and the status of child survival.
Due to missing date and gestation data, especially for stillbirths, birth histories reported in 1993 and 1998 were excluded. The newer datasets were much less likely to be missing gestational age for each birth, which was important for defining stillbirths, as well as the year of birth which was important for removing duplicates across panels. Births reported in the 2000, 2007, and 2014 panels were extracted to produce birth histories for mothers for births ending in stillbirth or live birth. All self-reported miscarriages were deleted, but miscarriages with gestation over 7 months were recoded as stillbirths. Similarly, self-reported stillbirths with gestation under 7 months were recoded as miscarriages and excluded. Duplicates reported in subsequent panels were removed based on year of birth and outcome. Births that were missing gestation and other covariates were excluded (l.0%). For the analysis, only multiparous women were included in the final sample. Data were right-censored since some mothers had not completed their reproductive span at the time of the final survey. The data were also panel unbalanced, meaning that mothers had different numbers of births in the sample.
Primary risk factors
Birth histories of mothers were used to calculate two primary risk variables for each birth: any history of stillbirth and any history of neonatal death. For stillbirth history, a binary variable was constructed with a value of 1 if the mother had a prior stillbirth and 0 if not. For neonatal mortality history, a binary variable was constructed with a value of 1 if the mother had a prior live birth that resulted in neonatal death and 0 if not.
There were two outcomes evaluated in separate models: stillbirth, defined as a pregnancy with seven months (or 28 weeks) or longer gestation that ended in a fetal death, and neonatal mortality, defined as death in the first 28 days of life per World Health Organization (WHO) recommendations . For neonatal mortality, a one-month endpoint was used to capture deaths that were reported in months, which was a majority of the sample. Gestational age was self-reported by the woman during the survey.
Covariates were measured at the level of birth, mother, and household. For example, pregnancy duration and parity were specific to birth, education was specific to the mother, and urban residence was specific to the entire household.
Number of births to date
Calculated as the number of births (live and still) that a woman reported across the full birth history. Each birth was assigned a value based on this history, grouped into three categories: none, 1–2 birth, and 3 or more births.
Age of mother at birth
The age of a mother at birth was calculated as the difference in years between her birth year and the year of each one of her births. The age of the mother at the time of the survey was not used in the analysis. Any ages below 15 years were excluded.
Education level was measured at the time of the survey. It was introduced as a categorical variable in the following four groups: none, elementary, junior high, senior high, college, or religious/vocational. ‘Religious’ education included Madrasah Islamic schools.
Whether a woman lived in an urban or rural cluster was based on her IFLS cluster, which was listed as urban or rural in the sampling frame .
Additional covariates that would have been useful but excluded because they were missing data for a significant proportion of women included access to antenatal care, place of birth, and skilled care at birth. Data on maternal body mass index (BMI) and household per capita expenditures were also available in the survey but only measured at the time of the interview. For many births, these would be far removed from the year of birth and were not used to avoid spurious correlations.
This study uses Random Effects (RE) models to estimate the underlying unmeasured subject/mother-specific risk of having either a stillbirth or a neonatal death, given observed group-level risk factors and covariates among women with at least two births in the sample. RE models describe variation in mother-specific responses according to unmeasured characteristics. Using the RE approach allows the mother-birth relationship to have mother-specific effects assuming this relationship varies by mother but is similar for births to the same mother. In RE models, all unmeasured factors are assumed to contribute to heterogeneity regardless of whether they can ever be observed. This variance is allowed to be random and uncorrelated between mothers and observed variables and allows for the estimation of the effect of maternal independent variables that remain constant across births (such as education and urban/rural residence) regardless of the number of observations in each cluster. The combined variance of unobserved heterogeneity is output as rho in the model statistics . It can be interpreted as the variation in the odds of having a stillbirth (or neonatal death in a neonatal model) that was not explained by the observed variables in the model.
Model building and standard error adjustment
The model-building strategy was based on the evaluation of biological and socioeconomic factors from relevant public health literature. Biological factors included age of the mother at birth, parity, female birth (neonatal model only), and any prior history of a stillbirth or neonatal death. Socioeconomic factors included urban or rural residence and education level. Some independent variables' exposure timing can produce causal estimates, while others can only be interpreted as correlates. In particular, the age of the mother at birth, parity, and history of stillbirth or neonatal death could be established as occurring before birth. Therefore, estimates for these variables can be interpreted as predictive of the outcome. Mothers’ education level and urban residence could only be correlated with the outcome as they were measured at the time of interview. However, they could not be evaluated as predictive or causal as it was not possible to determine whether these characteristics were true at the time of birth. Male birth would be associative but not necessarily causal in the absence of other information but established research has shown higher mortality among male newborns .
Standard errors were adjusted for robust estimates for maternal clustering. Clustering within communities was not adjusted to avoid adding greater complexity to an already complex dataset and survey design. Shared community-level factors that impact stillbirth and neonatal mortality that may exist would be included in maternal clustering estimates, introducing bias in interpreting maternal effects. Hierarchical clustering affects standard errors but does not influence effect estimates. All analyses were performed in Stata statistical software version 17 . Statistical significance was set at 0.05 and p-values were 2-tailed. Calculations were weighted to account for survey design and attrition over time and to adjust estimates for the Indonesian population in the panel year.