Data sources
The first primary data source, the Intercensal Population Survey (SUPAS), is conducted every 10 years at the mid-point between decennial population censuses [2]. The 2015 SUPAS collected a wealth of household and individual data relevant to the present article. Sample respondents were chosen using a stratified, two-stage cluster sampling scheme. The primary sampling unit was the census block, a geographically defined unit containing 80-120 households. A total of 40,728 census blocks were randomly chosen at the first stage of sample selection with probability proportional to estimated size (PPES) and allocated to provinces proportionally to provincial population size. As one priority of the 2015 SUPAS was to collect data on maternal mortality, a special sampling scheme was used to select households at the second stage of sample selection. In each selected census block, a sample of 16 households was chosen by first selecting with certainty households that reported maternal deaths in the previous five years (maximum of 8 households), and then selecting a random sample of the remaining households of size needed to yield a total sample of 16 households per census block. The definition of a maternal death used was women 15 to 54 years of age who were pregnant at the time of death or who died within two months post-delivery.
An important limitation of the maternal mortality data collected in SUPAS data should be noted. Because the information collected on maternal deaths was limited to place of death; age at the time of death; and whether the death occurred during pregnancy, during delivery or post-delivery, we lack information on risk and protective factors preceding reported deaths as would be required in order to undertake analyses with individual women as the unit of analysis. Instead, we undertook an ecological analysis in which we related maternal mortality risk and protective factors measured at the community level to the ratio of maternal deaths to live births in each census block during the five years 2010-2015.
The second main data source, the 2014 Village Potential Statistics (PODES) [16], is a census of villages that provides detailed information on the roughly 65,000 villages in Indonesia and the sub-districts and districts in which they are located. Three types of questionnaires were used: village-level, sub-district-level and district/city-level. Data were collected on population, environment, housing and settlements, educational facilities, social and cultural activities/institutions, recreation and entertainment, health facilities, nutrition and family planning, transportation and communication, land and its use, economy, security and information on village heads. Our main interest in the PODES data was information that described the supply environment for maternal health services.
Operationalization of variables
The following variables were extracted from SUPAS for all ever-married women of reproductive age: age, parity, number of births in the previous five years, age and parity at time of all births in the previous five years, contraceptive use at the time of data collection, highest educational attainment, household economic status, and residence. We created a set of census block-level indicators that measured community-level maternal mortality risk and protective factors. These included the proportion of ever-married women of reproductive age who were using contraceptives; proportion of births at elevated risk due to too young or too old maternal age (i.e., under age 20 or over age 40); proportion of births at elevated risk due to high parity (i.e. parity four and above); proportion of women with primary-level education or lower, and proportion of women residing in households that were in the lowest two household wealth quintiles (classified as “very poor” and “poor”). Urban-rural location and island group (Java-Bali vs. others) were also included as variables to capture unmeasured differences in development and sociocultural factors. The operational definitions for all variables are provided in Table 1.
Table 1: Operational definitions of variables used in the analyses
Variable
|
Definition
|
Contraceptive prevalence rate
|
The proportion of women using a contraceptive method in a given census block = Number of ever married women using contraception divided by the number of ever married women.
|
Contraceptive prevalence category
|
Coded 0 (Low) if contraceptive prevalence was less than 40%, 1 (Middle) if between 40% and 59%, 2 (High) if 60% or above
|
Proportion of high-risk births: maternal age
|
Coded 0 (Low) if proportion of women below age 20 or above age 40 in census block were less than 5% and 1 (High) otherwise.
|
Proportion of high-risk births: parity
|
Coded 0 (low) if a census block had less than 5% of births to parity 4 or above, coded 1 (Middle) if a census block had ³5% to 25% of births to parity 4 or above, and coded 2 (high) if a census block had >25% of births to parity 4 or above.
|
Proportion of low educated mothers
|
Low educated mother is defined as women of reproductive age having junior high school or less education. or less. Coded 0 (High educated) if a census block had less than 1% low-educated mothers; Coded 1 (Middle educated) if 1-40%; and Coded 2 (Low educated) if > 40%.
|
Proportion of low socio-economic households
|
Household socio-economic status is divided into five quintile categories: (1) very poor, (2) poor, (3) middle, (4) rich, and (5) very rich. Low socio-economic household is defined a household in category ‘very poor’ or ‘poor’. Coded as 0 (Rich) if a census block had less than 1% low socio-economic households; Coded 1 (Middle) if the proportion of poor or very poor household was 1-49%; Coded 2 (Poor) if the proportion of very poor/poor households was 50% or more.
|
District hospital population density
|
Number of hospitals per 1,000,000 district population. Coded 0 (Low) if the block census district hospital population density was less than 5 per 1,000,000 population; Coded 1 (Middle) if more than 15 per 1,000,000 population; and Coded 2 (High) if 5-15 per 1,000,000 population.
|
Sub-district health center population density
|
Number of community health centers per 100,000 sub-district population. Coded 0 (Low) if the block census sub-district health center density was 5 or less per 100,000 population; Coded 1 (High) if more than 5 sub-district health centers per 100,000 population.
|
Sub-district physician population density
|
Number of physicians per 100,000 sub-district population. Coded 0 (Low) if the block census sub-district physician density was 8 or less per 100,000 population; Coded 1 (High) if more than 8 physicians per 100,000 population.
|
Village midwife population density
|
Number of midwives per 10,000 village population. Coded 0 (Low) if the block census village midwife density was 0.85 or less midwives per 10,000 population; Coded 1 (High) if more than 0.85 midwives per 10,000 population.
|
Village TBA population density
|
Number of traditional birth attendants per 10,000 village population. Coded 0 (Low) if the block census village TBA density was less than 1 TBAs per 10,000 population; Coded 1 (Middle) if 1-4 TBAs per 10,000 population; Coded 2 (High) if more than 4 TBAs per 10,000 population.
|
Island group
|
Coded 0 if the census block was located on the islands of Java or Bali; Coded 1 otherwise.
|
Urban-rural
|
Coded 0 if the block census was located in an urban area; Coded 1 otherwise.
|
Information on contraceptive use in SUPAS is limited to contraceptive status at the time of SUPAS data collection. This required the assumption that community contraceptive prevalence at the time of SUPAS data collection reflected census block differences in contraceptive practice during the five-year period prior to the SUPAS. At the national level, contraceptive prevalence was stagnant during this reference period (61% in 2007 and 64% in 2017 [17]. While there is certain to have been some variability in the rate of change in contraceptive prevalence across subnational geographic units, we postulate that these were unlikely to have sufficiently dramatic to invalidate the assumption that community contraceptive prevalence measured in the 2015 SUPAS provides a valid proxy measure of relative levels of community contraceptive use during the 2010-2015 period.
PODES data were used to construct a series of variables describing the local supply environment for maternal health services in the form of population densities. These included the number of hospitals in the district in which sample census blocks were located per 1,000,000 population, the sub-district density of public health centers and physicians per 100,000 population, and the village density of midwives and traditional birth attendants (TBA) per 10,000 population. On the basis of these densities, we classified the access of respondents in a given census block to each type of health system asset as being high, medium or low. Further details may be found in Table 1.
Statistical Analyses
To measure the net impact of contraceptive use on maternal mortality, we estimated a series of log-linear regressions with census block maternal mortality ratios (MMRs) in the five years preceding the 2015 SUPAS as the dependent or outcome variable. The unit of analysis in all regressions was census blocks (n=40,728). The sampling weights calculated by the Indonesia Central Statistics Bureau (BPS), which corrected for unequal probabilities of selection of households, were applied to the data during analysis. Because of the skewness of the dependent variable and the large number of census blocks with no maternal deaths, we used a natural log transform with a small constant (one) added to MMR in each census block as the dependent variable in the analyses; that is,
Log-linear model: Ýi = a + ∑bXi + ei
Where:
Ýi = Ln [1 + ((MMRi)*100,000)]
MMRi = (MDi / LBi) * 100,000
MDi = number of maternal deaths 2010-2015 in census block i,
LBi = number of live births 2010-2015 in census block i,
Xi = Vector of independent variables,
α and β are regression coefficients to be estimated, and
ei = error term for census block i.
Visual inspection indicated that the distribution of the transformed dependent variable was improved, though not yet normalized. However, the distributional assumptions underlying the regressions become less of an issue with large sample sizes [18]. With large sample sizes, commonly used test statistics (e.g., p-values) rather quickly approach zero, and thus solely relying on p-values can lead to overstating the practical significance of empirical results [19]. Accordingly, we base our interpretation of results on effect sizes and their confidence intervals.
In the regression analyses, we first assess bivariable associations between the variables enumerated in Table 1 and maternal mortality, then a multivariable model with all variables included, followed by the identification of the most parsimonious model statistically. The latter was accomplished by backward elimination [20-21]. In order to facilitate reader understanding of the results, we back-transform the log coefficients produced in the regressions for some of the key results in the Results and Discussion section of the article using the formula eβ-1 (where β is the regression coefficient in log form). In addition to estimating the main effects of, we also estimated a regression model that included an interaction term to test the proposition that the effects of contraceptive use would be larger in districts where the maternal health infrastructure was less robust. To do this, we estimated the effects of contraceptive use within categories of the hospital population density variable, the latter being a proxy indicator for local health infrastructure.