Study Design and Setting
Given the focus on improving maternal and child health, the Myanmar Demographic and Health Survey (MDHS) 2016 was the first nationally representative cross-sectional survey conducted in Myanmar. Data were collected from 12,885 women from the sampled households based on stratified two-stage cluster sampling design from December 2015 to July 2016. Using the 2014 Myanmar census sampling units, 442 clusters (123 urban, 319 rural) were selected in the first stage from 4,000 clusters based on the probability proportional to the size. In the second stage, 30 households from each selected cluster were selected in the first stage by using systematic random sampling. The overall response rate was approximately 98%. The survey was funded by the United States Agency for International Development and implemented by the Ministry of Health and Sports, Myanmar, in coordination with the Millennium Development Goals. Technical support was provided by ICF international. Detail of the survey sampling procedure has been published in the MDHS report [17].
Characteristics of Participants
A total of 3,249 under-five children was included in the final analysis based on their retrospective birth histories after limiting to singleton births living with their mothers at the time of the survey and excluding children with missing information on SFU (Figure 1)[17,18]. The inclusion criteria were: i) children born within five years before the date of survey (only last child and singleton births were considered in case of multiple children in five years); ii) most recent children with information of survival status (alive/death at the time of the survey); iii) children with the date of death if applicable; iv) children with complete information of household cooking fuels use[17].
Measures of Child Mortality Outcomes
We considered neonatal mortality (deaths occurred during the first 28 days of life), infant mortality (deaths occurred during the first one year (0-11 months) of life), and under-five mortality (deaths occurred during the first five years (0-59 months) of life) as outcome variables [17,19,20].
Measures of HAP Exposure
The analysis was carried out for two exposure variables: Solid Fuel Use (clean fuel vs. solid fuel) and level of exposure to SFU induced HAP (non-exposure, moderate exposure, and high exposure). The MDHS collected information on the types of cooking fuels by asking women- what type of fuel does your household mainly use for cooking? Responses were coded as clean fuel =0 (if responses were electricity, liquid petroleum gas, and natural gas) and solid fuel =1 (if responses were coal, lignite, charcoal, wood, straw/shrubs, grass, and agricultural crop). Children's levels of exposure to HAP were generated from the women's responses to the place of cooking and the type of cooking fuel use. The responses were categorized as non-exposure =0 (if women reported not using solid fuel), moderate exposure =1 (if women reported using solid fuel, but in a separate building or outdoors), and high exposure =2 (if women reported using solid fuel inside the house).
Confounder Adjustment
Different sociodemographic factors contributing to neonatal, infant, and under-five child mortality were included as confounders (Figure 2). These were age at child deaths, child sex, parental education, interval of last two succeeding births, breastfeeding status, household wealth quintiles, urbanity, geographic regions, and seasons (Figure 2). The birth interval variable was generated based on women's response to the birth date of the last two children and categorized by following the World Health Organization guidelines[17]. The wealth quintile was reconstructed from the women's household durable and non-durable assets (e.g., televisions, bicycles, sources of drinking water, sanitation facilities, and construction materials of houses) using principal components analysis, excluding the types of cooking fuels as this was the main exposure of interest [17,21].
Note: HAP is exposure, and child mortality is the outcome. The minimal and sufficient adjustment set contains child age, child sex, breastfeeding status, maternal education, household wealth quintiles, urbanicity, geographic region, preceding birth interval, and season. This figure was constructed through DAG (http://www.dagitty.net/dags.htm).
Statistical Analysis
Descriptive statistics were reported as frequency and percentage to characterize the demographic profile of the study sample. Differences in neonatal, infant, and under-five child mortality across sociodemographic factors were presented using the chi-square test. The associations between exposure to HAP and child mortality outcomes were investigated using both univariable and multilevel Poisson regression models. As an additional analysis, effect modification by sex of the child was also tested for all models. The univariate models included only the exposure variable and the outcome variable. These associations were then progressively adjusted for potential confounders in the multivariable models, including child age, child sex, breasting status, maternal education, household wealth quintiles, urbanicity, geographic region, preceding birth interval, and season. However, birth weight and wasting were not adjusted in the models as they are likely to be on the causal pathway between exposure to HAP and mortality [22–24].
Furthermore, information on exact birth weight was unavailable for most of the children[17]. Multilevel Poisson models with robust error variance to minimize the overestimation of binary outcome were developed for complex survey design effects, adjusting clustering effects, individual and household characteristics of the children[18,21]. Results were reported as adjusted relative risks (aRRs) with 95% confidence intervals (CIs). All statistical tests were two-sided, and a p-value < 0·05 was considered statistically significant.