Data and sample
This study is based on data from the 2011 and 2016 Ethiopian Demographic and Health Survey (EDHS). The EDHS are five-year periodic national representative household surveys that collects retrospective information on a wide variety of health, socio-economic and demographic factors for the population across all region with the aim of improving maternal and child health in Ethiopia. The 2011 and 2016 EDHS used stratified two-stage cluster sampling design to select respondents for the study. Elaborate details of the survey protocols and their deigns have been reported elsewhere [22, 23]. Data were obtained through personal interviews with women in the child-bearing age 15–49 years. The EDHS consist of three components: the household questionnaire, the woman's questionnaire and the man's questionnaire. From the woman’s questionnaire, data for child mortality along with related variables were extracted. A total sample 11 654 children from 2011 survey and 10 641 from 2016 were examined. Information on children were taken from birth history supplied by mothers.
Treatment and Outcome
The health outcomes in our study is under-five child mortality which refers to death of children before reaching the age of five. The information is captured through the full birth’s history recalled by the interviewed women and recorded in the surveys. We define treatment as utilisation of antenatal care. Therefore, we have two groups. One group is composed of women who attended at least one ANC, considered as the treatment group. The other group includes women who never attended ANC throughout their pregnancy, known as the control group. The rate of under-five mortality from 2011 to 2016 for the control group will vary due to several possible unknown factors. The variation of this rate at the treatment group will be due to the same factors plus the variation in the utilisation of ANC.
Statistical Analysis
We adapted a two-stage research design to improve comparison between the treated and control groups. First, using propensity score matching, we matched mothers on a number of individual and household characteristics that would affect their likelihood of utilising ANC. Second, we assess the impact of ANC on under-five mortality among those coming from similar households. This minimizes the effects from uncontrolled factors that affect both utilisation of ANC and child health.
Our analyses begin by matching individuals with household characteristics that provide equal probability of utilising ANCs across groups. To this end, we used propensity score matching. Propensity score matching is a statistical technique that seeks to address the primary drawback of causal inferences from observational research designs where no standardized methods have been used to establish control groups [24]. This technique involves forming matched sets of control and treatment individuals who’s propensity score are similar [25]. If a matched sample has been established, the treatment effect can be assessed by comparing the outcomes directly between treated and control subjects in the matched sample [26].
The demographic and socio-economic covariates entered into the propensity score includes maternal age, age at first birth, child sex, birth order, birth size, birth interval, family size, residence status, region, wealth index and religion. Women that utilised ANC were matched to women that did not utilised ANC based on a logit regression performed using calipers of width equal to 0.2 and a nearest neighbourhood matching method with ratio of 1:1. We used a Chi-Square test to access the balance for all covariates before and after matching, with a 5% level of significance or more considered indicative of imbalance.
The Difference-in-Differences (DID) method was used to analyse the effect of ANC on under-five mortality. The DID is a quasi-experimental approach that compares outcome changes over time between a group involved in intervention (treatment group) and a group that is not (control group) [27]. While the DID method typically uses panel data to estimate the causal impact of policies or programmes, repeated cross-sectional data from the same areas has also been used in the literature [28–30].
We apply the DID method using the linear probability model:
To enables us to estimate the differences in under-five mortality for treatment and control groups. Where, Yit refers to the binary indicator whether child i born in year t died or not prior to reaching the age of five (under-five mortality). The variable Treatmenti is a dummy with 1 indicating mother had ANC and 0 otherwise. The variable Timet is also a dummy variable coded 0 for 2011 and 1 for 2016. The DID estimate β3 of effect of ANC, is an interaction between Treatment and Time. The vector Xi is a vector of variables controlled by propensity score matching and Zi is a vector of additional covariates to adjust for the remaining imbalance from our matching procedure. To account for the complexity of the survey design, the primary sampling unit, strata and person weight were incorporated in the regression models to adjust for the standard error. All statistical analysis where carried out using SAS version 9.4.