This study followed the standard guidelines for reporting observational studies using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE).
This study utilized secondary data from the most recent and available Demographic and Health Survey (DHS) conducted in 26 SSA countries between 2013 and 2019 (Supplementary Table 1). The DHS is a nationally representative household survey with similar data collection instruments and study designs conducted in LMICs with the primary goal of generating estimates for indicators that are comparable across the sub-region. The DHS provides data for a wide range of monitoring and impact evaluation indicators in the areas of population, health, and nutrition. Specifically, the DHS collects data on family planning (knowledge and use of contraceptives), maternal health (antenatal, delivery, and postnatal care), household wealth, parity, education, place of residence, and demographics, amongst other variables with sample sizes (usually between 5,000 and 30,000 households) and typically are conducted about every 5 years, to allow comparisons over time. The survey employs a multi-stage stratified cluster sampling design where the index country is stratified into distinct geographical regions or provinces during the first phase of the design. The first phase of sampling involves the random sampling of clusters or enumeration areas (EA) using probability proportional to the size of the EA and the subsequent sampling of a fixed number of households within each of the sampled enumeration areas using a systematic random sampling approach. A complete household listing was carried out to update the sampling frame before the random sampling of households. Trained field data collectors were assigned to these sampled enumeration areas for the household survey. Details on the study design and procedures for data collection have been published elsewhere.18
The DHS data is publicly available upon reasonable written request at the DHS website (https://dhsprogram.com/data/available-datasets.cfm).
This study did not need ethical approval since we only analyze secondary data from the DHS program where all study participants have been de-identified.
The primary outcome measure in this study was contraceptive use. Contraceptive use as defined by DHS was among women of reproductive age who currently use any standard method of contraceptive (traditional or modern). Contraceptive use was classified as a binary variable that takes the value of 1 if the woman is currently using a modern contraception method and a value of 0 if otherwise. The modern methods include women who use female sterilization (tubal ligation, laparotomy, voluntary surgical contraception ), male sterilization (vasectomy, voluntary surgical contraception ), the contraceptive pill (oral contraceptives), intrauterine contraceptive device (IUD), injectables (Depo-Provera), implant (Norplant), female condom, the male condom (prophylactic, rubber), diaphragm, contraceptive foam and contraceptive jelly, lactational amenorrhea method (LAM), standard days method (SDM), country-specific modern methods. Respondents mentioned other modern contraceptive methods (including cervical cap, contraceptive sponge, and others), but do not include abortions and menstrual regulation.19
Exposure to FPM was defined as individual women of reproductive age who heard or saw FPM on the radio, on television, in a newspaper or magazine, or on a mobile phone in the past few months.19
Variables considered as possible confounders were selected based on an extensive literature review of factors that could potentially influence access to FPM and contraceptive use among women of reproductive age. The following variables were accounted for in all the multivariable models: the age of the household head (categorized as ≤29, 30-39, 40-49, 50-59, and 60+), sex of the household head (male or female), household wealth Index (poorest, poorer, middle, richer, richest), place of residence (rural or urban), religion (Islam, Christian or Others), respondent age (15-19, 20-29, 30-39, 40-49), marital status (widowed, never married, married or divorced), educational level (no formal education, primary, secondary, higher), currently working (no, yes), children ever born (no child, 1 child, 2 children, 3+ children).20,21
We explored the trend of FPM and CU between 2013 and 2019 using tools from time series and estimated the weighted prevalence of FPM and CU over the period. Factors contributing to CU and FPM were assessed using the Poisson regression model with a cluster-robust standard error that generates prevalence ratios and their respective confidence intervals. Sensitivity analysis of the point estimates and corresponding confidence interval (CI) was conducted using the multivariable binary logistic regression model that reports odds ratio and CI. The Poisson model was preferred to the logistic regression model as the odds ratio may overestimate the prevalence ratio, the measure of choice in cross-sectional studies.22 Augmented inverse-probability weighting (AIPW) was used to estimate the average treatment effect of FPM from cross-sectional data. The AIPW estimator is classified among the estimators with the doubly-robust property as it combines aspects of regression adjustment and inverse-probability-weighted methods to reduce bias associated with the impact estimate. The model accounted for sampling weight and used cluster-robust standard errors to address the methodological challenges (stratification, clustering, weighting) associated with complex survey design. Since different impact estimation procedures may lead to slightly different impact estimates especially when the data originates from crossectional studies instead of the more rigorous experimental design, sensitivity analysis of the impact estimate was conducted using endogenous treatment effect models, inverse probability weighting, propensity scores, and nearest-neighbor matching techniques. Estimating the impact of an intervention, program or policy becomes difficult due to endogeneity. For instance, genetic predisposition, personal values, conservative lifestyle, religious beliefs, and other unmeasured confounders may simultaneously affect exposure to family planning messages and utilization of contraception.13 The standard regression models (e.g., Poisson, Negative Binomial, binary logistic, probit, and ordinary least square assume that these unmeasured covariates do not correlate with both the outcome measure (contraceptive use) and exposure to FPM. This assumption is largely violated in the context of observational data where both the outcome and exposure are usually measured at the same time and may correlate with unobserved confounders. We anticipated these problems, and as part of the sensitivity analyses that were conducted, we used endogenous treatment regression models to address endogeneity. Having radio or television was used as the instrumental variable since it met the exclusion restriction criteria recommended for instrumental variable regression analysis (that is, having a radio or television sets influence the ability to listen to FPM directly, it does not influence the use of contraceptives directly, but only through the family planning message and we assume that it is not influenced by other factors). All statistical analyses were conducted using Stata version 17 (StataCorp, College Station, Texas, USA) and a p-value of less than 0.05 was considered statistically significant.