Data source and setting
The data analyzed in this study are from a cross-sectional study conducted by Amref Malawi in October 2023 in Mangochi and Chikwawa districts in Malawi. The 2015-16 demographic health survey in Malawi (2017) showed that Mangochi District had a contraceptive prevalence rate of approximately 31%, significantly below the national average of 58%. In contrast, Chikwawa District had a contraceptive prevalence rate of 58.7%, suggesting a relatively higher contraceptive uptake [29]. The primary study collected data on women’s perspectives on their health status and behavior, including water, sanitation, and hygiene, and that of their families. Additionally, the study explored changes and effects of development interventions in the lives of community members over the years that Amref Malawi implemented in the two districts. The data were collected for 10 days in October 2023 using a structured questionnaire (Supplementary File 1) programmed on the Kobo Toolbox electronic data collection tool administered by trained research assistants.
The sample size in the primary study was 418 participants. The study was conducted in Mangochi and Chikwawa districts in Southern Malawi, two sites that have actively implemented Amref Malawi’s projects for over 10 years. The study sites provide valuable data on the effect of Amref’s work around sexual, reproductive, maternal, newborn, adolescent, and child health, including water, sanitation, and hygiene interventions.
Mangochi is a fast-growing rural district with a population of 1,148,611 (approximately 603,111 females), while Chikwawa is a predominantly rural district with an estimated population of 564,684, of whom 287,794 are women [30]. The primary study used a multi-stage sampling approach to select women of reproductive age that ranged from 15 to 49 years. The two districts had a total of 11 zones: 5 from Mangochi and 6 from Chikwawa. All the health zones were purposively selected for the study. A health zone usually has several health facilities based on the population density and the geographical boundaries. The number of villages served by a health zone fluctuates. However, each village is approximated to have around 100 to 500 households. The first phase of the sampling was a random sample of two health facilities within each health zone (overall, 22 health facilities). This was followed by a random sampling of three villages in the sampled health facilities, and then the selection of households within each selected village. In Mangochi, nine households from each village were randomly selected to participate in the study. Conversely, in Chikwawa, four households from each village were randomly selected. The selection was done using the list of household records that were kept by community health workers.
We used Microsoft Excel to generate random numbers for selecting health facilities, followed by villages, and then households. This sampling strategy resulted in a total of 418 women: 274 from Mangochi and 144 from Chikwawa.
Ethical considerations
The primary study received ethical approval from the Malawi National Social Sciences and Humanities Committee (reference number P.05/20/480) and was carried out in line with the principles of the Declaration of Helsinki. For this analysis, ethical approval was waived by the ethics committee as the analysis of the already collected data posed no potential risks to the participants, and there was no direct contact with them. We ensured the privacy and confidentiality of participants by anonymizing the data. Data were analyzed and presented in aggregate form, hence posed no risk to the participants. In the primary study, participants provided written informed consent before data collection.
Study design and population
We used a quasi-experimental study design as there was no randomization of the intervention. A randomized trial is considered the gold standard for evaluating cause-effect relationships. However, in this study, randomization was impractical because the source of modern contraceptive health education was determined by how the program was designed and implemented. This phenomenon is referred to as program selection, where individuals are exposed to interventions based on programmatic decisions rather than random assignment. Given these considerations, a quasi-experimental study design was the most suitable approach. The lack of randomization introduces selection bias and confounding. Therefore, inverse probability of treatment weighting was used to remove systematic differences between the two groups and achieve comparability on observed covariates, allowing the emulation of a randomized trial. We included data for women of reproductive age, 15 to 49 years.
Study variables
Intervention
The intervention was exposure to health education from either a community health worker, healthcare providers such as nurses, or non-healthcare providers such as traditional leaders (chiefs), religious leaders, youth leaders, peers, teachers, or through the media (radio, television, magazines, and social media handles like Facebook or X). Youth leaders delivered health education every week through Youth Centers. Traditional and religious leaders were used on a need basis. In Malawi, community health workers undergo 12 weeks of standardized training that covers both health promotion and contraceptive services. They receive the same level of training on contraception as professional healthcare providers. As part of their role, they provide various contraceptive methods, including Depo-Provera and oral contraceptive pills. Non-health workers underwent a one-day orientation on modern contraceptives, lasting approximately 4–6 h. Both health workers and non-health workers delivered basic contraceptive education, including information on the types of contraceptives available to women of reproductive age, where to access them, and the benefits of contraception.
To determine whether women received health education and identify the source of that information, they were asked to indicate whether the education was provided by a health worker or a non-health worker. The data collection process captured a comprehensive list of sources from which individuals could have received information about contraceptives. Based on this information, individuals were categorized as having received contraceptive information from either a health worker or a non-health worker. Women who reported receiving information from a community health worker or healthcare provider were considered as having received health education from a health worker, while those who reported the other sources were considered as having received it from a non-health worker. These responses were respectively coded as 1 and 0.
The intervention group consisted of women who received health education from health workers, while the comparison group comprised women who received health education from non-health workers.
Outcome variable
The primary outcome was modern contraceptive use measured as a dichotomous variable (yes vs. no). Women were asked to report if they started using modern contraceptives after receiving contraceptive information. Affirmative responses were coded as 1; otherwise, 0.
Covariates and their selection
We selected covariates based on the conditional independence assumption—variables that were associated with the outcome and those that differed systematically between the intervention and comparison groups. The covariates included age, religion, employment, educational level, parity, the partner’s educational level, and marital status. We categorized age as 15–24 and 25–49 years to denote adolescent girls and young women, and mature women, respectively. Level of education was categorized as none or no formal education, primary, or post-primary (tertiary, vocational, and secondary school). Marital status was measured as married or single (not married). Parity was measured as the total number of live births a woman had at the time of the data collection and was categorized as < 5 vs. ≥ 5. Antenatal care (ANC) attendance was measured as the number of ANC visits the woman had had at the most recent pregnancy and then categorized as < 4 visits vs. ≥ 4. Religion was categorized as Christian and others (Muslim and traditional religion). Employment status was measured as unemployed and employed.
Statistical analysis
Baseline characteristics of women were reported using frequencies and percentages for categorical variables, while continuous variables were summarized using mean and standard deviation. These characteristics were summarized by exposure to health education status to examine potential imbalance or confounding bias. Associations between covariates and the treatment variable were assessed using Pearson’s chi-square tests. Inverse probability treatment weighting was utilized to remove selection bias and confounding and to create a comparable group based on observed covariates. First, the propensity scores were computed as the probability of being exposed to the intervention given baseline characteristics, by regressing the intervention as a function of the covariates. Second, inverse probability treatment weights were computed as the inverse of the propensity score (1/propensity score) for the intervention group and (1/ [1-propensity score]) for the comparison group, creating a pseudo-population in which the covariates were equally distributed across the two groups. As such, individuals in the intervention group with a lower probability of exposure and those in the comparison group with a higher probability of exposure received larger weights, and therefore their relative influence on the comparison was increased. We checked covariate balance using an absolute standardized mean difference (SMD) and considered an SMD of less than 10% or 0.1 as indicative of balanced covariate. We then performed a binary logistic regression analysis to estimate the effect of the intervention on the outcome, adjusting for the inverse probability treatment weights. We reported the odds ratio (OR) and the 95% confidence interval (CI).
For a robustness check of the causal estimate, we computed the unadjusted association between the intervention and the outcome using a binary logistic regression analysis. We performed sensitivity analyses to assess the robustness of the causal effect to unmeasured confounders or hidden bias—bias that was not identified or removed by the analytic approach—using Wilcoxon’s signed rank test.
A very large change in the lower/upper odds bounds was needed before a change in statistical significance was interpreted as suggestive of robust findings. The statistical analyses were carried out using the R Programming Language and Statistical Software (R Version 4.4.1).
link
