Leveraging the panel data structure, we conduct a cluster analysis on the contraceptive intention trajectories that women follow in the first year postpartum. Cluster analyses allow researchers to construct groups based on similar characteristics or patterns within the data [17, 23]. These clusters may be structured hierarchically or non-hierarchically. We use a non-hierarchical method known as k-means clustering [24, 36] because it is well suited for large datasets and it allows individuals to move from one cluster to another, as we depict in Fig. 1. This differs from hierarchical approaches, where an individual cannot move to another cluster once they are assigned.

The ‘elbow method’ allows us to visually determine the optimal number of clusters (k) for this dataset based on the k-means clustering approach, by finding the inflection point at which the number of clusters introduced produces diminishing returns to the within-cluster sum of squared errors [24]. In the elbow plot in Fig. 2, k is increased incrementally and plotted against the total within sum of squares (TWSS). TWSS is a measure of the variability of the observations within each cluster. A cluster that has a small sum of squares is more compact than a cluster that has a large sum of squares. The more compact the cluster, the more similar data points within that cluster are to one another, on the dimension used for clustering.

Optimal number of clusters using the elbow method (A) and the gap statistic (B)

In Fig. 2, we see a gradual decrease in the within-cluster-sum of squared errors on the x-axis and a sharp inflection point when k = 3. This point represents the “elbow” at which we see diminishing returns in the TWSS value as we incrementally add more clusters. As with any cluster analysis, determining the number of clusters makes a difference in the interpretability.

The elbow method works when a strong inflection point emerges. However, we also see in Fig. 2 a dip in TWSS at the k = 8 point. Thus, we supplement the elbow method with the gap statistic method which allows us to compare a heterogeneous, clustered data structure to a hypothetical ‘null’ data structure [47]. The ‘gap’ is the degree to which the real data falls below the expectation of a ‘null’ or normally distributed data structure (i.e. data with no clustering whatsoever). The gap statistic would then identify the number of clusters that create the largest gap between the real and null distributions. We find two inflection points. At k = 8, the gap statistic is maximized, but we also see a similar inflection point at k = 3, reflecting what we found in the elbow method. For the purpose of this analysis, and of interpreting our results, we restrict the algorithm to three clusters. However, we include supplementary materials for a k = 8 clustering algorithm and discuss possible tradeoffs for the two approaches in the limitations section.

Based on the three-cluster algorithm, we examine which trajectories fall into which clusters. Figure 3, below, shows the probabilities of cluster membership for each state possible for all four waves of the survey. Distinct patterns emerge. Cluster 1 is primarily composed of individuals who do not intend to use a contraceptive method and actualize those intentions to not use during the first year postpartum. For brevity moving forward, we call members of this cluster “Actualized Non-Users.” Cluster 2 is characterized by individuals who intend to use a contraceptive method within the first year, but do not go on to actualize those intentions to use. We label this cluster the “Aspiring Users.” Cluster 3 represents respondents who both express an intent to use PPFP and actualize those intentions. We label this cluster “Actualized Users.”

Probabilities of cluster membership based on intent-to-use states across all four waves of the PMA Ethiopia cohort survey

While the dynamics of Actualized Non-Users and Actualized Users are informative, the cluster of Aspiring Users represents the addressable burden for PPFP in Ethiopia. That is, this is the group of women who have expressed intent to use contraceptives postpartum but who likely face barriers to access, which may be addressed with targeted intervention.

Using the results of the cluster analysis, we assign respondents to the clusters identified based on the probabilities outlined in Fig. 2. Importantly, although the results of the clustering method are probability-based (e.g. what is the probability that a woman who intends to use at 6-weeks and uses at 6-months will be included in Cluster k), we use a binary categorization of member or non-member, for the sake of simplifying our interpretation. That is, for a woman whose trajectory matches one of the clusters, she is considered a member, rather than being assigned a value between 0 and 1 to determine how close to a particular cluster she might be. Because of the necessity of categorizing women into mutually exclusive clusters, we allow for women who follow the same trajectory as identified in the clustering algorithm, but who answer in any follow up that they are unsure of their intentions, to be categorized into the cluster, nevertheless. This binary categorization of respondents into clusters necessarily excludes a substantial number of women who do not follow the trajectories identified in the cluster analysis, a limitation that could be more deeply explored with further research. With these simplified clusters, we can examine the individual-level characteristics associated with actualization, as well as with women who were unable to actualize their intent to use. Full descriptive statistics of all three clusters as well as the women who were not assigned a cluster are available in the Supplementary Materials, in table A2. For brevity, we present selected characteristics that differentiate the three clusters in Table 2, below.

Examining the individual-level characteristics of respondents in each cluster reveals several distinct patterns. Overall, Aspiring Users were evenly distributed across region and wealth, but 44% had no education (relative to 80% of Actualized Non-Users and 20% of Actualized Users). Contact with the healthcare system is an important factor distinguishing these clusters—while all women were very unlikely to receive PPFP counseling at both ANC and PNC visits, women who expressed intent to use (both Aspiring Users and Actualized Users) were slightly more likely to receive these multiple counseling touchpoints. Over half of women in all clusters did not receive PPFP counseling at any visit. Preference for home birth was highest among Actualized Non-Users (73%), low for Aspiring Users (28%) and lowest for Actualized Users (12%).

Using the same data and cluster categories, we now turn to the multinomial model results, to estimate the significance of individual factors on a woman’s intent-to-use trajectory. The outcome of the multinomial regression model is cluster membership. Figure 4 below maps coefficients on the x-axis and variables on the y-axis. The vertical dotted line represents the reference, which is the Actualized Non-Users cluster. Point estimates for Aspiring Users (circle) and Actualized Users (triangle) that overlap the dotted line are not significantly different from Actualized Non-Users. Point estimates for Aspiring Users and Actualized Users that overlap is not significantly different from one another. Because we are primarily interested in identifying the drivers of not actualizing a stated intent to use, we will focus our main discussion on the variables significantly different for Aspiring Users.

Multinomial regression model coefficients. Reference category is Cluster 1—Actualized non-users

Only two factors were associated with being an Aspiring User rather than actualizing either use or non-use are: (1) this pregnancy being unintended; (2) region. Two additional factors were associated with a significantly higher likelihood of being an Actualized User rather than an Aspiring User or Actualized Non-User: first birth, and wealth. All other factors are either not statistically significant or significant for both Aspiring Users and Actualized Users—suggesting that these factors are significant for developing an intent to use, but not necessarily sufficient for actualizing that intent.

This pregnancy being “mistimed” is associated with higher intentions—with significant higher probability of being an Actualized User or an Aspiring User than an Actualized Non-User, but no difference between intending—as an Aspiring User—and actualizing that intent—as an Actualized User. However, the pregnancy being “unintended” is associated with being an Aspiring User rather than either an Actualized User or an Actualized Non-User. Women who declared their pregnancy truly unintended (i.e., did not want a pregnancy versus wanted a pregnancy but not now) were more likely to intend to use a contraceptive method during the first year postpartum but subsequently not use a method.

Regional differences are striking. Women in Tigray, Oromiya, and the Afar and Amhara regions are more likely to be Aspiring Users, and in the case of Tigray and Oromiya, significantly less likely to be Actualized Users. Perhaps this is not altogether unsurprising, as each of these regions has faced challenges in health care delivery over the same period (e.g., conflict in Tigray).

While we do not see a significant wealth difference between Actualized Non-Users and Aspiring Users, we see a familiar wealth gradient for Actualized Users. This suggests that while wealth does not significantly impact a woman’s intent to use, it certainly seems to play a role in actualizing her intentions. Similarly, if this birth was a woman’s first, she was more likely to be an Actualized User than any other trajectory.

PPFP counseling is a topic that has garnered much attention in Ethiopia, especially with the expansion of the healthcare extension workers nationwide (cite). However, we find that while receiving PPFP counseling in both ANC and PNC visits was significant for women’s expressed intentions, it did not distinguish between women who were Aspiring Users and those who were Actualized Users.