A Hierarchical Bayesian approach to small area estimation of health insurance coverage in Ethiopian administrative zones for better policies and programs

Health insurance is vital to the realization of universal healthcare and the prevention of disease. Community-based health insurance (CBHI) is the most effective means to provide universal health coverage (UHC) with enough financial protection against healthcare expenses, promote equitable access to high-quality healthcare, raise financial security, and improve social cohesion and solidarity [1‐3]. Out-of-pocket medical expenses, however, lead to lower healthcare service utilization and higher catastrophic medical expenses. Certain countries have implemented CBHI programs to reduce out-of-pocket expenditures and enhance access to healthcare services, particularly for people living in rural areas or working in the unorganized sector [4, 5].

Around 100 million people are forced into extreme poverty yearly due to out-of-pocket health spending, and over 930 million individuals worldwide spend at least 10% of their household income on healthcare [6]. CBHI is becoming more popular in low- and middle-income countries to enhance healthcare utilization and shield households from financial catastrophes brought on by out-of-pocket medical charges [7]. A significant issue in Africa has been the catastrophic nature of healthcare financing arrangements for the poor, particularly the rural population [8]. Out-of-pocket charges can significantly hinder access to quality medical treatment in Sub-Saharan African nations[9]. Most sub-Saharan African nations, including Ethiopia, have adopted the CBHI program to increase the population's fair access to health services. However, enrollment rates and renewals for CBHI in sub-Saharan Africa are persistently low [10].

In 2019, it was estimated that 28% of the Ethiopian population nationwide had health insurance, primarily community-based health insurance [11, 12]. In other words, nearly 82 million Ethiopians, or about 72% of the country's total population, lacked health insurance enrollment. According to [13], a lack of health insurance is significantly associated with worse health outcomes and restricted access to medical care. People without insurance may experience health problems due to their lack of coverage. This could also affect the quality and organization of local healthcare systems [14].

The Sustainable Development Goal for Health (SDG3) aimed to ensure healthy lives and promote well-being for people of all ages [15]. SDG3 includes a sub-target (3.8) on achieving UHC, including financial risk protection and access to quality, essential health services. One of the SDGs for 2030 is UHC, which means everyone can access a full range of high-quality health services when and where they need them without financial hardship [16]. This goal still needs to be accomplished in six years. Although the UHC service coverage index (SDG indicator 3.8) has increased from 45 in 2000 to 67 in 2019, 30% of the world's population still lacks access to essential health services [17]. Ethiopia's health insurance coverage in 2019 was 28% [12], far below the SDG target of 80% by 2030. The inability to reach other SDG health-related targets results from this low coverage rate compared to other nations in the region. Additionally, 31% of all medical spending in 2016–17 was attributed to catastrophic out-of-pocket costs, which have remained high over time [18].

To achieve UHC in Ethiopia, it is essential to understand the impact of health insurance on healthcare utilization and financial protection. Studies have highlighted the potential factors related to maternal health-seeking behavior and the association between health insurance membership and antenatal care utilization in Ethiopia [19]. Additionally, the impact of CBHI on universal health coverage in Ethiopia has been evaluated, indicating positive impacts on health service issues by reducing catastrophic costs. However, the low universal health coverage for a limited population suggests a broader national scheme covering the entire population [20].

Aside from that, coverage varies according to administrative regions and other sociodemographic factors [12]. According to a recent report by [17], despite improvements in healthcare coverage nationally, intra-country disparities are obscured by aggregate data, making it challenging for UHC to address inequalities.

The studies above compiled national and regional health insurance coverage estimates through household surveys. These surveys are used to obtain accurate statistical information about a specific target population in broad regions, such as the nation, the state, the province, and so on [21, 22]. However, estimates obtained from local surveys are often unreliable due to inadequate sample sizes. In the case of Ethiopia, estimates for lower administrative levels, such as zones, are unavailable due to small sample sizes. To address this issue and produce more precise estimates, the small area estimation (SAE) model combines survey and census data.

Small areas refer to subpopulations or unplanned domains in large-scale sample surveys. In these cases, model-based methods, also known as the SAE approach, have consistently attracted attention for their ability to provide appropriate estimates for such a small area or domain. The 2030 Agenda for Sustainable Development Goals (SDGs) emphasizes the importance of decentralized statistics for micro-level planning, policy formation, and targeted social advancement. SAE is indispensable for reconciling the demand for reliable, representative, and disaggregated official statistics [21].

In Ethiopia, the demographic and health survey (DHS) program only includes subnational administrative levels like regions. In other words, the zones, Ethiopia's third administrative layer, were not the focus of the DHS program's design when providing accurate health-related information. For public health policymakers, research that only uses survey data at the lowest administrative levels, like zones, may not be reliable and practical since the estimates have large standard errors and coefficients of variation. As a result, additional sample data is needed to generate precise projections at the zone level [21]. We cannot increase the sample size due to financial and operational constraints.

There has been a growing demand for small area statistics in both the public and private sectors. Currently, SAE is experiencing high demand, with occasional nationwide spikes. In order to allocate resources and focus initiatives at lower-level administrations, SAE is a crucial technique for estimating health, education, and environmental indicators [23, 24]. The DHS program, the World Bank group, the International Monetary Fund (IMF), and the World Health Organization (WHO) were all very concerned about how to disaggregate DHS data and provide it to third-level government administrations [25‐27]. The DHS host nations often request the DHS Program for subregional indicator estimates for development planning and policymaking [28]. In order to facilitate the mapping of health and economic inequality, the United Nations Population Fund (UNFPA) works to strengthen countries' capacities to produce and distribute high-quality, disaggregate data on population and development indicators, as well as demographic discrepancies, and to support humanitarian programming [29].

Estimating health-related parameters at unplanned geographic levels is crucial for successful health promotion and initiatives, improved healthcare delivery, and population-specific health planning and management. The lack of a readily available, adequate dataset for characteristics of interest in small areas has a detrimental effect on the ability of local and national organizations to manage significant health issues and associated risks in the community [12, 15, 18]. Developing a method for simulating reliable small-area statistics could solve this problem.

In the Ethiopian DHS, some zones have very few or no samples available. However, estimating the characteristics of interest, such as health insurance coverage, at the zone level for effective healthcare planning and resource allocation is crucial [30, 31]. To address this need, the study aims to estimate the level of health insurance coverage in Ethiopia at the zone level. The SAE approach's strength lies in using additional data sources, including census and administrative data [22, 32]. Therefore, this study used the HB approach to SAE, which combines data from the 2007 population census and the 2019 Ethiopian DHS.

Accurately estimating health insurance coverage at the zone level is vital in Ethiopia for several reasons. Firstly, it allows policymakers and healthcare providers to allocate resources more efficiently by providing precise information on the areas in need of healthcare services. Secondly, understanding the distribution of health insurance coverage ensures that healthcare services are accessible to all individuals, regardless of their location, thereby promoting equitable healthcare access. Thirdly, precise estimates of health insurance coverage at the zone level provide valuable insights for evaluating the effectiveness of existing policies and planning future interventions. Lastly, by estimating health insurance coverage at the zone level, healthcare providers can identify zones with high uninsured populations, which helps to design strategies to increase coverage, leading to a more efficient and equitable healthcare system. This knowledge can also help reveal geographic disparities and enable healthcare providers to address them more effectively [33, 34].

The Fay-Herriot model and its extensions have been used in various studies in the SAE framework. These include exploring spatiotemporal models using undernourished children in studies such as [34‐36] and the Bayesian spatiotemporal approach to quantify SARS-CoV-2 [37]. Spatial SAE is also employed by [33, 38, 39]. In order to explore binary outcomes of response variables for SAE, a logit-normal model [40], logit-normal spatial model [41], HB logit-normal model [8, 42], HB logit-normal spatial model [43], zero-inflated Binomial [44], and zero-inflated negative Binomial model [45] are used under the Fay-Herriot model.

Although the SAE approach is popular in literature, it has yet to be widely practiced in Ethiopia. Therefore, this study aims to implement the HB SAE approach using the CBHI binary outcomes from the 2019 EMDHS and auxiliary variables from the 2007 population census datasets [46]. The main objective of this research is to estimate the community-based health insurance (CBHI) coverage at the zonal level in Ethiopia using the HB SAE approach. The study is expected to offer important insights into the spatial distribution of health insurance coverage, which is vital for making informed policy decisions and developing effective healthcare interventions at the zone level.

The paper is structured as follows: Section two explains the data sources, sampling design, and methods applied in the study, namely the Fay Herriot model under the different linking models and the HB SAE model. Section three presents the results and discussions, while Section four concludes.

This research focused on estimating the zone-level CBHI in Ethiopia by utilizing the 2019 EMDHS and the 2007 population census data. The CBHI enrollment status is represented using binary outcomes, where zero indicates non-insured individuals and one represents those enrolled in CBHI [65]. For binary outcomes, we employed the Logit Normal model with the HB SAE method [8, 22, 41, 50]. This method is advantageous because it simplifies model specifications and is relatively easy to compute using the MCMC technique [44]. This study utilizes the GVF to effectively smooth unreliable and noisy estimates of sampling error variances for SAE obtained through direct surveys. The results indicate that GVF helps to smooth out unreliable and noisy estimated variances, which is consistent with previous SAE literature [59, 60, 65‐67]. Figure 2 shows the direct survey estimated variance (red color) and the GVF predicted variance (blue color). The GVF predicted variance is smoother compared to the variances obtained through direct survey estimates.In HB SAE applications, model diagnostics for small area estimates are typically employed. These techniques are used to verify the assumptions made by the model [59, 68]. The Shapiro–Wilk normality test and Q-Q plot are the most commonly used methods to establish normality [53, 69]. The model assumptions of the random area effect \(\nu\) are assumed to have identically and independently distributed with mean zero and variance (\(\sigma_{\nu }^{2}\) = 0.3243). The results of the Shapiro–Wilk test indicate that the residuals at the zonal level (W = 0.950 and p-value = 0.085) support the normality assumption for the data on health insurance coverage. This suggests that the normality assumption for the health insurance coverage data is adequately met. Similar studies in Ethiopia about children under the age of five have agreed with the normality assumptions [31, 34].

The bias diagnostics plot is the second diagnostic measure used to evaluate the reliability of HB estimates [68]. It helps identify potential sources of bias and assesses the model's validity. When estimating health insurance coverage through HB, it should be comparable to direct survey estimates. Figure 3 shows the direct survey estimates on the y-axis and the HB estimates on the x-axis. The graph depicts a strong relationship between both estimates, which is consistent with various literature sources [33, 53]. The direct survey-based and HB estimates agree, indicating that they are consistent and aligned. To put it differently, the estimates obtained from the HB method and the direct survey method are consistent with each other.

Additionally, three widely used diagnostic measurements, such as percent CV, root MSE, and 95% credible intervals, are also used to assess the accuracy of the HB estimates [69]. Even though there are no precise cutoff values for CV to determine precision, the lower the CV, the more precise and consistent the estimates [70]. According to the condensed tables and figures (Fig. 6 and Table 3), the HB estimates have a lower CV in each zone. The HB estimates are relatively more accurate than the corresponding direct estimates in all zones, according to the CV, root MSE, and 95% CI, which are consistent with the different study results [57, 71, 72].

The results of this study found that auxiliary variables such as marriage and the number of children of more than five are statistically significant predictors of CBHI prevalence. This finding explicitly revealed that respondents had more than five children in the household, which is a significant effect of the CBHI scheme. Other studies are consistent with our results [3, 73, 74]. The number of children over five significantly impacts the CBHI prevalence, which is consistent with the studies [2, 3]. Similarly, the studies conducted in South Gondar zones of Ethiopia show that those having more than five children in the household were two times more likely to use the CBHI scheme when compared with those who have fewer than five children [2, 3].

This could be because families with more children require more medical care, such as immunizations, which could explain why they seek enrollment in CBHI programs. However, recent research contradicts earlier studies in Kenya, which suggested that the number of children did impact CBHI enrollment [75]. Under UHC, healthcare should be accessible to all without causing financial hardship [17]. Achieving UHC is essential for long-term development and reducing poverty. It is also critical in any effort to combat social inequality and improve access to healthcare [5, 16].

The study also demonstrates that auxiliary variables, like marital status, can significantly influence the possibility of CBHI adoption in Ethiopia. This suggests that evaluating and enhancing the uptake of CBHI schemes requires an understanding of the demographic and social factors of the population. This finding aligns with a study in the North Shoa Zone of Ethiopia by [76], which also found a significant association between marital status and CBHI utilization.

Using the estimation from the disaggregated level HB estimates, we were able to identify disparities in CBHI coverage between Ethiopian administrative zones. It was found that the Amhara and Tigray regional zones have better CBHI coverage than other regions of the nation, as shown in Fig. 7. This is because the Ethiopian government approved and initiated the CBHI practice in the Tigray and Amhara regions before rolling it out nationwide to improve healthcare quality [49]. As a result, the prevalence of CBHI in Ethiopia differs from region to region, with higher enrollment rates in the Amhara and Tigray regions than in others [77]. This finding agrees with a study by [2], which also reported higher enrollment rates in these regions. Similarly, another study by [48] found that the overall pooled prevalence of CBHI enrollment in Ethiopia was 45.5%, with better coverage in the Amhara and Tigray regions. Our research also confirms that there are variations in health insurance coverage across different administrative zones [77, 78]. It is evident that health insurance coverage varies geographically.

Our study is unique in that it provides synthetic estimates for small areas with zero sample sizes. These estimates are generated by either combining data from different sources or using statistical models to infer values for smaller areas based on larger, more reliable datasets. Synthetic estimates help make informed decisions when direct measurements are unavailable [22, 50]. In Table 4, the HB model-based SAE produced CBHI estimates for non-sampled zones. We used the HB model-based SAE to estimate CBHI coverage in non-sampled zones, which borrowed auxiliary variables from census datasets. This approach is consistent with most studies in SAE contexts [39, 78, 79].

The SAE approach is a statistical method used to estimate the proportion of socioeconomic indicators at the local or domain level. This method has been extensively studied and adopted as a target by national statistics offices in developed countries. However, there is a need for more research on how to apply SAE in developing countries such as Ethiopia. For example, in order to track progress towards achieving the SDGs, it is essential to have disaggregated estimates at lower administrative levels or specific domain levels. This initiative aligns with the global framework of indicators for monitoring the SDGs, which suggests that information should be disaggregated not only geographically (in subregions such as provinces, municipalities, or districts) but also by income group, gender, age, race, ethnicity, immigration status, and disability status [80].

This study found that by combining data from the 2019 EMDHS with additional information from the 2007 population census, more accurate estimates of health insurance coverage at the zone level were obtained using the SAE approach. The research also suggests that relying solely on survey data for estimating the CBHI coverage in Ethiopia's administrative zones is less precise than when auxiliary variables are also taken into account.

Ethiopia is strongly committed to achieving UHC, which means providing high-quality healthcare services that are accessible and equitable to all. Our study has highlighted the differences across Ethiopian zones regarding their coverage of CBHI schemes. We have also identified some areas with high coverage and some with low coverage. This information can be very useful in improving local planning and policy decision-making, which can ultimately lead to an increase in the coverage of CBHI schemes across Ethiopia.

Overall, this study stands out due to its innovative approach of integrating data from two sources, namely the 2007 population census and the 2019 EMDHS, to improve the accuracy of health insurance coverage estimation at the zone level. The study employed the HB modeling framework in SAE to estimate health insurance coverage at the lower administrative level (zones), which can help policymakers gain a more nuanced understanding of coverage disparities and target resources and interventions more effectively. Furthermore, this paper's novelty lies in estimating the non-sampled zones, which will receive equal attention from policymakers as the sampled zones. This study contributes to evidence-based decision-making in healthcare planning and resource allocation, making it valuable for researchers and policymakers.