Socioeconomic and Regional Determinants of Inclusive Insurance Participation in Indonesia

Abstract

Inclusive insurance plays a critical role in reducing household vulnerability in developing countries such as Indonesia. This study investigates the factors influencing inclusive insurance participation across regencies in the Special Region of Yogyakarta Province using multinomial logistic regression and stereotype logistic regression. Insurance participation status is classified into three categories: uninsured, government-subsidized, and insured-without-support. Socioeconomic, demographic, and regional characteristics are examined. The results indicate that households with higher spending, higher education, and formal employment are less likely to be uninsured or to rely on government-subsidized insurance. Urban residence has varying effects across regencies. Furthermore, the results from the stereotype logistic regression model suggest that the uninsured group is conceptually closer to the government-subsidized group than to the insured-without-support group. These findings highlight the need for targeted, region-specific policies to expand coverage and facilitate transitions toward stable, non-subsidized insurance, thereby promoting inclusive insurance in Indonesia.

1. Introduction

Insurance plays an important role in reducing vulnerability and strengthening financial resilience against unexpected risks. By providing protection against events such as illness, accidents, or natural disasters, insurance enables individuals and households to manage risk more effectively and maintain livelihood stability. In developing countries such as Indonesia, inclusive insurance is particularly important, as a large share of the population remains exposed to economic and environmental shocks while having limited access to formal risk-sharing mechanisms. Inclusive insurance refers to providing affordable and appropriate insurance products to underserved populations, including low-income, vulnerable, and lower-middle-income groups (Cheston 2018). It often covers specific risks faced by these groups, such as agricultural losses, health emergencies, and natural disasters, which can have significant impacts on their livelihoods and assets. Improving participation in inclusive insurance is critical not only for protecting vulnerable households from financial shocks but also for promoting equitable access to social protection, enhancing economic stability, and supporting broader development goals.

Indonesia, with a population of more than 275 million, exhibits significant disparities in access to insurance due to its diverse socioeconomic conditions and high exposure to various risks. The country is geographically prone to natural disasters, including earthquakes, floods, volcanic eruptions, and tsunamis, which can severely impact communities and local economies. This vulnerability arises from Indonesia’s location within the Pacific Ring of Fire, an area of intense tectonic activity where several tectonic plates converge (Andreastuti et al. 2018). In addition, Indonesia has around 150 active volcanoes (Skoufias et al. 2021), many of which are also part of the Pacific Ring of Fire.

Although the government has introduced several initiatives to expand insurance coverage, including the National Health Insurance program and the development of micro-insurance products, participation rates remain low compared to the country’s total population. As shown in Figure 1, the inclusive insurance rates remained low, at below $20\%$ during the years 2019–2024. According to the National Financial Literacy and Inclusion Survey conducted by the Indonesian Financial Services Authority, this rate is measured by the proportion of respondents who use non-subsidized insurance products or services.

Understanding the determinants of insurance participation is therefore essential for identifying the barriers and enabling factors that influence individuals’ and households’ decisions to obtain coverage. Factors such as consumption expenditure, education, employment type, and geographic location may significantly shape insurance behavior and access. Analyzing these determinants provides valuable insights for policymakers and insurers in designing more effective, inclusive, and sustainable insurance strategies. Ultimately, improving insurance participation contributes not only to enhanced financial security at the household level but also to broader national goals of reducing inequality and strengthening social and economic resilience.

Several studies have assessed the relationship between individual or households’ socioeconomic characteristics and the insurance participation. Atake (2020) showed that the education level of the household head, household size, and total household expenditure affected the decision to seek health care. In addition, the type of insurance, the share of expenditures allocated to food, the distance to the nearest health center, and waiting time significantly influenced the choice of provider. Mulenga et al. (2021) found that age, education, and professional occupation were positively associated with health insurance coverage, while being self-employed in the agricultural sector negatively influenced coverage in Zambia. Being married and having a clerical occupation increased the probability of health insurance enrollment for women, whereas employment in the services, skilled, and unskilled manual sectors increased the probability for men. Furthermore, residing in rural areas reduced the likelihood of having health insurance. Similarly, Salari et al. (2019) reported that education, wealth, marital status, and age were positively associated with enrollment in the National Health Insurance Scheme (NHIS) in Ghana. Household expenditure, used as a proxy for income, was also found to have a statistically significant and positive correlation with NHIS enrollment. The probability of NHIS enrollment additionally varied by occupational status.

This study aims to analyze the determinants of insurance participation in Indonesia. By examining how socioeconomic, demographic, and regional characteristics influence insurance participation, this research seeks to provide empirical evidence on the factors that promote or hinder inclusion in insurance markets. The findings are expected to contribute to the formulation of policies that enhance the accessibility and affordability of insurance products, thereby supporting the development of a more inclusive and resilient insurance system in Indonesia.

In this study, the response variable representing insurance participation consists of three categories: uninsured, government-subsidized, and insured-without-support. Multinomial logistic regression is widely used in actuarial applications to model categorical outcomes with more than two unordered categories. For example, Khemka et al. (2017) applied multinomial logistic regression to analyze how changes in the unemployment rate affect claim incidence in Disability Income Insurance (DII) in Australia, while Kurowski (2021) used multinomial logistic regression to examine the relationship between financial and debt literacy and household overindebtedness during the COVID-19 pandemic.

Although insurance participation categories are often treated as nominal, they may reflect an implicit hierarchy related to financial security, access to services, and socioeconomic status. Conventional multinomial logistic regression assumes that outcome categories are strictly unordered (Jong and Heller 2008), which limits its ability to capture such latent structure. To address this limitation, we incorporate stereotype logistic regression into our analysis. Originally proposed by Anderson (1984), stereotype logistic regression allows intermediate ordinal patterns to emerge from the data without being strictly imposed a priori, providing a more flexible and informative modeling framework. By comparing multinomial and stereotype logistic regression models, this study contributes to the literature by evaluating whether an underlying ordering structure exists and whether accounting for it improves model interpretation and efficiency.

To the best of our knowledge, this study provides a detailed empirical analysis of household insurance participation in Indonesia using both multinomial and stereotype logistic regression. The results reveal a clear socioeconomic gradient among insurance participation categories: the uninsured group is substantially closer to the government-subsidized group than to those insured without support. Moreover, the study identifies two complementary policy pathways for reducing the uninsured population: expanding government-subsidized insurance for low-income households as a short-term goal, and supporting transitions to self-financed coverage through broader socioeconomic development as a long-term goal.

The remainder of the paper is organized as follows. Section 2 describes the dataset, demographic profile of the study area, and variables used in the study. Section 3 presents the multinomial and stereotype logistic regression models and discusses the interpretation of their parameters. Section 4 applies these methods to examine insurance participation in Indonesia. Finally, Section 5 discusses the main findings and proposes policy-relevant pathways for expanding inclusive insurance participation.

2. Data

This study is based on data from the National Socioeconomic Survey (known as SUSENAS), conducted by Statistics Indonesia in March 2023. The survey covers 345,000 sample households across 34 provinces and 514 districts/municipalities in Indonesia. The main sampling frame consisted of 40 percent of the census blocks from the 2020 Population Census, selected using Probability Proportional to Size (PPS), with the number of heads of households from the population list recap of the 2020 Population Census used as the measure of size. These selected census blocks already had assigned strata codes. Prior to forming the main sampling frame, all census blocks from the 2020 Population Census were stratified according to the urban–rural classification. In the March 2023 SUSENAS, 10 households per census block were selected using systematic sampling, resulting in a total sample of 345,000 households from 34,500 census blocks (BPS-Statistics Indonesia 2023). PPS sampling is a technique used to select samples from a finite population in which a size measure is available for each population unit prior to sampling and in which the probability of selecting a unit is proportional to its size (Skinner 2016).

Since the cost of accessing data for the entire country is too high, this study focuses on one province, the Special Region of Yogyakarta. This province is a major tourism destination in Indonesia and is home to an active volcano, Mount Merapi. As a result, the tourism industry faces risks such as volcanic eruptions, which can disrupt operations and affect local employment. Insurance can help stabilize this key sector.

According to BPS-Statistics D.I. Yogyakarta Province (2024), the Special Region of Yogyakarta Province has an area of 3170.65 km² which consists of:

• Kulonprogo Regency, with an area of 577.22 km² (18.21%);
• Bantul Regency, with an area of 511.71 km² (16.14%);
• Gunungkidul Regency, with an area of 1475.15 km² (46.53%);
• Sleman Regency, with an area of 573.75 km² (18.10%);
• Yogyakarta Municipality, with an area of 32.82 km² (1.04%).

The total population of Special Region of Yogyakarta Province in 2023 is 3.736 million people. The largest population in 2023 is in Sleman Regency with a population of 1.157 million people, while the smallest population is in the Yogyakarta Municipality with a population of 0.376 million people (see Figure 2). Then, the population density for each regency can be seen in

Table 1. Population density by regency in 2023.

Regency/Municipality	Abbreviation	Population Density (Persons/km2)
Kulon Progo	KP	750
Bantul	BT	1972
Gunungkidul	GK	509
Sleman	SL	2017
Yogyakarta	YK	11,426

.

According to Table 1, Yogyakarta Municipality has the highest population density, with 11,426 persons/km², while Gunungkidul Regency has the lowest. Yogyakarta Municipality serves as the capital as well as the administrative and economic center of the Special Region of Yogyakarta Province. In contrast, Gunungkidul Regency is located farthest from the provincial capital in terms of both distance and travel time relative to the other regencies.

Figure 3 shows the map of the Special Region of Yogyakarta Province, showing the percentage of population who experienced health problem within the past month by regency in 2023.

The data indicate that a considerable proportion of the population in each regency experienced health problems during the reference period, highlighting the importance of health insurance coverage. In 2023, more than 30% of residents in Kulonprogo Regency (31.96%), Bantul Regency (30.97%), and Yogyakarta Municipality (31.65%) reported health issues, while Sleman Regency (22.71%) and Gunungkidul Regency (22.91%) also had notable shares of individuals facing health problems. These figures suggest that a significant portion of the population may require medical attention at any given time, underscoring the need for comprehensive health insurance to ensure access to healthcare services and protect households from medical expenses.

The list of variables used in this study is presented in

Table 2. Description of variables in the study.

Variable Name	Description
Response Variable
Insurance participation	uninsured, government-subsidized, insured-without-support
Independent Variables
Food consumption expenditure per capita	measured in 100,000 IDR (Indonesian Rupiah)
Age	age of the head of household (in years)
Gender of head of household	female, male
Education of head of household	elementary school and below, high school, college and university
Job of head of household	informal, formal
Marital status of head of household	not married, married
Number of household members	total number of people living in the household
Classification of residence	rural, urban
Regency of residence	Gunungkidul, Kulonprogo, Bantul, Sleman, Yogyakarta

. The data come from 3963 households in the Special Region of Yogyakarta Province.

One of the independent variables is food consumption expenditure, which we use as a proxy for income. In developing countries such as Indonesia, food consumption expenditure has been widely used as a proxy for income because it serves as a more reliable indicator of household living standards. Meyer and Sullivan (2003) noted that consumption is better measured than income for low-income families, particularly where income reporting is prone to underestimation or nonresponse. In many developing countries, income data are difficult to measure accurately because earnings often come from informal activities, seasonal employment, or self-employment, which can lead to substantial reporting errors in household surveys. As a result, researchers frequently rely on consumption-based measures of welfare, which tend to be reported more consistently and better reflect long-term living standards. Consequently, consumption expenditure is widely used in poverty and welfare analysis as a measure of household economic status (Deaton and Zaidi 2002). Hall and Mishkin (1982) also found that food expenditure represents one of the most reliable measures of consumption available in household surveys. For these reasons, food consumption expenditure has been widely adopted in development economics as a practical and empirically supported proxy for household economic status when reliable income data are unavailable.

3. Methods

This study analyzes the effects of several independent variables on the status of household insurance participation in the Special Region of Yogyakarta Province using multinomial and stereotype logistic regressions.

3.1. Multinomial Logistic Regression (MLR)

MLR (Frees 2010) is used to analyze categorical response variables with more than two nominal outcomes. MLR estimates the odds of belonging to a given category relative to a baseline category.

Let Y denote a response variable with m categories and let $\mathbf{x}$ be a vector of p independent variables. In this study, the response variable has $m=3$ categories. We code the response variable Y as $Y=c_1$ (government-subsidized), $Y=c_2$ (uninsured), and $Y=c_0$ (insured-without-support), with $Y=c_0$ used as the baseline category. We denote the two logit functions by

$$f_1\left(\mathbf{x}\right)=ln\left({\displaystyle \frac{Pr(Y=c_1\mid \mathbf{x})}{Pr(Y=c_0\mid \mathbf{x})}}\right)={\beta }_{10}+{\beta }_{11}{x}_1+{\beta }_{12}{x}_2+\cdots +{\beta }_{1p}{x}_p={\beta }_{10}+{\mathbf{x}}^{\prime }{\mathit{\beta }}_{\mathbf{1}},$$

(1)

and

$$f_2\left(\mathbf{x}\right)=ln\left({\displaystyle \frac{Pr(Y=c_2\mid \mathbf{x})}{Pr(Y=c_0\mid \mathbf{x})}}\right)={\beta }_{20}+{\beta }_{21}{x}_1+{\beta }_{22}{x}_2+\cdots +{\beta }_{2p}{x}_p={\beta }_{20}+{\mathbf{x}}^{\prime }{\mathit{\beta }}_{\mathbf{2}}.$$

(2)

It follows that the conditional probabilities of each category of response variable given the independent variable vector are

$$Pr(Y=c_0\mid \mathbf{x})={\displaystyle \frac{1}{1+e^{f_1\left(\mathbf{x}\right)}+e^{f_2\left(\mathbf{x}\right)}}},$$

(3)

$$Pr(Y=c_1\mid \mathbf{x})={\displaystyle \frac{e^{f_1\left(\mathbf{x}\right)}}{1+e^{f_1\left(\mathbf{x}\right)}+e^{f_2\left(\mathbf{x}\right)}}},$$

(4)

and

$$Pr(Y=c_2\mid \mathbf{x})={\displaystyle \frac{e^{f_2\left(\mathbf{x}\right)}}{1+e^{f_1\left(\mathbf{x}\right)}+e^{f_2\left(\mathbf{x}\right)}}}.$$

(5)

The model parameters are estimated by maximum likelihood estimation. Since we are interested in potential interactions between predictors, we consider the following full model that includes all individual independent variables and the following interaction terms:

• Food consumption expenditure × class of residence: The effect of household consumption expenditure on insurance participation may differ by class of residence. Higher consumption expenditure (as a proxy for household income) may have a stronger positive association with insurance uptake among urban households than among rural households.
• Age × job: Among individuals with formal employment, increases in age are associated with a higher likelihood of insurance participation. This suggests that the positive effect of age on insurance uptake is more pronounced for formally employed workers than for those without formal employment.
• Education × job: Occupational outcomes are often structured by educational attainment. Thus, the effect of job type on insurance category may differ across education levels.
• Education × gender: The effect of educational attainment on insurance participation may vary by gender.
• Education × class of residence: The effect of education on insurance participation may differ across classes of residence.
• Additionally, interactions between regency and each predictor were included to allow the backward selection on MLR to identify the best fitted model.

To build the final model, a backward selection procedure is employed. The backward procedure begins with the full model and iteratively selects or removes independent variables based on their Akaike Information Criterion (AIC). A lower AIC value indicates a better model fit.

3.2. Stereotype Logistic Regression (SLR)

In addition to classification, we are interested in the conceptual similarities among the three classes. In particular, given the distinction between individuals in the government-subsidized group and those in the insured-without-support group, our interest lies in assessing how similar or dissimilar the uninsured group is relative to the other two. To address this question within a regression framework, we employ SLR, introduced by Anderson (1984). SLR can be viewed as an extension of ordinal regression that is especially useful when the class ordering is unclear or subject to investigation.

The SLR model expresses the log-odds of each outcome relative to a baseline category as a linear combination of predictor variables scaled by ${\varphi }_j$:

$$ln{\displaystyle \frac{Pr(Y=j\mid \mathbf{x})}{Pr(Y=m\mid \mathbf{x})}}={\alpha }_j-{\varphi }_j{\mathbf{x}}^{\prime }\mathit{\beta },j=1,2,\dots ,m-1,$$

(6)

where m is the baseline or reference category, ${\alpha }_j$ is the intercept term, $\mathit{\beta }$ is a vector of predictor coefficients shared across all outcome categories, and ${\varphi }_j$ is the category score for outcome j which are used to ensure ordinality of the outcomes if the following condition is satisfied:

$$1={\varphi }_1>{\varphi }_2>\cdots >{\varphi }_{m-1}>{\varphi }_m=0.$$

(7)

The $\varphi$ parameters describe how each outcome category is positioned along an underlying scale in the SLR model. Specifically, ${\varphi }_j$ scales the effect of each predictor on the log-odds of being in category j relative to the reference category, such that higher $\varphi$ values indicate a stronger influence of the predictors. In this way, $\varphi$ represents the relative placement of each category, such as different insurance groups, along the underlying scale defined by the predictors. The reference category is assigned $\varphi =0$, establishing a baseline against which all other predictor effects are interpreted.

In this study, there are three categories: $Y=1$ for government-subsidy, $Y=2$ for uninsured, and $Y=3$ for insured-without-support. The insured-without-support category is used as the reference category. Accordingly, we set ${\varphi }_1=1$ and ${\varphi }_3=0$, and estimate ${\varphi }_2$. If ${\widehat{\varphi }}_2>0.5$, indicating that it is closer to ${\varphi }_1$, the uninsured group shares characteristics more similar to the government-subsidy group. In contrast, if ${\widehat{\varphi }}_2<0.5$, indicating that it is closer to ${\varphi }_3$, the uninsured group shares characteristics more similar to the insured-without-support group.

By jointly estimating $\varphi$ and $\beta$, the SLR model captures the ordinal structure while allowing for non-uniform spacing between categories. Note, however, that SLR imposes a common set of regression coefficients $\beta$ across all categories. This constraint reduces model flexibility relative to MLR. MLR estimates separate sets of regression coefficients for each non-reference category, allowing maximum flexibility in capturing category-specific effects. In contrast, SLR introduces a structured form of parsimony by constraining the regression coefficients to follow a common direction across categories while allowing category-specific scaling parameters. This formulation preserves the ordinal structure of the outcome and reduces the number of parameters to estimate, thereby enhancing interpretability. Thus, MLR prioritizes flexibility and predictive fit, whereas SLR emphasizes structural ordering and parsimony.

Parameter estimation in the SLR is performed via maximum likelihood using the rrvglm function in the VGAM package in R 4.4.0 (Yee and Moler 2025). The initial model includes the full set of predictors and selected interactions, and model selection is guided by AIC.

4. Results

4.1. Distribution of Insurance Participation Across Household Characteristics

Table 3. Characteristics of the sample.

Variable	Category	Insurance Participation
		Uninsured	Government-Subsidized	Insured-Without-Support
Gender of head of household	Female	86	436	194
		12.01%	60.89%	27.09%
	Male	383	1968	896
		11.80%	60.61%	27.59%
Education level of head of household	Elementary school and below	155	1221	97
		10.52%	82.89%	6.59%
	High school	272	1028	612
		14.23%	53.77%	32.01%
	College and university	42	155	381
		7.27%	26.82%	65.92%
Job of head of household	Informal	332	1793	573
		12.31%	66.46%	21.24%
	Formal	137	611	517
		10.83%	48.30%	40.87%
Marital status of head of household	Not married	132	568	280
		13.47%	57.96%	28.57%
	Married	337	1836	810
		11.30%	61.55%	27.15%
Classification of residence	Rural	153	960	159
		12.03%	75.47%	12.50%
	Urban	316	1444	931
		11.74%	53.66%	34.60%

presents insurance participation by the gender, education level, employment type, marital status of the household head, and classification of residence. Overall, government-subsidized insurance is the predominant coverage.

Insurance participation does not differ meaningfully by gender or marital status of the household head. Male- and female-headed households exhibit similar enrollment patterns across both subsidized and non-subsidized schemes, and married and unmarried household heads participate at comparable rates.

In contrast, insurance participation varies substantially by education and employment status. Households headed by individuals with elementary education or less and engaged in informal employment are disproportionately represented in government-subsidized insurance. Conversely, households with college- or university-educated heads show markedly higher participation in non-subsidized insurance, reflecting the influence of both educational attainment and labor-market position on the ability to obtain coverage independently of government support.

Insurance participation also differs by place of residence. Rural households are more frequently enrolled in government-subsidized insurance, whereas urban households are more likely to possess non-subsidized coverage, indicating spatial disparities in socioeconomic conditions and access to contributory insurance schemes.

4.2. Results from the MLR Model

A backward selection procedure on the MLR model is conducted to examine the effects of the independent variables on insurance participation status. The final model ($\mathrm{AIC}=5775.581$), presented in

Table 4. Final selected model from the MLR model with the base category “insured-without-support”.

	Coefficient of Government-Subsidized Model					Coefficient of Uninsured Model
Variable	KP	BT	SL	YK	GK	KP	BT	SL	YK	GK
Intercept	7.667	4.742	4.828	5.044	6.509	5.452	3.721	3.186	5.408	5.668
Food consumption expenditure per capita	−0.250	−0.116	−0.152	−0.099	−0.097	−0.103	−0.028	−0.058	−0.061	−0.058
Age	−0.041	−0.029	−0.014	0.004	−0.041	−0.052	−0.037	−0.031	−0.058	−0.064
Education_high school × job_informal			−1.670					−1.253
Education_college and university × job_informal			−2.438					−1.997
Education_high school × job_formal	−3.599	−2.655	−2.192	−1.754	−2.727	−3.346	−2.368	−1.264	−2.192	−2.385
Education_college and university × job_formal	−4.990	−4.046	−3.583	−3.145	−4.118	−4.694	−3.716	−2.612	−3.540	−3.733
Number of family members			−0.062					−0.115
Class of residence_urban	−0.397	0.547	−1.068	−2.006	−0.541	−0.069	−0.023	−0.818	−1.704	−1.445

Coefficients shown in bold and italics are identical across regencies, indicating no statistically significant interaction with regency. Coefficients that differ across regencies indicate statistically significant regency interaction effects at the 0.05 significance level.

, includes several interaction terms between regency and selected independent variables. To facilitate interpretation of these interactions, the results are reported separately for each regency.

Table 4 presents the final MLR model that estimates the likelihood of participating in government-subsidized insurance or being uninsured relative to the base category, insured-without-support. The gender and marital status of the household head are excluded from the final model, indicating that these factors are not statistically significant after controlling for other characteristics. This result is consistent with previous studies, including Chauluka et al. (2022) in Malawi and Kimani et al. (2014) in Kenya, which reported no significant association between the gender of the household head and the ownership of health insurance. Similarly, Okumu et al. (2024) found that marital status did not significantly influence enrollment in Kenya’s National Health Insurance Fund.

Across the five regencies, several consistent determinants of insurance participation emerge. Higher food consumption expenditure per capita is consistently associated with lower odds of both receiving government-subsidized insurance and being uninsured, relative to being insured-without-support. This finding aligns with Birhanu et al. (2025), who reported that better food security increased willingness to join community-based health insurance in Ethiopia.

Age displays predominantly negative associations with the likelihood of being uninsured and generally negative or small effects on the likelihood of receiving government-subsidized insurance. Similar age-related patterns were reported by Duku (2018) in Ghana. The slight positive coefficient observed in Yogyakarta Municipality (0.004 for the subsidized category) is negligible and does not alter the overall trend.

Education and employment status jointly play a central role in distinguishing insurance participation categories. Individuals with a high school education or higher who work in informal jobs exhibit markedly lower odds of being in both the subsidized and uninsured groups. The effects are even stronger among those in formal employment, with large negative coefficients indicating that better-educated, formally employed individuals are substantially less likely to rely on government subsidies or to lack insurance coverage. These results are consistent with Kimani et al. (2014), who found that formal-sector employment and attainment of secondary education or higher increased the likelihood of owning health insurance.

The number of family members shows a uniform negative association with the likelihood of being in the subsidized or uninsured categories. Larger households are less likely to fall into either group relative to being insured-without-support. Similar associations between household size and insurance uptake were documented by Sendekie et al. (2024) in Northwest Ethiopia and Macha et al. (2014) in Tanzania.

The effect of urban residence varies substantially across regencies, indicating important geographic heterogeneity in insurance participation dynamics. In several regencies, urban residents are less likely to receive subsidized insurance or to be uninsured, suggesting broader access to employment-linked coverage or improved socioeconomic resources. In contrast, the positive coefficient for government-subsidized insurance in Bantul suggests that urban residents there are slightly more likely to receive subsidies, potentially reflecting local socioeconomic conditions or program targeting.

In summary, the final model identifies food consumption expenditure per capita, age, education–employment interactions, household size, and class of residence as significant determinants of insurance participation. Higher household expenditure and larger family size are consistently associated with lower odds of receiving government-subsidized insurance and being uninsured. Older household heads are less likely to be uninsured, with smaller and more variable effects on subsidized participation. Strong negative coefficients for education–job interactions indicate that individuals with higher education, particularly those in formal employment, are substantially less likely to rely on government subsidies or to remain uninsured. Urban residence exhibits heterogeneous effects across regencies, with some areas showing a reduced likelihood of subsidized participation or uninsured status, underscoring important regional differences in insurance participation patterns.

4.3. Results from the SLR Model

An SLR model is used to examine the latent ordering of the insurance participation categories. The results reveal a clear socioeconomic hierarchy underlying these categories: households with government-subsidized insurance occupy the lowest position on this gradient, the uninsured form an intermediate group, and those insured-without-support represent the highest socioeconomic position. The estimated category scores ($\widehat{\varphi }$) confirm this ordering, with ${\varphi }_1=1$ for the government-subsidized group, ${\widehat{\varphi }}_2=0.673$ ($95\%$ CI: 0.602–0.745) for the uninsured group, and ${\varphi }_3=0$ for the insured-without-support. These estimated category scores reveal that the uninsured group lies substantially closer to the government-subsidized group than to those insured-without-support. This indicates that uninsured households share socioeconomic and structural characteristics more similar to subsidized beneficiaries than to insured-without-support households. Moreover, the narrow confidence interval for the uninsured category indicates that its position differs significantly from both the government-subsidized and the insured-without-support groups.

The final model after model selection (AIC = 5885.759) is presented in

Table 5. Final selected model from the SLR model with the base category “insured-without-support”.

	Estimated Coefficient of					Estimated Coefficient of
	Government-Subsidized Model					Uninsured Model *
Variable	KP	BT	SL	YK	GK	KP	BT	SL	YK	GK
Intercept	7.643	5.886	4.489	3.928	6.430	4.018	2.836	1.894	1.517	3.202
Food consumption expenditure per capita	−0.243	−0.096	−0.133	−0.104	−0.103	−0.164	−0.064	−0.090	−0.070	−0.069
Age × job_informal	−0.041	−0.034	−0.021	−0.001	−0.039	−0.028	−0.023	−0.014	−0.001	−0.026
Age × job_formal	−1.994	−1.116	−0.585	−0.276	−1.119	−1.342	−0.751	−0.394	−0.186	−0.753
Education_high school × job_informal			−1.894					−1.275
Education_college and university × job_informal			−2.010					−1.353
Education_high school × job_formal	−4.445	−3.573	−3.056	−2.767	−3.572	−2.991	−2.405	−2.057	−1.862	−2.404
Education_college and university × job_formal	−5.206	−4.335	−3.818	−3.528	−4.333	−3.504	−2.918	−2.570	−2.375	−2.916
Number of family members			−0.074					−0.050
Class of residence_urban			−0.455					−0.306
Gender_female × education_high school			−1.894					−1.275
Gender_female × education_college and university			−2.010					−1.353
Gender_male × education_high school			−1.636					−1.101
Gender_male × education_college and university			−2.578					−1.735

* Estimated Coefficient of Uninsured Model $={\widehat{\varphi }}_{uni}$ × Estimated Coefficient of Government-Subsidized Model. Coefficients shown in bold and italics are identical across regencies, indicating no statistically significant interaction with regency. Coefficients that differ across regencies indicate statistically significant regency interaction effects at the 0.05 significance level.

.

Using insured-without-support as the reference category, the results indicate that higher education, formal employment, urban residence, and greater food consumption per capita are consistently associated with a lower likelihood of being government-subsidized or uninsured across all regencies. Age interacted with formal employment also shows a negative effect, suggesting that older individuals in formal jobs are less likely to rely on subsidized insurance or to be uninsured. The effects of family size and gender interactions with education are modest but negative, indicating that larger households and males with higher education are more likely to be insured-without-support.

When compared with the MLR model, the overall patterns are similar: both models highlight education, job type, urban residence, and expenditure as key predictors. The results suggest that the relationships between the predictors and the response are mostly robust to the ordering assumption regarding the class, as well as to model complexity.

5. Discussion

In this study, insurance participation is measured by households’ enrollment status: government-subsidized, uninsured, or insured-without-support. These outcomes are analyzed using MLR and SLR. While MLR yields a lower AIC (5775.58) than SLR (5885.76), indicating better in-sample fit, this difference should be interpreted in light of their distinct modeling objectives. Although the AIC penalizes model complexity, MLR remains favored because it relaxes the proportional-odds constraint and estimates separate coefficients for each category comparison, thereby maximizing flexibility in capturing category-specific variation. In contrast, SLR imposes an ordering structure and proportional constraints across categories. Although this restriction results in a higher AIC, it enhances parsimony and interpretability by revealing the systematic progression from subsidized to uninsured to fully insured status. Thus, the observed AIC difference reflects a fundamental trade-off between predictive fit and structural clarity. MLR prioritizes predictive accuracy by allowing greater parameter flexibility, whereas SLR emphasizes interpretability and theoretical coherence by leveraging the ordinal nature of the outcome. Importantly, SLR provides additional explanatory value by identifying whether covariates consistently shift households along an ordered participation gradient, offering substantive insight into insurance mobility that MLR does not explicitly impose. We also considered an adjacent-category logit model, a more flexible extension of SLR, which produced an AIC of 5780.93, comparable to that of MLR. However, we do not pursue this model further, as its additional complexity substantially complicates interpretation without providing commensurate substantive insight.

Inclusive insurance means making insurance affordable and suitable for people who are not well covered by traditional insurance, especially low-income and vulnerable groups, while also including the lower middle class. In line with this concept, the MLR results show that uninsured households are more likely to have lower food consumption (used as a proxy for income), informal employment, lower education levels, and to vary across regencies. These results indicate that lower insurance participation is associated with limited ability to pay, unstable employment, and unequal access across regions. The SLR results reveal a clear socioeconomic gradient in insurance participation. Households receiving government-subsidized insurance are the most socioeconomically disadvantaged. Uninsured households fall in an intermediate position, while households insured without support are the most advantaged. This pattern suggests that insurance participation primarily reflects households’ socioeconomic capacity rather than random enrollment. Overall, these findings indicate that reducing the number of uninsured households should be a key policy priority for improving inclusive insurance participation.

There are two main pathways for reducing the uninsured population. The first pathway moves uninsured households into government-subsidized insurance (Path 1), while the second pathway moves uninsured households directly into insurance without support (Path 2). These pathways are illustrated in Figure 4.

The SLR results indicate that uninsured households are socioeconomically closer to the subsidized group than to the insured-without-support group. This suggests that most uninsured households lack the capacity to pay full premiums and are therefore more likely to access insurance through subsidies. Consequently, Path 1 represents the most practical and effective short-term strategy. Path 2 remains important, but it primarily applies to households with higher education, more stable employment, and higher living standards. Expanding subsidized insurance, however, requires substantial government funding, raising concerns about long-term sustainability.

For this reason, a secondary, long-term goal is necessary: supporting transitions from government-subsidized insurance to insured-without-support. Unlike the primary goal, which can be achieved relatively quickly through subsidies, this secondary goal depends on improvements in households’ socioeconomic conditions. Both the MLR and SLR results show that higher education, formal employment, urban residence, and greater food consumption expenditure are key factors associated with participation without support. These shared factors suggest that achieving this long-term goal requires broader improvements in education, employment quality, and living standards. Policies that support skills development, labor market formalization, and income growth can gradually increase households’ ability to finance their own insurance, thereby reducing reliance on subsidies over time.

Based on these findings, several policy implications emerge. To achieve the primary goal of reducing the uninsured population, policymakers should prioritize expanding government-subsidized insurance for households with low income, informal employment, and limited education. This can be facilitated through simplified enrollment procedures, improved outreach, and automatic enrollment linked to other social assistance programs. At the same time, uninsured households with greater capacity should be encouraged to enroll in insurance without support through information campaigns, flexible premium arrangements, and temporary incentives.

To support the secondary goal, insurance policy should be integrated with broader social and economic development strategies. Investments in education, job training, and formal employment can strengthen households’ ability to transition from subsidized insurance to self-financed coverage. Gradual and well-monitored exit mechanisms from subsidies are important to avoid coverage loss while ensuring that public resources remain focused on households that need them most.

Overall, the findings suggest that inclusive insurance in Indonesia should be approached as a dynamic process. Short-term policies should focus on reducing the uninsured population through targeted subsidies, while long-term policies should strengthen households’ capacity to participate without support. This combined approach can expand coverage, improve equity, and maintain fiscal sustainability within the insurance system. However, this study is limited to the Special Region of Yogyakarta due to cost constraints associated with obtaining the data. As socioeconomic conditions, fiscal capacity, and disaster risk characteristics vary across Indonesian provinces, the findings may not be fully generalizable to other regional contexts. Therefore, extending the analysis to other provinces with different structural and risk profiles or to the national level remains a direction for future research.

Furthermore, while this study identifies the main pathways of insurance participation, future research should examine reverse or alternative transitions within the system using panel data and dynamic transition frameworks. For example, Markov models (either a discrete-time Markov chain or a continuous-time Markov process) or multi-state survival analysis (see, e.g., Cook and Lawless 2018) could be employed to track movements among subsidized, uninsured, and non-subsidized insurance statuses over time. Such approaches would allow the quantification of transition and stationary probabilities and the identification of key triggering events, such as income shocks, job loss, health shocks, or policy changes, that drive shifts in coverage. Households may move from non-subsidized to subsidized insurance following adverse economic shocks or may transition from subsidized coverage to uninsured status due to administrative or eligibility barriers. Modeling these dynamics across diverse provincial settings would provide a more comprehensive understanding of insurance mobility and vulnerability, thereby enabling the design of adaptive programs that respond to both upward and downward movements in coverage and support region-sensitive policies.