Impact of Public Health Insurance on Reducing Out-of-Pocket Payment for Psychiatric Care Among Cross-Border Migrants in Thailand: A Two-Part Model and Machine Learning Approach

Abstract

Purpose

Thailand has long implemented public insurance schemes for cross-border migrants, though psychiatric care is not explicitly included in the schemes’ benefit package. As evidence related to the association between public insurance and migrants’ out-of-pocket (OOP) expenditure for psychiatric services remains limited, this study aims to assess if and to what extent the insurance is associated with OOP spending among migrants utilizing psychiatric-related care in Thailand.

Patients and Methods

A cross-sectional analysis of routine service data was applied. We analyzed nationwide service utilization data of cross-border migrants receiving psychiatric care in public hospitals between 2018 and 2022. A two-part model (TPM) with machine learning was employed. The first part used a decision tree classifier to estimate the probability of incurring OOP spending, and the second part applied a non-negative least squares ensemble to predict the magnitude of OOP spending.

Results

A total of 4,559 migrant patients accounted for 13,744 visits (inpatient and outpatient care combined), of which 84.1% were uninsured. The mean OOP spending per visit was 449 Baht, ranging from zero to 109,075 Baht. Public insurance was associated with about 50% reduction in the likelihood of incurring any OOP payment. After integrating both parts of the model, the estimated mean OOP spending per visit was 471.3 Baht (95% confidence interval [CI] = 433.8–511.3) among the uninsured, compared with 7.4 Baht (95% CI = 6.7–8.1) among the insured—about 98% OOP spending avoidance.

Conclusion

Public insurance for cross-border migrants in Thailand was associated with reduced OOP spending for psychiatric services at the point of care. Inclusion of psychiatric services in migrant health insurance benefit packages may be considered as an option to help reduce inconsistent interpretations of coverage. Further research examining the financial burden of psychiatric illnesses among migrant households would be valuable. Ongoing monitoring of migrant-related policy changes may help inform future policy decisions.

Introduction

Mental health is one of the key public health concerns globally. The World Health Organization (WHO) reported that about 970 million people were suffering from mental disorders.¹ Besides, 16% of global disability-adjusted life years (DALY) were attributable to mental disorders.² Cross-border migrants are among the populations most affected by mental health problems. While some remain healthy, many—particularly refugees, asylum seekers, and undocumented migrants—face limited access to healthcare.³

Thailand is one of the main destination countries for migrants in Southeast Asia, hosting over 5 million non-Thai nationals.⁴ Approximately half of this population consists of migrant workers from Cambodia, Lao PDR, Myanmar, and Vietnam (CLMV). Notably, this figure does not include undocumented migrants, whose true numbers remain uncertain.

Regarding health security, Thailand has two main insurance schemes for migrants: the Social Security Scheme (SSS) and the Health Insurance Card Scheme (HICS). The SSS, introduced in 1991 and governed by the Ministry of Labor (MOL), covers documented migrants in the formal sector and is financed through tripartite contributions from employees, employers, and the government. In contrast, the HICS, introduced in 2004 and managed by the Ministry of Public Health (MOPH), primarily targets CLMV migrants in the informal sector who are registered under temporary leniency policies. It is financed through an annual premium and an upfront health check fee. Both schemes provide comprehensive benefits, including outpatient, inpatient, and high-cost care, and are delivered mostly through public hospitals under the Ministry of Public Health.⁵

However, the benefit does not cover treatment for “psychiatric diseases”. Note that the MOPH and the MOL have not indicated the list of specific diagnoses according to the International Classification of Diseases, 10^th Revision (ICD-10) that are not covered by the mentioned public insurance schemes. Another key feature of the HICS is that about 60% of the insurance revenue is pooled at each registered hospital. The MOPH only manages a small portion (about one-fifth) of the premium earmarked for high-cost treatment.⁶ This means, in practice, migrants with psychiatric diseases can be covered by the HICS based on the discretion of the healthcare providers as most of the insurance funds are independently managed by the registered hospitals.

Studies on psychiatric diseases among migrants in Thailand are neglected. Though some exist, most focus on prevalence and epidemiological profiles of migrant patients. For example, Sikhiwat et al⁷ reported a 239% prevalence of depression among CLMV migrants. Aung et al⁸ highlighted that longer duration of stay in the host country and being female were significantly associated with depression.

Since no definitive list exists of psychiatric conditions covered under public insurance schemes for migrants, we adopted the broader term psychiatric-related illness throughout the analysis. In addition, evidence on the financial burden associated with psychiatric service utilization in Thailand remains extremely limited. To date, no studies have specifically examined how public insurance schemes—particularly HICS and SSS—are associated with migrants’ out-of-pocket (OOP) spending at the point of care. These gaps in knowledge highlight the limited understanding of both the scope of coverage and the extent of financial protection for psychiatric care among migrant populations.

Therefore, this study aims to examine the extent to which public insurance for migrants is associated with reductions in OOP spending from service utilization psychiatric-related illnesses for cross-border migrants in Thailand.

Materials and Methods

Study Design and Data Source

A cross-sectional study on real-world nationwide hospital utilization data was conducted. We obtained service utilization data for CLMV migrants from the Health Data Center (HDC) of the MOPH for the period, 2018–2022. The HDC is a national data collection system that compiles service utilization records from patients attending MOPH-affiliated hospitals. We analyzed both inpatient admissions and outpatient visits. The inclusion criteria comprised all psychiatric-related visits made by migrant individuals, regardless of insurance status. These were defined as records coded using ICD-10 codes beginning with F and G codes (eg., F41.0 [panic disorder], F15.2 [mental and behavioral disorders due to use of other stimulants], F20.0 [paranoid schizophrenia], and G47 [sleep disorders]). The inclusion of both F and G codes was intended to maximize coverage of psychiatric-related illnesses, given the lack of specificity in insurance regulations regarding which diagnoses are included under psychiatric conditions.

With the above approach, we had retrieved 344,979 outpatient visits and 10,661 inpatient visits. Duplicate visits were identified using a combination of encrypted citizen identity number, visit date, and date of birth, while unique individuals were identified using encrypted citizen identity number and date of birth. Subsequently, the de-duplicated dataset comprised 26,941 individuals and 87,792 visits. Records with implausible extreme values suggestive of miscoding (for instance, visits with health expenditures exceeding 200,000 Baht) and those with missing OOP spending were excluded. Finally, the final dataset included 4,559 individuals and 13,744 visits, Figure 1.

Figure 1.

Flow of data collection and cleaning process.

Since all records meeting the inclusion criteria were analyzed, sample size calculation was not relevant. Nonetheless, based on a previous study⁹ in Thailand reporting mean OOP expenditure for outpatient care (not limited to psychiatric services) of 420 Baht (standard deviation [SD] = 806) among the uninsured and 22 Baht (SD = 147) among the insured, a minimum sample size of approximately 70 observations (35 per group) would be required using a two-sample comparison of means. This is substantially lower than the final analytical sample.

Variable Management

We collected the following variables: age, sex, ICD-10 diagnoses, nationality, place of visit (province), date of visit, occupation (such as general workers, students, and farmers), insurance status (insured either by the SSS or the HICS versus uninsured), and payment by the patients (in Baht [33 Baht ~ US$ 1]). The unit of analysis is visit. Yet, in some parts of descriptive analysis, we grouped multiple visits within a person together to analyze the demographic profile of the patients.

Data Analysis

The analysis is divided into two main strands. In the first strand, demographic data were analyzed. Categorial data were presented in terms of frequency and percentage while numerical data were presented by mean, SD, median and range. Comparison of key characteristics between insured and uninsured participants was conducted using the Chi-square test for categorical variables (such as sex and diagnosis codes) and the Mann–Whitney U-test for continuous variables (such as age and OOP spending).

In the second strand, we used the concept of TPM to address the main study objectives. The TPM is one of the widely used econometric techniques for analyzing healthcare expenditures, suited for datasets exhibiting excessive zero expenditure observations and a highly skewed distribution among those with positive expenditures.^10,11

Conventionally, the TPM uses a logit or probit model to estimate the probability of zero expenditure (first part), while employing an ordinary least square (OLS) or generalized linear model (GLM) to evaluate the level of positive expenditures (second part). This two-part approach effectively accounts for the fact that excess zeros may arise from a mechanism different from that causing positive expenditures. However, in recent years, novel analytic techniques, such as machine-learning algorithms, have been increasingly applied in public health research. As a result, we integrated machine-learning algorithms into the TPM framework to improve predictive power, rather than relying solely on conventional methods.

In the first part, a decision-tree–based algorithm (CatBoost classifier) was employed to estimate the probability of incurring OOP spending, alongside other machine-learning classifiers. The CatBoost algorithm was selected as the primary classifier and regressor for several reasons. First, it natively handles categorical features (such as nationality, occupation, and province) using an advanced symmetric tree structure, which reduces the need for extensive pre-processing. Second, CatBoost is less prone to overfitting compared to other gradient boosting decision trees. This was particularly crucial given the highly skewed nature of healthcare expenditure data.¹² We defined an occurrence of OOP spending as a payment of at least 30 Baht. This cutoff was selected based on the historical context of Thailand’s Universal Coverage Scheme (UCS), which, when introduced in 2002, required a 30-Baht copayment per visit and was widely known as the “30-Baht scheme.”¹³ Although this mandatory fee was abolished in 2006,¹⁴ some healthcare facilities continue to request a 30-Baht visiting fee on a voluntary basis. In practice, such payments may also be applied to beneficiaries of insurance schemes other than the UCS.

Predictor variables included demographic characteristics (age, sex, age group, nationality, and occupation), diagnostic information (diagnosis code, and diagnosis type), service-related features (facility, province, and visit type), insurance status (insured versus uninsured), and temporal indicators (year, month, day of week, and weekend flag). The same set of predictors was used in the second part of the TPM. Model optimization for the CatBoost algorithm was conducted using systematic grid search and randomized search strategies to identify optimal hyperparameters. Tuned parameters included learning rate, tree depth (restricted between 4 and 10), L2 leaf regularization, and the number of iterations with early stopping based on validation loss.

To evaluate model performance, we applied stratified k-fold cross-validation. This stratification ensured that the imbalance between the insured and uninsured groups and the proportion of zero (below 30-Baht) versus positive expenditure visits was consistently maintained across all folds. Each iteration utilized 80% of the data for model training and 20% for validation.

Model performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (PR-AUC), accuracy, F1-score, sensitivity, specificity, log loss and Brier score. Variable importance was assessed through model-specific feature contribution measures.

In the second part of the TPM, regression algorithms including LightGBM, XGBoost, and CatBoost were trained on log-transformed OOP payment to estimate conditional payment amounts among the payers. Their outputs were combined using a non-negative least squares (NNLS) ensemble to enhance model stability and generalizability. After applying weight to multiple models, we reduced the variance associated with any single algorithm, resulting in a more stable prediction of the magnitude of OOP spending.

Model performance was evaluated using the coefficient of determination (R²), root mean squared error (RMSE), mean absolute error (MAE), and median absolute error (MedAE). For comparison, conventional TPM employing logit and GLM were also applied.

The expected OOP spending per visit was computed by multiplying the predicted probability of payment by the estimated conditional OOP payment. Differences in expected OOP spending between insured and uninsured migrants were summarized using mean differences and 95% confidence intervals (CI). We also performed a subgroup analysis limited to utilization records with ICD-10 codes beginning with F, since these codes are directly related to psychiatric diseases, whereas codes starting with G typically involve nervous system disorders.

All analyses were performed in Python (packages: pandas, numpy, scikit-learn, lightgbm, xgboost, catboost, statsmodels, and scipy) and Microsoft Excel.

Results

A total of 4,559 migrant patients were explored, resulting in 13,744 visits. Males outnumbered females (53% versus 47%). Mean age and median age of the patients were 36 and 35 years respectively. About four-fifths (84%) of the patients were uninsured. While most patients (62%) were general workers, about 10% were unemployed, Table 1.

Table 1.

Individual Characteristics of Migrants Utilizing Psychiatric Services in Thai Public Hospitals, 2018–2022

Demographic Profiles (N = 4,559)	Value
Sex–number (%)
Male	2,417 (53.0%)
Female	2,142 (47.0%)
Mean age (standard deviation)—years	36.2 (15.2)
Median age (P25, P75)—years	35 (26,45)
Insured status— number (%)
Insured	723 (15.9%)
Uninsured	3,836 (84.1%)
Occupation— number (%)
General worker	2,830 (62.1%)
Unemployed	446 (9.8%)
Student	240 (5.3%)
Gardener	145 (3.2%)
Others	898 (19.7%)

In terms of demographic characteristics by insurance status, uninsured migrants exhibited a higher proportion of males compared with insured migrants (55.7% versus 39.0%). The median age of insured migrants was slightly higher than that of uninsured migrants (41 years versus 34 years). Although statistically significant differences were observed across all demographic variables (P-value < 0.001), the overall occupation profiles of the two groups were relatively similar, with general workers comprising the largest proportion, Table 2.

Table 2.

Comparing Individual Characteristics of Migrants Utilizing Psychiatric Services in Thai Public Hospitals During 2018–2022 by Insurance Status

Demographic Profiles	Insured	Uninsured	P-value
Sex—number (%)
Male	282 (39.0%)	2,135 (55.7%)	<0.001
Female	441 (61.0%)	1,701 (44.3%)
Median age (P25, P75)—years	41 (30, 53)	34 (26, 43)	<0.001
Occupation-number—number (%)			<0.001
General workers	494 (68.3%)	2,336 (60.9%)
Unemployed	21 (2.9%)	425 (11.1%)
Students	50 (6.9%)	190 (5.0%)
Gardeners	8 (1.1%)	174 (4.5%)
Others	150 (20.8%)	711 (18.5%)

Note: Chi-square test was used in all variables, except for age where Mann–Whitney U-test was applied.

Regarding visit profiles, the peak of visits presented in 2019 (30%). F15.2 (mental and behavioral disorders due to the use of stimulants) was the most common diagnosis. Chiang Mai, Chiang Rai and Kanchanaburi—areas along Thailand-Myanmar border—were provinces which contributed to most of the visits. Mean OOP and median OOP spending per visit were 448.9 Baht and 100 Baht, respectively. The range of OOP spending per visit varied from zero to 109,075 Baht, Table 3.

Table 3.

Hospital Visit Profiles of Migrants Utilizing Psychiatric Services in Thai Public Hospitals, 2018–2022

Visit Profiles (N = 13,744)	Value
Number of visits over study period—number (%)
2018	2,954 (21.5%)
2019	4,075 (29.6%)
2020	3,059 (22.3%)
2021	2,022 (14.7%)
2022	1,634 (11.9%)
Amount of visits by insurance status of the users—number (%)
Insured	2,106 (15.3%)
Uninsured	11,638 (84.7%)
Top 5 most prevalent diagnoses—number (%)
F15.2 - Mental and behavioral disorders due to stimulants	1,126 (8.2%)
G44.2 - Tension-type headache	534 (3.9%)
G40.9 - Epilepsy, unspecified	480 (3.5%)
F20.0 - Paranoid schizophrenia	424 (3.1%)
G47.0 - Insomnia, disorders of initiating and maintaining sleep	382 (2.8%)
Top 5 most prevalent provinces–n (%)
Chiang Mai	3,814 (27.8%)
Chiang Rai	1,447 (10.5%)
Kanchanaburi	1,106 (8.0%)
Trat	982 (7.1%)
Roi Et	788 (5.7%)
Mean out-of-pocket payment per visit—Baht (standard deviation)	448.9 (2,751.6)
Median out-of-pocket payment per visit—Baht (P25, P75)	100.0 (0.0, 351.0)
Minimum out-of-pocket payment per visit—Baht	0
Maximum out-of-pocket payment per visit—Baht	109,075

Most visits occurred during 2018–2019 among both insured and uninsured migrants. F-code diagnoses were more predominant among uninsured migrants. Approximately four-fifths of insured migrants utilized healthcare services in hospitals located in border provinces. Statistically significant differences were observed in diagnosis code category and province of visit (P-value < 0.001), Table 4.

Table 4.

Comparing Hospital Visit Profiles of Migrants Utilizing Psychiatric Services in Thai Public Hospitals During 2018–2022 by Insurance Status

	Insured	Uninsured	P-value
Year—number (%)			0.053
2018	229 (31.7%)	1,028 (26.8%)
2019	196 (27.1%)	1,067 (27.8%)
2020	137 (18.9%)	770 (20.1%)
2021	93 (12.9%)	511 (13.3%)
2022	68 (9.4%)	460 (12.0%)
Diagnosis starting code—number (%)			<0.001
F code	394 (54.5%)	3,037 (79.2%)
G code	329 (45.5%)	799 (20.8%)
Province of visits—number (%)			<0.001
Border	598 (82.7%)	2,584 (67.4%)
Non-border	125 (17.3%)	1,252 (32.6%)

Note: Chi-square test was used in all variables.

Overall, the OOP spending per visit remained relatively stable throughout the study period. Among the uninsured patients, the median OOP spending per visit was 142 Baht (range = 0–109,075 Baht), whereas the insured patients typically paid zero Baht at the median (range = 0–10,664 Baht). The mean OOP spending per visit among uninsured patients was 526.1 Baht (SD = 2,981.3 Baht), about 23 times higher than that of insured patients (mean = 22.6 Baht, SD = 287.3 Baht). A clear difference in OOP spending was observed by insurance status, whereas differences by sex and diagnosis code category were less apparent, as shown in Figures 2 and 3.

Figure 2.

Out-of-pocket payment per visit in each year (2018–2022) sorted by insured status.

Note: The bold horizontal line within each box represents the median out-of-pocket payment. The vertical axis is displayed on a logarithmic scale to improve visualization. The red dashed line refers to the 30-Baht threshold.

Figure 3.

Out-of-pocket payment per visit over the study period (2018–2022) sorted by insured status, diagnosis codes, and sex.

Note: The bold horizontal line within each box represents the median out-of-pocket payment. The vertical axis is displayed on a logarithmic scale to improve visualization. The red dashed line refers to the 30-Baht threshold.

According to CatBoost classifier to model the first part of the TPM, we found that the AUC-ROC was 0.986 while the PR-AUC was 0.985. Accuracy, F1-score, sensitivity, and specificity were all over 0.90, corresponding with log loss and Brier score, numbering about 0.137 and 0.036 respectively, Figure 4.

Figure 4.

Area under receiver operating characteristic curve of CatBoost classifier to estimate the probability of incurring out-of-pocket spending.

Note: AUC-ROC = 0.986; PR-AUC = 0.985; Brier score = 0.036; Log loss = 0.137; Accuracy at 0.5 = 0.954; F1 at 0.5 = 0.958; Sensitivity = 0.979; Specificity = 0.925.

In the second part, CatBoost regressor outperformed other regression techniques, namely Light GBM and XGBoost regressors. When combining all techniques by NNLS ensemble, the model yielded R² of 83.4%, larger than all other models, and RMSE, MAE, and MedAE, were close to those of CatBoost regressor. After merging the first and the second parts (CatBoost classifier and NNLS ensemble), RMSE, MAE, and MedAE dropped markedly to 1,245.7, 80.9 and 0.6 respectively, while maintaining high R² as 79.5%. In contrast, conventional TPM yielded lower R² and higher RMSE, MAE, and MedAE, relative to CatBoost classifier combined with NNLS ensemble, Table 5.

Table 5.

Conditional Mean Out-of-Pocket Payment (Given Non-Zero) Estimated by Ensemble Regression Models

Model	R2 (%)	RMSE—Baht per Visit	MAE—Baht per Visit	MedAE—Baht per Visit
NNLS ensemble	83.4	1,508.9	129.6	6.2
CatBoost regressor	82.7	1,538.8	126.1	4.9
XGBoost regressor	79.8	1,663.6	162.8	9.0
LightGBM regressor	76.1	1,807.9	184.8	9.6
Two-stage*	79.5	1,245.7	80.9	0.6
Conventional approaches**	10.7	2,600.8	345.0	49.7

Note: Coefficient of determination (R²), root mean squared error (RMSE), mean absolute error (MAE), median absolute error (MedAE), generalized linear model (GLM), *Two-stage = CatBoost classifier+ NNLS ensemble, **Conventional approaches = Logit + GLM.

The final model suggested that the predicted probability of incurring payment reduced from 24.0% among insured migrants to 59.9% among uninsured migrants. Focusing only on the payers, the conditional mean OOP spending per visit was 38.5 Baht (95% CI = 36.2–41.0) among the insured—about 95% OOP spending avoidance. With both parts combined, the expected OOP spending per visit was 7.4 Baht for insured individuals, equivalent to 464 Baht saved or 98% less payment, Table 6.

Table 6.

Conditional Mean Out-of-Pocket Payment After Combining Both Parts of the Two-Part Model

Group	Predicted percent of Visits Incurring OOP Payment—% per Visit (95% CI) *	Predicted OOP Payment Among the Payers—Baht per Visit (95% CI) **	Predicted OOP Payment after Accounting for Probability of Incurring Payment—Baht per Visit (95% CI) ***
-Insured	24.0 (22.4–25.7)	38.5 (36.2–41.0)	7.4 (6.7–8.1)
-Uninsured	59.9 (59.1–60.8)	850.4 (777.5–934.0)	471.3 (433.8–511.3)

Note: *first part, **second part, ***combining both parts.

Abbreviation: OOP, Out-of-pocket.

The subgroup analysis yielded results consistent with those from the complete dataset. The predicted OOP spending in this regard was 394.8 Baht (95% CI = 368.0–423.1) for the uninsured and 8.4 Baht (95% CI = 7.4–9.5) for the insured—equivalent to approximately 98% saving, Table 7.

Table 7.

Conditional Mean Out-of-Pocket Payment After Combining Both Parts of the Two-Part Model by Restricting the Analysis Only to ICD-10 Codes Beginning with F

Group	Predicted Percent of Visits Incurring OOP Payment—% per Visit (95% CI) *	Predicted OOP Payment Among the Payers—Baht per Visit (95% CI) **	Predicted OOP Payment after Accounting for PROBABILITY of Incurring Payment—Baht per Visit (95% CI) ***
-Insured	26.4 (24.1–28.7)	40.4 (36.9–44.3)	8.4 (7.4–9.5)
-Uninsured	57.1 (56.2–58.0)	733.7 (682.7–787.1)	394.8 (368.0–423.1)

Note: *first part, **second part, ***combining both parts.

Abbreviation: OOP, Out-of-pocket.

Discussion

Overall, this study presents the outcome of public insurance arrangements in preventing upfront payment for psychiatric services among cross-border migrants in Thailand. The insured migrants are protected against OOP by 464 Baht on average, compared with the uninsured migrants. Based on our knowledge, this study is probably one of the first studies in Thailand or even in Southeast Asia that examines the financial benefit of public insurance on migrants suffering from psychiatric diseases.

Although few studies exist, most provide qualitative insights without quantifying the financial protection offered by insurance or focus on general insurance benefits without specific attention to psychiatric care. For instance, Tangcharoensathien et al¹⁵ highlighted that one of the key benefits of the HICS is the provision of financial risk protection for its beneficiaries. König et al¹⁶ conducted a study in Tak—a major city along Thailand-Myanmar border—and reported the advantage of a non-state private micro health insurance in enhancing health utilization among its migrant beneficiaries. Phaiyarom et al¹⁷ found that uninsured migrants tended to be younger and more likely to present with severe conditions compared with insured migrants. Moreover, insurance coverage and higher household socioeconomic status were associated with better access to non-communicable disease (NCD) services among asylum seekers in Bangkok.

There exist few studies that highlight the importance of expanding insurance benefits to cover migrants suffering from psychiatric illnesses. A systematic review by Satinsky et al¹⁸ demonstrated cost of care as one of the main deterrents for low mental care utilization among refugees and asylum seekers in European countries. McDaid and Park¹⁹ underpinned that an investment in mental health services by a host country would reduce financial burden and long-term societal cost for refugees and displaced people.

Interestingly, our study identifies a difference between the design of the insurance policy and its implementation. While, de jure, the insurance policy does not explicitly include psychiatric treatment for migrants, de facto, some local healthcare providers may provide psychiatric services as part of the insurance coverage at their discretion. A possible explanation is that most of the insurance revenue, such as payroll for the SSS and annual premiums for the HICS, is collected and managed by healthcare providers. On one hand, this arrangement allows flexibility of the providers to exercise their own discretion to adjust the insurance benefit coverage. On the other hand, it reflects a lack of policy awareness and monitoring in the field of neglected diseases like psychiatric illnesses, raising concerns about the sustainability of the policy. This presumption is supported by the fact that the HICS regulation on psychiatric services has never been changed since 2004.^5,6 At the time when the HICS was launched, it aimed at economic migrants, or in other words, those who are fit for work. With this logic, migrants with psychiatric illnesses are excluded from insurance enrollment. However, it appears that the policy is not aware of the fact that some migrants, though not presenting with psychiatric illnesses upon enrollment, may develop psychiatric diseases at any time during their stay in the host country. Besides, the service burden for migrants at the hospitals has never reduced as cross-border migrants have continued to be the main driver of the Thai economy during the last two decades.⁴

Incoherence between policy intentions and actions for migrants occur in several areas, not only psychiatric services. A clear example is the nationality verification (NV) which is the main legalization process for undocumented migrants. The idea is, under leniency policy, undocumented CLMV migrants are allowed to register with the Thai Government. Upon registration, they will be issued a work permit and enrolled in the insurance. When the NV is completed, a once-undocumented migrant will become fully documented. If those migrants wish to continue working in Thailand, they need to return to their home countries and come back again legally through the agreement between Thailand and the sending countries (which restrict quota of imported migrants, varying year by year). However, in practice, the NV process is very lengthy, and many migrants and Thai employers do not fully comply with the policy due to cost and convenience reasons, much less the poor enforcement by the responsible authorities.²⁰

Another prominent example illustrating the gap between policy intentions and actual practice is the provision of basic vaccination for undocumented migrant children. Although these children are officially required to pay for vaccination services, some health facilities provide them free of charge by utilizing surplus vaccines originally allocated for Thai children. Kosiyaporn et al²¹ reported that this practice occurs because vaccine uptake among undocumented migrant children is not high enough to exhaust the available surplus. This is further reinforced by the attitudes of local healthcare providers, who consider vaccination a fundamental health right regardless of immigration status.

With respect to policy implications, the findings suggest that formal inclusion of psychiatric services within the benefit packages of public insurance schemes for migrants may be beneficial, potentially improving the sustainability of psychiatric service provision and reducing reliance on discretionary decision-making by local healthcare providers. First, this inclusion is unlikely to impose substantial additional financial burden on healthcare facilities, as the beneficiaries are existing insurance members who have already contributed premiums at enrollment. Second, it does not appear to conflict with work permit regulations that restrict certain psychiatric conditions at the point of entry into the insurance scheme; rather, it may serve as a safety net by supporting migrant workers’ health while minimizing financial barriers to care. Third, it may provide a basis for extending healthcare benefits to underserved populations, including undocumented migrants, who often face barriers to accessing psychiatric services. This direction is broadly aligned with the “leave no one behind” principle of the Sustainable Development Goals (SDGs), to which Thailand has committed, as well as the vision of the National Health Security Office to promote equitable access to healthcare for all individuals residing in Thailand.²²

A major strength of this study is the use of real-world data from all hospitals affiliated with the MOPH nationwide. Another key strength is the application of advanced analytical approaches, including decision-tree algorithms such as the CatBoost classifier and machine-learning–based regression techniques like the NNLS ensemble, which effectively manage complex real-world data influenced by multiple determinants and demonstrate better model fit compared to conventional TPM.

However, some limitations remain. First, since we relied on hospital-based data, our findings do not capture migrants with psychiatric problems who did not seek care in hospitals. Therefore, the results should be interpreted with caution, as we lack information on overall healthcare utilization among all migrants. Moreover, migrant population is highly diverse. Some migrants have high socioeconomic status, such as those entering the country for medical tourism, while others live in deprived conditions. Second, due to the nature of cross-sectional study, causal inference cannot be established, although it is generally accepted that insurance entitlement influences OOP payment. Third, though the policy focus is on CLMV migrants, it is undeniable that there are other non-Thai nationals that utilize public hospitals. In general, these population account for a small number of services used compared to Thais and CLMV migrants. In several tourist provinces, there are notable numbers of non-Thai nationals who do not fall into the CLMV category. This group is quite diverse and includes both vulnerable individuals such as stateless people and refugees, as well as more prosperous residents like expatriates and retirees from Western countries. Fourth, the variables included in the model were limited to those routinely collected by hospitals. Consequently, some important factors influencing patients’ ability to pay and their likelihood of accessing healthcare were not captured, such as household socioeconomic status or financial support from migrants’ employers. Fifth, misclassification of some variables cannot be ruled out, especially for the insurance status. A patient whose work permit is expired may be still quoted as insured if the hospital providers do not seriously verify for his or her status and still rely on previously stored records. This situation also reflects room for improvement concerning data sharing between official authorities between the MOL which handles the employment record of migrants and the MOPH hospital database. Last, because the analytical model was carefully tailored to fit the data, there is a potential risk of overfitting. Therefore, the analysis should be updated whenever new data become available to maintain model validity and generalizability.

Additional studies that delve into the burden of psychiatric diseases amongst migrants beyond the health facility level is recommended. In this regard, a household survey on migrants’ health status is of huge value. Such an approach would facilitate the collection of additional variables not routinely captured in health facility data, including household wealth, duration of residence in Thailand, and levels of support from employers and family members. These data would also be useful for accurately estimating the psychiatric service burden on public health facilities, as well as the burden from other related illnesses, such as substance abuse and intoxication. Further research aimed at gaining a better understanding of how local providers make decisions regarding the provision of free psychiatric care to different groups of migrants would be of substantial value. The observed variation in discretionary practices may stem from multiple factors, including providers’ awareness of insurance regulations, their perceptions of the financial status of the facilities in which they work, and prevailing norms of care provision within their catchment areas. In addition, regular monitoring of migrant-related policy changes is essential, and any major shifts should prompt a re-examination of the data to inform future policy decisions.

Conclusion

Public insurance for cross-border migrants in Thailand contributed to a reduction in OOP for psychiatric illnesses at the point of service by an average of approximately 464 Baht, after accounting for both the probability of incurring payment and the conditional mean of amounts paid among payers. Clearer inclusion of psychiatric services within the benefit packages of public insurance schemes for migrants may help reduce inconsistent interpretations of service eligibility by local healthcare providers and support equitable access to care. However, the findings should be interpreted with caution, as the cross-sectional design limits causal inference and the absence of household-level data restricts assessment of the broader financial burden of psychiatric illnesses among migrants. Further studies exploring household-level burden are warranted, and regular updates to the data are important, particularly during periods of policy change, as significant shifts may prompt re-examination of these findings.

Funding Statement

This study is financially supported by the Health Systems Research Institute, Thailand.

Abbreviations

AUC-ROC, Area under the receiver operating characteristic curve; CI, Confidence interval; CLMV, Cambodia, Lao PDR, Myanmar and Vietnam; DALY, Disability-adjusted life year; GLM, Generalized linear model; HDC, Health Data Center; HICS, Health Insurance Card Scheme; ICD-10, International Classification of Diseases, 10^th Revision; MAE, Mean absolute error; MedAE, Median absolute error; MOL, Ministry of Labor; MOPH, Ministry of Public Health; NCD, Non-communicable disease; NNLS, Non-negative least square; NV, nationality verification; OLS, Ordinary least square; OOP, Out-of-pocket; PR-AUC, Area under the precision-recall curve; RMSE, Root mean squared error; SD, Standard deviation; SDG, Sustainable Development Goal; SSS, Social Security Scheme; TPM, Two-part model; UCS, Universal Coverage Scheme; WHO, World Health Organization.

Data Sharing Statement

Data are available from the corresponding author upon reasonable requests due to privacy and ethical restrictions.

Ethics Approval and Consent to Participate

The study received ethical approval from the Institute for the Development of Human Research Protections (IHRP), an official ethics review authority under the Health Systems Research Institute (HSRI), whose governing body is chaired by the Minister of Public Health, with approval letter number 110/2565. As the study involved secondary data analysis only, informed consent from individual participants was not required, as acknowledged by the IHRP. All personal data were encrypted during data retrieval and were not presented in any form that could allow the identification of individual participants. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare that there is no conflict of interest in this work.