Does vertical integration increase the costs for primary care inpatients? Evidence from a national pilot programme in China

Abstract

Objective

To assess the impact of vertical integration (VI) within County-Level Integrated Health Organisations (CIHOs) on the costs of primary care inpatients.

Methods

This study assessed Xishui, a national pilot county for CIHOs, using inpatient claims data. The treatment group comprised 10,118 inpatients from 5 vertically integrated township health centres (THCs), while the control group consisted of 21,165 inpatients from 19 non-vertically integrated THCs. The periods from July 2020 to December 2021 and January 2022 to December 2022 were defined as pre- and post-policy intervention, respectively. The primary outcome variables were total health expenditures (THS), out-of-pocket (OOP) expenditures, and the proportion of OOP expenditures. Propensity score matching was employed to align inpatient demographics and disease characteristics between the groups, followed by a difference-in-differences analysis to evaluate the outcomes.

Findings

VI significantly increased THS (β = 0.1337, p < 0.01) and OOP expenditures per case (β = 0.1661, p < 0.001), but the increase in the proportion of OOP expenditures per case was not significant (β = 0.0029, p > 0.05). For the basic medical insurance for urban and rural residents, THS per case (β = 0.1343, p < 0.01) and OOP expenditures (β = 0.1714, p < 0.001) significantly increased. For the basic medical insurance system for employees, THS per case also increased significantly (β = 0.1238, p < 0.01), but the change in OOP expenditure proportion per case was not significant (β = 0.1020, p > 0.05). The THS per case led by Xishui County People’s Hospital, the leading county medical sub-centre (CMSC), significantly increased (β = 0.1753, p < 0.01), whereas the increase led by Xishui County Traditional Chinese Medicine Hospital was not significant (β = 0.0742, p > 0.05). Increases in OOP expenditures per case were significant in CMSCs led by the People’s Hospital and the Traditional Chinese Medicine Hospital (β = 0.1782, p < 0.01 and β = 0.0757, p < 0.05, respectively).

Conclusion

VI significantly increased THS and OOP expenditures for primary care inpatients. However, VI could exacerbate economic disparities in disease burden across different insurance categories.

Text box 1. Contributions to the literature
• China is piloting a new vertically integrated model at the PHC, called the ‘County Medical Sub-centre’ (CMSC).
• Utilising inpatient claims data, our study explored the impact of the CMCS reform on the expenditures of primary care inpatients.
• VI significantly increased THS and OOP expenditures for primary care inpatients.
• VI could exacerbate economic disparities in disease burden across different insurance categories.

Introduction

Primary health care (PHC) is a crucial platform for efficient, effective and equitable universal health coverage (UHC) [1]. A robust PHC system strengthens the overall health system by enhancing its effectiveness, efficiency and equity [2]. The development of PHC has been prioritised in health service systems globally [3]. Integrated Delivery Systems (IDS) are utilised as a key strategy to strengthen PHC [4]. In 2016, the World Health Organisation, the World Bank and the Chinese government advocated for the establishment of a people-centred integrated health care system in China [5]. Following this recommendation, the Chinese central government issued its first national-level guidance on developing an integrated health care service system in April 2017.

China has designated county-level integrated health organisations (CIHOs) as the primary organisational form for IDS development [6]. These organisations aim to strengthen primary care by vertically integrating (VI) secondary services and PHC [7]. CIHOs encompass various integration aspects, including structural, functional, normative, interpersonal and process integration [8, 9]. In contrast to the United States, where IDS development is primarily driven by payment reform, China’s CIHOs are largely propelled by governmental administration, reflecting a more passive approach [10]. According to the National Health Commission, as of October 2023, China has more than 1800 pilot counties for CIHOs. Since the national pilot for CIHOs began in 2019, both positive and negative effects on capacity enhancement and cost control have been observed, primarily due to the limited service capabilities of county hospitals [11]. To address these inadequacies, the National Health Commission introduced a new organisational model, namely, the ‘county medical sub-centre (CMSC)’ in 2021 [12].

Nonetheless, limited evidence exists on the impact of CIHOs on the costs for primary care inpatients. Previous studies primarily focused on capacity enhancement, operational efficiency and development of a tiered healthcare system and reported mixed results [8]. Health security administrations are concerned that the establishment of CIHOs may lead to high costs for primary care inpatients. This concern is theoretically grounded because two competing theories, namely, strategic-based and efficiency-based theories, suggest that VI could increase or decrease medical costs for these patients [13,14,15,16]. Evidence from developed countries typically links VI with increased prices for medical services, heightened diagnostic and treatment intensity and higher overall medical expenditures, even at the primary care level [17,18,19,20].

Compared with the U.S., UK and Germany, China’s CIHOs show significant differences in integration goals, driving forces, organisational forms and member types. Medical service prices in China are regulated by the government, so the usual bargaining power dynamics may not fully explain the cost effects of VI [21, 22]. Given the government regulation of medical service prices, the impact of VI on the costs for primary care inpatients remains poorly understood in China. This paper aims to fill this knowledge gap by using medical claim data from a national pilot county and employing a natural experimental design. Our study assesses the effect of VI on the costs for primary care inpatients and contributes to the expanding literature on the cost effects of VI in low- and middle-income countries (LMICs).

Institutional background

Xishui County, located in Guizhou Province, southwestern China, administers 24 THCs. In 2023, the county reported a GDP of RMB 26.98 billion and a permanent population of 580,000. The development of Xishui’s CIHOs underwent two phases:

In the first phase, Xishui adopted a technical collaboration model to establish a loosely CIHO. The Xishui County People’s Hospital led 14 THCs, while the Xishui County Traditional Chinese Medicine Hospital led 10 THCs. However, due to the limited capabilities of county hospitals, the geographic dispersion of THCs, and weak connections among CIHO members, the loosely CIHO reform in Xishui did not achieve the intended policy goals.

In the second phase, Xishui adopted the vertically integrated CMSC model. The concept of CMSC was first proposed by the Xishui County Health Commission in September 2021 and was officially implemented in January 2022. Based on geographical location, service population and health resource distribution, Xishui was divided into five healthcare service areas (HSAs). In each HSA, one THC with stronger health service capabilities was designated as a CMSC. The Xishui County Health Commission regards the CMSCs as branches of the county hospital, which exercises unified management of personnel, finances, and resources over them, achieving full vertical integration without changing the nature of the THCs they belong to. The Xishui County People’s Hospital vertically integrated 3 THCs, while the Xishui County Traditional Chinese Medicine Hospital vertically integrated 2 THCs. Other THCs not included in the integrated management maintained a technical collaboration relationship with the county hospital under the framework of the loosely CIHO, without undergoing vertical integration. Additionally, after vertical integration, both the CMSCs and the non-vertically integrated THCs maintain the same price schedule and medical insurance reimbursement levels.

According to the monitoring and evaluation results from the National Health Commission, the CMSC reforms in Xishui have significantly achieved multiple policy goals, including strengthening primary care, promoting a tiered healthcare system and improving medical quality. In November 2022, Xishui was designated as one of the 12 national comprehensive pilot zones for primary health and wellness by the National Health Commission.

Data and methods

Data source

The data for this study were obtained from the health security administrations of Xishui County. These data include patient demographics (age, gender, insurance type); dates of medical service (admission and discharge dates); characteristics of the disease (type of treatment, main diagnosis, comorbidities, surgical intervention and length of hospital stay); institutional characteristics (institution name, code and level); and hospitalisation costs (total health expenditures (THS), out-of-pocket (OOP) expenditures and coinsurance). The data cover the period from January 1, 2020 to December 31, 2022. The study period was specifically defined from July 1, 2020 to December 31, 2022, to exclude the impact of COVID-19.

We focused on the top five diseases (acute bronchitis, lumbar disc herniation, arthritis, abdominal pain and bronchitis) treated in the THCs in Xishui. Cases were selected based on the ICD-10 codes: J20.900 for acute bronchitis, M51.202 for lumbar disc herniation, M13.900 for arthritis, R10.402 for abdominal pain and J40.x00 for bronchitis. The selection revealed that 10,118 inpatients from the 5 CMSCs and 21,165 inpatients from the 19 general THCs were included, accounting for 20.77% of all hospitalisations across the 24 THCs. This proportion indicates the representativeness of our sample. To ensure privacy, all identifiable patient information, such as names, ID numbers, addresses, and inpatient numbers, was removed before the data were made available to the research team.

The specific mapping process is illustrated in Fig. 1.

Fig. 1

Schematic diagram of the process of data screening in Xishui, July 2020 to December 2022. Notes N represents sample size, CMSC represents county medical sub-centre, THC represents township health centre, ICD-10 represents International Classifcation of Diseases-10th revision

Measures

Outcome variables

We defined THS, OOP expenditures and proportion of OOP expenditures per inpatient case as the outcome variables. Given the typically skewed distribution of medical expenditures, we normalised the data by using the natural logarithm of THS and OOP expenditures.

Explanatory variable

In this study, inpatients in the five CMSCs were considered the treatment group, while those in the remaining 19 THCs formed the control group. July 2020 to December 2021 and January 2022 to December 2022 were set as the pre- and post-policy intervention periods, respectively, to exclude the impact of the COVID-19 pandemic. A patient from the CMSC group discharged after January 1, 2022 was coded as 1; all other patients were coded as 0.

Control variables

Control variables included patient age, gender (coded as female = 0, male = 1), insurance classification (basic medical insurance for urban and rural residents (URRBMI) = 0, basic medical insurance system for employees (UEBMI) = 1), repeat hospitalisation (no = 0, yes = 1) and length of stay.

Statistical analysis

Difference-in-difference method (DID)

Our main empirical strategy employs a difference-in-differences(DID) analysis with fixed effects on hospitals and months, following the approach of Lai et al. [21]. The constructed regression model is as follows:

$$\eqalign{ Y{\rm{iht}} = & \alpha + \beta E{\rm{ventiht}} + \theta X{\rm{iht}} \cr & + \mu t + \nu h + \varepsilon {\rm{iht}} \cr}$$

(1)

where Event_iht is a dummy variable representing the state of the VI reform. Coefficient β captures the average treatment effect of the VI reform. The dependent variable Y_iht represents the THS, OOP expenditures and proportion of OOP expenditures for a specific hospitalisation case i at hospital h and time t. Control variables X_iht include patient demographics (age, gender, insurance type, repeat hospitalisation) and disease characteristics (main diagnosis, length of stay). µ_t represents a year-month dummy variable to control for flexible time effects, ν_h represents a series of hospital dummy variables to control for unobserved time-varying heterogeneity among hospitals and ε_iht represents a robust error term at the hospital-month level.

Propensity score matching method (PSM)

In China, primary care lacks a mandatory gatekeeper system, which gives patients the freedom to choose their hospitals. Additionally, the establishment of CMSCs was not based on randomised clinical trial interventions but rather on a non-randomised experimental approach, which did not adhere to the principles of randomness and homogeneity [23]. To address these issues and reduce potential sample selection biases between the control and treatment groups, we employed Propensity Score Matching (PSM). Specifically, we used two-nearest neighbour matching based on disease type for matching.

$$\:Logit({D}_{\text{i}}=1)=\alpha\:+\beta\:{X}_{i}+{\epsilon\:}_{i}$$

(2)

where D_i represents the dummy variable for the intervention of the vertical integration policy (it is coded as 1 for inpatients affiliated with the CMSCs and 0 for inpatients at other THCs). X_i includes matching variables, such as age, gender and insurance type, whether the patient has multiple hospital visits, the main diagnosis and length of stay.

Parallel trend test

A prerequisite for the double-difference approach is that the trends in the THS, OOP expenditures and proportion of OOP expenditures in the treatment and control groups were parallel before the VI reform. We use an event study methodology to test for parallel trends in the treatment and control groups following the approach of D’Haultfoeuille et al. [24].

$$L{\rm{og}}\,Yit\,\, = \alpha + \sum\limits_{{\rm{n}} = - 14}^{14} {\beta nDID_{it}^n + \gamma Xit + \mu i + \lambda i + \varepsilon i} t$$

(3)

DID is generated as a relative year policy variable with reference to the year of the VI reform, where CMCS variable equals 1, while the variables for THC are always 0. The month of the VI reform is set as the baseline time for the event study; β_n is the regression coefficient relative to the baseline month; X_it includes control variables. If the pre-reform coefficients β_n are not significantly different from 0, then the assumption of parallel trends is met, and the estimated results ofβ_n within the 95% confidence interval can be plotted.

Sensitivity analysis

We conducted robustness checks by using two methods. First, we replaced 2-nearest neighbour matching with alternative methods, such as Mahalanobis, radius and kernel matching, followed by DID estimation. Second, we changed the estimation method by employing double-machine learning models, specifically using random forest and Lasso algorithms for re-estimation.

All analyses were conducted using Stata 17.0 for Windows, with significance levels set at 0.1%, 1%, and 5%.

Results

Descriptive statistics

Table 1 presents the results of the descriptive statistical analysis. After implementing VI, the THS per inpatient case and OOP expenditures in the treatment group increased by RMB 181.15 and RMB 11.77, respectively, while the proportion of OOP expenditures decreased by 2.30%. In the control group, THS per inpatient case and OOP expenditures decreased by RMB 47.73 and RMB 5.98, respectively, with the proportion of OOP expenditures increasing by 0.08%. The control group had a higher average age, higher rates of repeat hospitalisation, and a larger proportion of female patients compared with the treatment group. Conversely, the treatment group had longer hospital stays and a higher proportion of patients with UEBMI.

Table 1 Descriptive statistics for treatment and control groups in Xishui, July 2020 to December 2022

Baseline regression

Table 2 reports the estimated effects of VI on the costs per inpatient case. Column 1 details the impact of VI on the THS per case, which significantly increased by 14.30% (= exp^(0.1337) − 1, p < 0.01). Column 2 presents the impact of VI on OOP expenditures per case, showing a significant increase of 18.07% (= exp^(0.1661) − 1, p < 0.001). Column 3 examines the impact on the proportion of OOP expenditures per case, which increased by 0.29%, although this increase was not statistically significant (p > 0.05).

Table 2 Impacts of VI on costs for primary care inpatients in Xishui, July 2020 to December 2022

Heterogeneity analysis

Different of insurance type

Table 3 reports the test results for the impact of VI on the costs per inpatient case across different insurance type. For URRBMI, VI led to a significant increase in THS per case by 14.37% (= exp^(0.1343) − 1, p < 0.01) and in OOP expenditures by 18.70% (= exp^(0.1714) − 1, p < 0.001), as shown in columns 1 and 3, respectively The proportion of OOP expenditures per case increased by 0.34%, which was not statistically significant (p > 0.05), as reported in column 5. For UEBMI, the THS per case increased by 13.18% (= exp^(0.1238) − 1, p < 0.01) and OOP expenditures by 10.74% (= exp^(0.1020) − 1), with the latter not reaching statistical significance (p > 0.05), as detailed in columns 2 and 4. The proportion of OOP expenditures per case decreased by 0.28%, with no significant change (p > 0.05), as indicated in column 6.

Table 3 Impacts of the VI on the costs for primary care inpatients by insurance type in Xishui

Different leading hospitals

Table 4 presents the test results for the impact of VI on costs per inpatient case across different CMSCs led by various hospitals. For the CMSC led by the Xishui County People’s Hospital, the THS per case increased significantly by 19.16% (= exp^(0.1753) − 1, p < 0.01), while the Traditional Chinese Medicine Hospital-led CMSC was smaller, with a non-significant increase of 7.70% (= exp^(0.0742) − 1, p > 0.05), as shown in columns 1 and 2. For OOP expenditures, there was a significant rise of 19.51% (= exp^(0.1782) − 1, p < 0.01) at the CMSC led by Xishui County People’s Hospital and a 7.86% increase (= exp^(0.0757) − 1, p < 0.05) at the Traditional Chinese Medicine Hospital, as detailed in columns 3 and 4. Columns 5 and 6 outline the changes in the proportion of OOP expenditures, which decreased by 0.42% at the former and increased by 0.19% at the latter, though neither change was statistically significant (p > 0.05).

Table 4 Impacts of the VI on the costs for primary care inpatients by leading hospitals in Xishui

Propensity score matching results

Balance test

After matching, we conducted a balance test to assess the effectiveness of the matching process. The null hypothesis for this balance test asserts that there are no systematic differences between the control and treatment groups. Table 5 indicate that the p-values for all covariates after matching were greater than 5%. Additionally, the absolute values of the standard deviations for the matching variables are all below 10%, suggesting that the null hypothesis cannot be rejected. This evidence supports the effectiveness of the PSM matching.

Table 5 Balance test of PSM for treatment and control groups in Xishui, July 2020 to December 2022

Kernel density plot

Kernel density plots are utilised to assess the quality of PSM. A larger overlap between the kernel density plots of the treatment and control groups indicates higher quality of the matching. As shown in Fig. 2, the kernel density distributions of the treatment and control groups nearly overlap after PSM, which indicates a high quality of matching.

Fig. 2

Kernel density plot of PSM for treatment and control groups in Xishui, July 2020 to December 2022

Parallel trend test

Figure 3 displays the event study results for the 12 periods before and after the treatment. Prior to the VI reform, no significant differences were observed in THS, OOP expenditures, and the proportion of OOP expenditures between the treatment group and control group. Following the VI reform, the effects of expenditures associated with VI began to gradually emerge. The findings confirm that, before the VI reform, the treatment and control groups met the criteria for the parallel trends test, validating the methodological approach.

Fig. 3

Parallel trends test of impacts of the VI on the costs for primary care inpatients in Xishui, January 2021 to December 2022. Notes OOP: out-of-pocket

Sensitivity analysis

To enhance the robustness of our conclusions, we employed various matching methods, including Mahalanobis, kernel, and caliper matching, as well as estimation methods using random forest and Lasso machine learning models. Table 6 presents the regression results obtained with these methods. The results closely align with the baseline regression findings, demonstrating the robustness of our study.

Table 6 Sensitivity checks of impacts of VI on costs for primary care inpatients in Xishui, July 2020 to December 2022

Discussion

This study is the first in China to utilise inpatient claims data and a quasi-natural experimental design to evaluate the impact of VI on the costs of primary care inpatients. The findings indicate that VI significantly increases the THS and OOP expenditures per case, although the change in the proportion of OOP expenditures per case is not significant. Potential mechanisms that might explain these cost increases under a profit maximisation objective include cost internalisation, budget gaming, capability internalisation and increased market power. However, the increases in THS and OOP expenditures are more pronounced for those insured under the URRBMI, who are economically and health-wise more disadvantaged, compared with those insured under the UEBMI.

Given that China has not yet established a mandatory gatekeeper system and medical insurance benefits are insufficient to promote tiered diagnosis and treatment, the share of primary care funding from medical insurance has been declining annually [25, 26]. Before VI, county hospitals and THCs operated with distinct utility functions, with county hospitals significantly surpassing THCs in terms of service capacity and reputation. Moreover, unlike integration motivated by efficiency and strategic reasons, the VI in the construction of CIHOs in China is predominantly driven by administrative directives. Theoretically, VI can affect medical prices, treatment intensity and healthcare costs through mechanisms such as bargaining power, double marginalisation, market concentration and vertical monopolies [14, 27,28,29,30,31].

Our estimates indicate that after VI, the THS per case increased significantly by 14.30% (p < 0.01), suggesting a substantial cost increase for primary care inpatients. This finding aligns with studies by Curto [32], Post [33], Ho [34] and Baker [35]. However, medical service pricing in China is government regulated, and CMSCs and other THCs adhere to consistent medical insurance payment prices and reimbursement levels. The differences in bargaining power and payment across locations are unlikely to be the primary explanatory mechanisms for changes in THS. In the Chinese context, the cost increase effects of VI on hospitalisation may relate to changes in treatment combinations driven by factors, such as cost internalisation, budget gaming and capability internalisation, under a profit maximisation objective. Firstly, in contrast to American Accountable Care Organisations (ACOs), CIHOs in Xishui County have not established a shared savings mechanism [36], and profit maximisation remains a dominant function. Secondly, during the VI process, county hospitals assume responsibility for human and equipment resources for CMSCs [37], incurring monitoring, coordination and cooperation costs [38] without corresponding compensation mechanisms, significantly internalising costs and increasing the financial risk for county hospitals [15]. This finding amplifies the community’s motivation to maximise revenue. Thirdly, budget gaming rules lead to the maximisation of medical insurance revenue during VI, with medical insurance still applying a single institution budget model to county-level health organisations. Strategic budget overspending occurs as each entity competes for a larger share of next year’s budget [39]. Fourthly, capability internalisation allows county hospitals managing CMSCs to bypass essential drug lists and enables the use of expensive drugs [8]. Hence, capitated payment system with shared savings should be established to promote payment for health outcomes and foster accountability for population health and costs within medical communities.

In the context of universal medical insurance coverage in China, reducing patients’ ’OOP expenditures is a primary focus of health insurance policies [40]. Our estimates reveal that VI has led to a significant increase of 18.07% in OOP expenditures per case (p < 0.001) and a marginal increase of 0.29% in the proportion of OOP expenditures, though the latter is not statistically significant (p > 0.05). These results indicate that VI significantly raises the cost burden on primary care inpatients. This finding contrasts with the conclusions of Su [41] and others who investigated the impact of VI under a capitated total amount prepaid system on OOP expenditures for inpatients among URRBMI. Hence, a capitated payment system with shared savings should be established.

Furthermore, our study explores the impact of VI on different insurance classifications. Previous research highlighted an unequal economic burden of disease between China’s URRBMI and the UEBMI, with URRBMI individuals more likely to incur catastrophic health expenditures [42, 43]. Our findings indicate that VI led to significant increases in THS per case for URRBMI and UEBMI, with a more pronounced increase observed for URRBMI. Although VI significantly raised the OOP expenditures per case for URRBMI, it did not have a significant effect on UEBMI. These results suggest that VI may further exacerbate economic disparities in disease burden across different insurance classifications.

In China, different levels of healthcare providers are still strongly motivated to compete for maximum profits [44]. The cost for a single hospitalisation at tertiary hospitals is significantly higher than those at secondary hospitals, and tertiary hospitals also handle a far greater volume of services [45]. Specifically, the market dominance of the Xishui County People’s Hospital, a tertiary hospital, is significantly stronger than that of the secondary-level Traditional Chinese Medicine Hospital. In the context of VI, such market dominance is often linked to increases in medical service prices and healthcare costs [46, 47]. Our findings show that after VI, the THS per case at the CMSC led by the Xishui County People’s Hospital increased significantly by 19.16%, while no significant increase was found at the CMSC led by the Traditional Chinese Medicine Hospital. Additionally, the increase in OOP expenditures per case at the CMSC led by the Xishui County People’s Hospital was significantly higher than that at the CMSC led by the Traditional Chinese Medicine Hospital. These results suggest that market forces significantly amplify the cost-increasing effects of VI in Xishui.

A limitation is that we only investigated the short-term effects of VI due to data availability. Future research should utilise long-term claims data to continuously monitor whether the cost-containment effects of VI persist over time.

Another potential limitation is that there may be spillover effects from the VI reform on control group. Although we used PSM to reduce bias between the treatment and control groups, the control group within the same county might also be affected by VI. Therefore, caution should be exercised when applying the findings of this study, particularly when considering the application of Xishui County’s VI experience to other counties in China. Future studies could use neighboring counties that have not implemented VI as a control group to more accurately estimate the cost-containment effects of VI.

In conclusion, within the regulatory environment of medical service pricing in China, VI significantly increased THS and OOP expenditures for primary care inpatients. These cost increases may be attributed to mechanisms, such as cost internalisation, budget gaming, capability internalisation and market power. However, VI could exacerbate the economic burden of diseases across different insurance classifications. In this regard, a capitated payment system with shared savings should be established. Policy orientation should also be strengthened towards payment for health outcomes and enhanced accountability for population health and costs within medical communities.