How Do Health System Features Influence Health System Performance?

4. Does a strong primary care system improve performance?

Abstract

This chapter investigates the relationship between primary care system characteristics and avoidable hospital admissions across OECD countries, focusing on admissions for asthma, chronic obstructive pulmonary disease, and congestive heart failure. Using cluster analysis, countries were grouped into five clusters based on three key primary care features: gatekeeping roles, continuity of care, and financial incentives for providers. The analysis revealed that health systems with strong gatekeeping, high continuity of care, and substantial financial incentives for quality demonstrated lower rates of avoidable hospital admissions compared to other systems. While the findings suggest that primary care oriented systems may help reduce acute deterioration in patients with chronic conditions, the large confidence intervals in the results warrant careful interpretation as the relationship may be influenced by unmeasured health system features or social determinants of health.

This chapter looked at whether differences in avoidable hospital admissions for selected conditions – asthma, chronic obstructive pulmonary disease, and congestive heart failure – can be linked to primary care oriented systems.

A high-performing primary care system, which provides accessible and high-quality services, can improve population health (Stange, Miller and Etz, 2023[1]) and reduce acute deterioration in people living with asthma, chronic obstructive pulmonary disease, or congestive heart failure, widely prevalent long-term conditions. The evidence base for effective treatment is well established, and hospital admissions for these conditions are largely avoidable and are therefore used as a marker of quality and access to primary care (OECD/European Commission, 2024[2]).

There is a long-standing debate on the features of high-quality primary care. When defining a “good” primary care system, there is a broad consensus that such system should be characterised by its accessibility, the provision of comprehensive care tailored to patient needs, with strong co‑ordination across all healthcare levels and continuity of care, particularly for individuals living with chronic conditions (OECD, 2020[3]). Primary care delivery should rely on robust relationships between the interdisciplinary primary care team, the patient, their family, and the community. This system must be adept at co‑ordinating activities across different levels of healthcare, serving as the primary point of contact, especially for patients with complex care needs. Across the board, there is agreement that a well-functioning primary healthcare system improves care co‑ordination and health outcomes and reduces wasteful spending, by limiting unnecessary hospitalisations and their associated costs (OECD, 2020[3]).

Considering the above, Ambulatory Care Sensitive Conditions (ACSC) are widely recognised as an indicator of access, quality and performance of the primary healthcare system (Agency for Healthcare Research and Quality, 2002[4]). ACSCs refer to acute or chronic health conditions that lead to avoidable hospitalisations if not effectively managed early in the outpatient primary care setting. Timely and adequate treatment for ACSCs, delivered at a primary care level is shown to reduce the need for hospital admissions, leading to better healthcare outcomes for patients.

Care and management of ACSCs is complex, often involving different care providers in different settings across the healthcare system, rendering the critical role of the primary care system as focal. Multiple studies have shown that continuity of care and good co‑ordination between healthcare professionals are strongly associated with lower rates of ACSC-related hospital admissions (Van Loenen et al., 2015[5]; Van Loenen et al., 2016[6]; Lyhne et al., 2022[7]). Research findings also indicate that hospital bed supply is an important factor in determining the number of ACSC-related hospital admissions. The availability of acute hospital beds in particular may influence admission decisions and lead to unnecessary ACSC-related hospital admissions, even when outpatient care could suffice (O’Cathain et al., 2013[8]; Van Loenen et al., 2016[6]).

In this set of analyses, the age‑sex standardised rate of potentially avoidable hospital admissions due to asthma, chronic obstructive pulmonary disease, and congestive heart failure per 100 000 population was used as output variable, based on the assumption that a lower rate can indicate higher quality of and better access to primary care (OECD, 2020[3]). The OECD also looked at potentially avoidable hospital admissions for two additional conditions – diabetes and hypertension. However, variations in coding practices for these conditions limit international comparisons. As a result, OECD discontinued the collection of hospitalisation data on hypertension, and further efforts are underway to improve the accuracy and international comparability of data related to hospital admissions for diabetes. Consequently, the analysis in this chapter focuses on chronic obstructive pulmonary disease, asthma and congestive heart failure.

While hospital-based measures provide valuable insights into system-level impacts, they do not fully capture the direct effects of primary care services on patient health and well-being. Future analyses could incorporate patient outcome variables, such as those defined by the OECD PaRIS (Patient-Reported Indicators Surveys) framework, to provide a more comprehensive evaluation.

Incentives for primary healthcare providers have demonstrated moderate effectiveness in reducing avoidable hospitalisations, particularly by enhancing chronic disease management and improving the quality of preventive care. Some studies suggest that these incentives can align provider efforts with desired care outcomes, leading to partial or positive impacts on care quality (Petersen et al., 2006[54]). However, the overall evidence remains mixed, as the effectiveness of such programmes varies significantly depending on their design, the scale of incentives, and the specific populations targeted. Despite this variability, incentive‑based approaches appear to encourage a shift toward more proactive and patient-centred care practices, highlighting their potential when thoughtfully implemented.

There is a range of ways to assess the performance of primary care at system level (World Health Organization and the United Nations Children’s Fund (UNICEF), 2022[9]; Kringos et al., 2013[10]; Gumas et al., 2024[11]). In this chapter, considering the key domains used by Barbara Starfield’s 4Cs framework (Macinko, Starfield and Shi, 2003[12]; Starfield, Shi and Macinko, 2005[13]) to assess the contribution of primary care to health outcomes, we aimed to identify indicators of primary care for which the data available through the OECD survey on Health System Characteristics cover a large part of OECD countries (Table 4.1): gatekeeping; population registered with a primary care provider and/or with regular doctor to go for care; financial incentives to primary care physicians to provide preventive care, to manage chronic disease and population risk factors, to uptake IT services and related to patient satisfaction. Those indicators are mainly related to the service‑delivery process (Kringos et al., 2013[14]).

Table 4.1. Description of indicators used for a cluster analysis on the strength of primary care

Domain	Indicator	Description
First contact	Role of primary care in the health system (gate‑keeping)	Show whether there is a requirement that primary care practitioners serve as gatekeepers to other levels of care. C: strong gatekeeping; B: limited gatekeeping; A: weak/no gatekeeping.
Continuity	Population registered with a primary healthcare provider and/or with a regular doctor to go for care	It captures the share of the population that has a regular doctor to go for care. A: limited part of the population; B: the majority of the population; C: almost the whole population.
Incentives	Financial incentives for primary care physicians	Indicates degree of financial incentives to primary care physicians to provide preventive care, to manage chronic disease and population risk factors, to uptake IT services and related to patient satisfaction. A greater number of positive responses indicate strong financial incentives. A higher score is then assigned.

First, a pure statistical data driven approach to clustering was conducted based on those three indicators (Box 4.1).

Box 4.1. Results of the data driven approach to clustering health systems based on financial incentives for quality of care and payment systems to providers

The data driven approach identified the following five clusters of health systems:

Figure 4.1. Clusters of health systems

At first sight, some of the clusters look plausible – most of them include countries which are neighbours of each other and/or which have aspects in common. However, there are interlopers that are difficult to explain (e.g. Czechia in cluster 1; Luxembourg in cluster 4).

Based on the data driven approach, some expert judgement was used to ensure that there are meaningful and identifiable policy differences to explain why health systems are grouped together. As a result, five clusters1 that display the following features were identified (Figure 4.2):

• Cluster 1: Australia, Estonia, Lithuania, the Netherlands and the United Kingdom report a strong gatekeeping role of primary care physicians, a high continuity of care and large financial incentives to primary care physicians.

• Cluster 2: Costa Rica, Finland, Norway and Slovenia report a strong gatekeeping role of primary care physicians and a high continuity of care. In contrast to cluster 1, financial incentives for quality to primary care physicians are generally weaker.

• Cluster 3: Canada, Chile, France, Latvia, Poland, Portugal, Spain and Sweden report either a strong gatekeeping and a limited care continuity or a limited gatekeeping and a strong care continuity as well as strong financial incentives for quality to primary care physicians.

• Cluster 4: Belgium, Germany, Ireland, Israel, Italy and Switzerland report either a strong gatekeeping and limited continuity of care or a limited gatekeeping and a strong continuity of care. In contrast to cluster 3, financial incentives for quality to primary care physicians are generally weaker.

• Cluster 5: Austria, Czechia, Greece, Hungary, Iceland, Japan, Korea and Luxembourg do not report a “gatekeeping” role of primary care physicians.

Figure 4.2. Groups of health systems with similar gate‑keeping function, continuity of care and financial incentives for primary care physicians to improve quality of care

Statistical methods

Similarly to the previous set of analyses (see Chapter 3), the model controls for several characteristics that influence health system performance in primary healthcare (Table 4.2). Control variables also accounted for the two features considered non-actionable – that is more complex to change – from a policy perspective: the overall type of coverage (residence‑based/single payer versus multiple insurers) (Paris, Devaux and Wei, 2010[15]) and the degree of decentralisation of spending autonomy in health (Dougherty and Phillips, 2019[16]). Finally, a dummy variable was used to capture the impact of COVID‑19.

Table 4.2. Control variables used in the panel regression model

Variable	Indicator	Reference
Wealth	GDP per capita (measured in Purchasing Power Parities)	(Macinko, Starfield and Shi, 2003[12])
Inequality	Gini index of household income distribution
Education	% of the population 25‑65 with tertiary education
Demographics	% of the population 65 years old or older	(Macinko, Starfield and Shi, 2003[12])
Risk factors	% of obese population Percentage of population 15+ daily smokers	(Macinko, Starfield and Shi, 2003[12])
Environmental hazard	Pollution (PM2.5 kg per km2)
Health system capacity	Hospitalisation rates (discharges per 100 000 population)/Hospital beds per 1 000 population Primary care physicians (General practitioners, gynaecologists, paediatricians) per 1 000 population.	(Macinko, Starfield and Shi, 2003[12])
Major disruptions to hospital admissions	A discrete variable to indicate the years of the COVID‑19 pandemic

Note: GNI – instead of GDP – was used for Luxembourg.

Key findings

In line with expectations, the panel regression analysis indicated that higher avoidable admission rates are reported in countries with higher inequality in income distribution, higher level of educational attainment, and in health systems with a higher number of hospital beds per 1 000 population. Avoidable hospital admissions rates were higher in the “multiple insurers and high decentralisation” group of health systems as compared to the “residence‑based/single payer and high decentralisation” group of health systems (Table 4.3).

Cluster 1 – which groups health systems with strong gatekeeping where most of the population regularly sees the same family doctor and large financial incentives for quality to primary care physicians – presents lower avoidable admission rates relative to the other clusters. This difference is statistically significant (Table 4.3).

Table 4.3. Variables with a statistical significant coefficient in the regression model

Variable	Estimate	Standard Error	P-value
Education	0.02	0.00	0.00***
Gini	3.29	1.36	0.02*
Obese	0.02	0.01	0.03*
Hospitalization rate	0.00	0.00	0.00***
Primary care physicians	‑0.16	0.08	0.05*
Post_Covid	‑0.18	0.05	0.00***
NonModChar3 (vs. NonModChar1)	0.72	0.18	0.00***
Cluster 2 (versus cluster 1)	0.52	0.11	0.00***
Cluster 3 (versus cluster 1)	0.40	0.12	0.00**
Cluster 4 (versus cluster 1)	0.57	0.10	0.00***
Cluster 5 (versus cluster 1)	0.54	0.11	0.00***

Note: Significant result at *0.05, **0.01, ***0.001 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered at country level. The outcome of the is log transformed. The model is also controlled for health behaviours, environmental hazards, rate of PCPs, COVID‑19 years and non-modifiable characteristics. The full model, together with the functional form, estimation method and assumption testing can be found in Annex 4.A.

Figure 4.3 shows the potential increase in performance for the other clusters of health systems should they adopt primary care oriented policies similar to the one reported for health systems in cluster 1. The large confidence intervals call for caution when interpreting those results.

Figure 4.3. Potential average decrease in avoidable hospital admissions by cluster

Note: 95% confidence intervals are displayed.

Sensitivity analyses

Sensitivity analyses were conducted to assess the robustness of results to using time fixed effects and to different output variables (see Annex 4.A). The direction and significance of coefficients of control variables remained robust across those specifications. However, the link between clusters and the output variable was not confirmed when using congestive heart failure admission rates alone, while only cluster 2 and 4 had a statistically significant effect when using asthma hospital admissions as the output variable.

Care continuity, strong gatekeeping, and large financial incentives for quality of care are related to lower rates of avoidable admissions

Although the effect of clusters on the outcome variable may be due to health system features that are not captured or controlled for in the analysis or to social determinants of health, results suggest that a primary care oriented system helps improve population health by reducing acute deterioration in people living with asthma, chronic obstructive pulmonary disease, or congestive heart failure. These are widely prevalent long-term conditions that may result in access to hospitals that should not occur in the presence of timely and effective healthcare.

Annex 4.A. Clustering and sensitivity analyses

Output variable

The output variable of this set of analyses is the age‑sex standardised rate of avoidable hospital admission rates for asthma, chronic obstructive pulmonary disease (COPD) and congestive heart failure per 100 000 population (Annex Figure 4.A.1)

Annex Figure 4.A.1. Avoidable hospital admission for asthma, chronic obstructive pulmonary disease and congestive heart failure by country, 2022 or latest available year

Source: OECD Health Statistics, January 2024.

Clustering

Three indicators were used to describe characteristics of primary care that could be influenced by actionable policy levers: gatekeeping; continuity of care; financial incentives for primary care physicians to improve quality of care (Annex Table 4.A.1).

Ward’s method was employed to group countries into clusters based on three indicators. Based on Silhouette scores and the visual inspection of the dendrogram, five clusters were identified to best represent groups of health systems (Annex Figure 4.A.2).

Annex Table 4.A.1. Score of indicators by cluster

Indicator	Average for numeric variables and dominance in categorical variables
	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Cluster 5
Continuity	C	Mostly C	Mostly B	Mostly C	A
Gatekeeping	C	C	Mostly C	Mostly B	A
Financial incentives to increase healthcare quality in primary healthcare	2.1	0	2.3	0	0.9

Note: Numeric variables a re‑scaled between 0 and 3, with 0 being the lowest score and 3 the highest. Role of primary care in the health system (gatekeeping): A: no requirement and no incentive for referral form PHC; B: financial incentives; C: referral form PHC is required. Continuity of care: A: limited part of the population; B: the majority of the population; C: almost the whole population.

Annex Figure 4.A.2. Dendogram

Five clusters were identified as containing elements that were similar among themselves and dissimilar to elements belonging to other groups (Annex Figure 4.A.3).

Annex Figure 4.A.3. Health systems by cluster. Data driven approach

Statistical approach

Pooled ordinary least squares regression

A pooled ordinary least squares regression was used to estimate the relationship between various socio‑economic, environmental, and health-related factors and age‑sex standardised avoidable hospital admissions (Equation 4). The dependent variables – rate of avoidable admissions – is operationalised as the logarithm to improve model fit. The explanatory variables include Gross Domestic Product (GDP) per capita, adjusted for purchasing power parity, which serves as a control for wealth of the country. The unemployment rate was also included in order to capture the economic activity. The Gini index was used to measure income inequality. Educational attainment was represented by the percentage of the population aged 25‑65 with tertiary education, which reflects the overall education level of the country. Health-related behaviours were captured through variables such as the percentage of the obese population and the percentage of daily smokers aged 15 and above. Environmental exposure was represented by the mean annual concentration of particulate matter (PM2.5 in mg per m^3) per squared kilometre.

In addition to the model variables outlined above, the model included care capacity indicators hospitalisation rates per 100 000 population (Hospitalisation rate) and Primary care practitioners (General practitioners, paediatricians and gynaecologist) per 1 000 population (PCP).

Additionally, Nonmodifiable health system characteristics (NonModChar) were included to account for aspects of the health system that are not subject to change in the short term. Finally, the model integrated the interest dummy variables (Clusters) representing the five clusters of countries. The analysis is conducted using the “plm” package in R.

Equation 4

${\mathrm{l}\mathrm{o}\mathrm{g}(Rate of avoidable admissions per 100 000 population}_{ it})=$

$\alpha +{\beta }_0{GDP}_{it}+{{\beta }_1 Education}_{it}+{\beta }_2{Beds}_{it}+{\beta }_3{Gini}_{it}+{\beta }_{4}Elderly+{\beta }_5{Obesity}_{it}+{\beta }_6{tobacco}_{it}+\mathrm{ }{\beta }_7{PCP}_{it}$$+ {{\beta }_{8}COVID}_{it}+{\beta }_9{NonModChar}_i+{\beta }_{10}{\mathit{C}\mathit{l}\mathit{u}\mathit{s}\mathit{t}\mathit{e}\mathit{r}\mathit{s}}_{\mathit{i}}{+\epsilon }_{it}$

With $i$ representing OECD countries and $t$ the year in the 2016‑22 period.

Assumption testing

Several tests were conducted to assess the underlying assumptions of the panel models. A Chow test was first conducted for poolability using the pooltest function (R package plm), which examines whether the panel data model benefits from individual effects beyond the pooled model. For both models, this test suggested potential individual effects across entities, significant at 5% error. Due to the static nature of the interest variables, they could not be adequately controlled by a fixed effects model at country level without losing these variables from the analysis, thus becoming one of the limitations of the approach. Nevertheless, the standard errors were clustered at the country level to account for potential country level fixed effects and not overestimate the significance of the results. Similarly, the significance was tested, and time fixed effects were included when appropriate by calculating F Test for Individual and/or Time Effects. In the final models of Chapter 3 and 4, time fixed effects were not significant when adding the COVID‑19 variable.

Additionally, Pesaran’s test of cross-sectional dependence was applied to investigate the independence of errors across panel units. For heteroskedasticity, a Breusch-Pagan test indicated the presence of heteroskedastic error terms in both models, which was addressed by computing robust standard errors using the Arellano method. The Breusch-Godfrey test was conducted for autocorrelation, which checks for serial correlation in the residuals, and assessed multicollinearity among independent variables using the Variance Inflation Factor.

The results of the main model are reported in Annex Table 4.A.2. The variables in the model were removed one by one, with the overall model conclusions remaining strong to the different specifications. The final model corresponds to the one with the highest R squared.

Annex Table 4.A.2. Panel regression results (robust estimates)

Variable	Estimate	SE	P-values
(Intercept)	2.61E+00	4.59E‑01	6.49E‑07
GDP	‑2.54E‑06	2.45E‑06	3.04E‑01
Education	1.91E‑02	4.73E‑03	1.85E‑04
Gini	3.29E+00	1.36E+00	1.88E‑02
Obese	2.12E‑02	9.19E‑03	2.53E‑02
tobacco	1.53E‑02	1.00E‑02	1.34E‑01
PM2.5_km	4.20E‑04	2.13E‑04	5.37E‑02
Hospitalization rate	6.16E‑05	1.17E‑05	2.92E‑06
PCP_per1000	‑1.59E‑01	7.83E‑02	4.76E‑02
Post_Covid	‑1.84E‑01	4.67E‑02	2.60E‑04
NonModChar2 (versus NonModChar1)	1.40E‑01	8.17E‑02	9.22E‑02
NonModChar3 (versus NonModChar1)	7.24E‑01	1.83E‑01	2.32E‑04
NonModChar4 (versus NonModChar1)	‑1.42E‑01	1.66E‑01	3.97E‑01
Cluster 2 (versus cluster 1)	5.18E‑01	1.06E‑01	1.07E‑05
Cluster 3 (versus cluster 1)	3.96E‑01	1.24E‑01	2.44E‑03
Cluster 4 (versus cluster 1)	5.71E‑01	1.05E‑01	1.50E‑06
Cluster 5 (versus cluster 1)	5.37E‑01	1.12E‑01	1.44E‑05
R2	0.83
R2adj	0.78
F-statistic	15.27

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

NonModChar1: Residence‑based or single payer + Relatively high spending power in health of subnational decision makers.

NonModChar2: Residence‑based or single payer + Relatively low spending power in health of subnational decision makers.

NonModChar3: Multiple insurers + Relatively high spending power in health of subnational decision makers.

NonModChar4: Multiple insurers + Relatively low spending power in health of subnational decision makers.

Sensitivity analyses

A series of sensitivity analyses were performed to test the significance of time fixed effects and the robustness of findings to different outcomes variables and time windows. Tested outcomes were the rates of avoidable admissions by condition (Asthma, COPD, Heart Failure, and a combination of Asthma and COPD), and substituting the hospitalisation rate with the rate of hospital beds per 1 000 population.

• Admission rates for Asthma: the model did not indicate the need for time‑fixed effects. Key control variables such as Gini, obesity, hospitalisation rate and smoking rates were positive and significant, while primary care physicians showed a negative correlation with asthma admission rates. Similar to the main model, cluster 4 and cluster 2 had a significant impact, showing a positive relation with admissions compared to cluster 1. On the contrary cluster 5 was associated with lower asthma admissions, but not statistically significant (Annex Table 4.A.3).

Annex Table 4.A.3. Panel regression results using admission rates for Asthma as output variable

Variable	Estimate	SE	P-values
(Intercept)	‑8.59E‑01	7.06E‑01	2.29E‑01
GDP	1.34E‑05	4.96E‑06	9.24E‑03
Education	‑2.09E‑02	6.87E‑03	3.67E‑03
Gini	6.13E+00	2.20E+00	7.52E‑03
Obese	4.61E‑02	1.50E‑02	3.41E‑03
tobacco	5.03E‑02	1.26E‑02	2.02E‑04
PM2.5_km	8.42E‑04	3.29E‑04	1.35E‑02
Hospitalization rate	7.64E‑05	1.39E‑05	1.24E‑06
PCP_per1000	‑1.99E‑01	8.75E‑02	2.74E‑02
Post_Covid	‑2.10E‑01	5.55E‑02	4.15E‑04
NonModChar2 (versus NonModChar1)	‑3.10E‑01	1.23E‑01	1.50E‑02
NonModChar3 (versus NonModChar1)	4.46E‑01	3.03E‑01	1.48E‑01
NonModChar4 (versus NonModChar1)	‑7.84E‑01	2.44E‑01	2.32E‑03
Cluster 2 (versus cluster 1)	5.99E‑01	1.51E‑01	2.27E‑04
Cluster 3 (versus cluster 1)	1.26E‑01	1.96E‑01	5.24E‑01
Cluster 4 (versus cluster 1)	5.48E‑01	1.78E‑01	3.31E‑03
Cluster 5 (versus cluster 1)	‑2.52E‑01	1.98E‑01	2.09E‑01
R2	0.92
R2adj	0.89
F-statistic	35.35

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

• Admission rates for COPD: the model revealed significant time‑fixed effects. Control variables displayed similar directional behaviour to the main model, but only Gini were significant. The direction and significance of the cluster variables remained robust (Annex Table 4.A.4).

Annex Table 4.A.4. Panel regression results using admission rates for COPD as output variable

Variable	Estimate	SE	P-values
(Intercept)	1.20E‑05	4.30E‑06	7.87E‑03
GDP	1.79E‑02	6.29E‑03	6.46E‑03
Education	1.00E+01	2.21E+00	4.27E‑05
Gini	7.53E‑03	1.78E‑02	6.74E‑01
Obese	‑1.07E‑02	1.34E‑02	4.29E‑01
tobacco	2.94E‑04	2.55E‑04	2.55E‑01
PM2.5_km	1.13E‑05	1.78E‑05	5.28E‑01
Hospitalization rate	‑1.45E‑01	1.02E‑01	1.65E‑01
PCP_per1000	9.82E‑02	1.58E‑01	5.38E‑01
NonModChar2 (versus NonModChar1)	7.47E‑01	3.09E‑01	1.97E‑02
NonModChar3 (versus NonModChar1)	8.68E‑02	4.08E‑01	8.33E‑01
NonModChar4 (versus NonModChar1)	8.41E‑01	1.64E‑01	5.71E‑06
Cluster 2 (versus cluster 1)	5.62E‑01	1.88E‑01	4.45E‑03
Cluster 3 (versus cluster 1)	9.67E‑01	2.32E‑01	1.35E‑04
Cluster 4 (versus cluster 1)	8.80E‑01	2.73E‑01	2.30E‑03
Cluster 5 (versus cluster 1)	1.20E‑05	4.30E‑06	7.87E‑03
R2	0.82
R2adj	0.74
F-statistic	14.40

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

• Admission rates for Asthma and COPD combined: the model did not require time‑fixed effects. While the direction of control variables remained consistent with the main model, their significance shifted, with only Gini and Education remaining significant, while smoking became significant and hospitalisation rate became insignificant. Clusters showed significant effects on admissions with the same direction as the main model (Annex Table 4.A.5).

Annex Table 4.A.5. Panel regression results using admission rates for Asthma and COPD as output variable

Variable	Estimate	SE	P-values
(Intercept)	1.01E‑05	4.22E‑06	2.40E+00
GDP	3.29E‑02	6.93E‑03	4.75E+00
Education	1.04E+01	2.39E+00	4.34E+00
Gini	1.69E‑02	1.57E‑02	1.08E+00
Obese	3.56E‑02	1.65E‑02	2.15E+00
tobacco	7.32E‑04	3.04E‑04	2.41E+00
PM2.5_km	‑1.90E‑05	1.74E‑05	‑1.09E+00
Hospitalization rate	‑1.43E‑01	8.47E‑02	‑1.69E+00
PCP_per1000	1.99E‑01	1.32E‑01	1.51E+00
Post_Covid	6.30E‑01	2.60E‑01	2.42E+00
NonModChar2 (versus NonModChar1)	3.41E‑01	3.61E‑01	9.45E‑01
NonModChar3 (versus NonModChar1)	1.17E+00	1.55E‑01	7.56E+00
NonModChar4 (versus NonModChar1)	3.37E‑01	1.65E‑01	2.04E+00
Cluster 2 (versus cluster 1)	7.63E‑01	2.34E‑01	3.27E+00
Cluster 3 (versus cluster 1)	8.43E‑01	2.87E‑01	2.93E+00
Cluster 4 (versus cluster 1)	1.01E‑05	4.22E‑06	2.40E+00
Cluster 5 (versus cluster 1)	3.29E‑02	6.93E‑03	4.75E+00
R2	0.83
R2adj	0.76
F-statistic	15.37

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

• Admission rates for Heart Failure: the model did not require time‑fixed effects. Moreover, the direction of controls variables remained generally stable compared to the main model. However, only hospitalisation rates and Education were found to be significant were found to be significant (associated with higher avoidable admissions). None of the clusters of characteristics related to the strength of primary care were found significant (Annex Table 4.A.6).

Annex Table 4.A.6. Panel regression results using admission rates for Congestive Heart Failure as output variable

Variable	Estimate	SE	P-values
(Intercept)	4.29E+00	1.02E+00	1.11E‑04
GDP	‑1.14E‑05	7.86E‑06	1.53E‑01
Education	1.69E‑02	7.42E‑03	2.74E‑02
Gini	‑3.93E+00	2.47E+00	1.18E‑01
Obese	9.01E‑03	1.70E‑02	5.99E‑01
tobacco	1.41E‑02	1.87E‑02	4.56E‑01
PM2.5_km	1.85E‑04	3.92E‑04	6.39E‑01
Hospitalization rate	1.11E‑04	2.51E‑05	5.58E‑05
PCP_per1000	‑5.50E‑02	1.19E‑01	6.45E‑01
Post_Covid	1.00E‑01	8.54E‑02	2.47E‑01
NonModChar2 (versus NonModChar1)	3.44E‑02	1.74E‑01	8.45E‑01
NonModChar3 (versus NonModChar1)	5.36E‑01	4.69E‑01	2.59E‑01
NonModChar4 (versus NonModChar1)	‑4.15E‑01	3.34E‑01	2.20E‑01
Cluster 2 (versus cluster 1)	‑1.34E‑02	2.22E‑01	9.52E‑01
Cluster 3 (versus cluster 1)	2.92E‑01	3.12E‑01	3.55E‑01
Cluster 4 (versus cluster 1)	1.10E‑01	4.09E‑01	7.89E‑01
Cluster 5 (versus cluster 1)	2.34E‑01	3.34E‑01	4.87E‑01
R2	0.71
R2adj	0.62
F-statistic	7.64

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

• Using the rate of hospital beds: the direction and significance of the interest variables in the model remains stable compared to the base model that used hospitalisation rates (Annex Table 4.A.7).

Annex Table 4.A.7. Panel regression results using the rate of hospital beds instead of the hospitalisation rate

Variable	Estimate	SE	P-values
(Intercept)	3.68E+00	3.50E‑01	2.09E‑15
GDP	2.28E‑06	2.05E‑06	2.71E‑01
Education	1.08E‑03	3.91E‑03	7.83E‑01
Gini	3.64E+00	1.02E+00	7.04E‑04
Obese	2.48E‑02	7.99E‑03	2.90E‑03
tobacco	‑1.20E‑02	6.78E‑03	8.19E‑02
PM2.5_km	‑1.74E‑04	2.07E‑04	4.05E‑01
Hospitalization rate	2.78E‑01	5.45E‑02	3.56E‑06
PCP_per1000	‑2.80E‑01	7.91E‑02	7.73E‑04
Post_Covid	‑2.55E‑01	4.50E‑02	4.18E‑07
NonModChar2 (versus NonModChar1)	‑1.21E‑01	1.03E‑01	2.44E‑01
NonModChar3 (versus NonModChar1)	5.93E‑01	1.71E‑01	9.79E‑04
NonModChar4 (versus NonModChar1)	‑6.14E‑01	2.29E‑01	9.37E‑03
Cluster 2 (versus cluster 1)	5.45E‑01	9.92E‑02	7.88E‑07
Cluster 3 (versus cluster 1)	3.81E‑01	1.02E‑01	3.89E‑04
Cluster 4 (versus cluster 1)	8.81E‑01	1.15E‑01	1.77E‑10
Cluster 5 (versus cluster 1)	2.63E‑01	1.21E‑01	3.30E‑02
R2	8.06E‑01
R2adj	7.56E‑01
F-statistic	1.61E+01

Note: Significant result at *0.05, **0.001, ***0.000 level. The model uses the Arellano method for heteroskedasticity-consistent standard errors (White) clustered in time (year) and group (Country). The outcome of the model is log transformed.

References

[4] Agency for Healthcare Research and Quality (2002), Guide to Prevention Quality Indicators: Hospital Admission for Ambulatory Care Sensitive Conditions.

[16] Dougherty, S. and L. Phillips (2019), “The spending power of sub-national decision makers across five policy sectors”, OECD Working Papers on Fiscal Federalism, No. 25, OECD Publishing, Paris, https://doi.org/10.1787/8955021f-en.

[11] Gumas, E. et al. (2024), Finger on the Pulse: The State of Primary Care in the U.S. and Nine Other Countries, https://doi.org/10.26099/p3y4-5g38.

[14] Kringos, D. et al. (2013), “The strength of primary care in Europe: an international comparative study”, British Journal of General Practice, Vol. 63/616, pp. e742-e750, https://doi.org/10.3399/bjgp13x674422.

[10] Kringos, D. et al. (2013), “Europe’s strong primary care systems are linked to better population health but also to higher health spending”, Health Affairs, Vol. 32/4, pp. 86-694, https://doi.org/10.1377/hlthaff.2012.1242.

[7] Lyhne, C. et al. (2022), “Interventions to Prevent Potentially Avoidable Hospitalizations: A Mixed Methods Systematic Review”, Frontiers in Public Health, Vol. 10, https://doi.org/10.3389/fpubh.2022.898359.

[12] Macinko, J., B. Starfield and L. Shi (2003), “The Contribution of Primary Care Systems to Health Outcomes within Organization for Economic Cooperation and Development (OECD) Countries, 1970–1998”, Health Services Research, Vol. 38/3, pp. 831-865, https://doi.org/10.1111/1475-6773.00149.

[8] O’Cathain, A. et al. (2013), “Hospital characteristics affecting potentially avoidable emergency admissions: National ecological study”, Health Services Management Research, Vol. 26/4, pp. 110-118, https://doi.org/10.1177/0951484814525357.

[3] OECD (2020), Realising the Potential of Primary Health Care, OECD Health Policy Studies, OECD Publishing, Paris, https://doi.org/10.1787/a92adee4-en.

[2] OECD/European Commission (2024), Health at a Glance: Europe 2024: State of Health in the EU Cycle, OECD Publishing, Paris, https://doi.org/10.1787/b3704e14-en.

[15] Paris, V., M. Devaux and L. Wei (2010), “Health Systems Institutional Characteristics: A Survey of 29 OECD Countries”, OECD Health Working Papers, No. 50, OECD Publishing, Paris, https://doi.org/10.1787/5kmfxfq9qbnr-en.

[1] Stange, K., W. Miller and R. Etz (2023), “The Role of Primary Care in Improving Population Health”, The Milbank Quarterly, Vol. 101/S1, pp. 795-840.

[13] Starfield, B., L. Shi and J. Macinko (2005), “Contribution of Primary Care to Health Systems and Health”, The Milbank Quarterly, Vol. 83/3, pp. 457-502, https://doi.org/10.1111/j.1468-0009.2005.00409.x.

[6] Van Loenen et al. (2016), “The impact of primary care organization on avoidable hospital admissions for diabetes in 23 countries”, https://doi.org/10.3109/02813432.2015.1132883.

[5] Van Loenen, T. et al. (2015), “Organizational aspects of primary care related to avoidable hospitalization: a systematic review”, https://doi.org/10.1093/fampra/cmu053.

[9] World Health Organization and the United Nations Children’s Fund (UNICEF) (2022), Primary health care measurement framework and indicators: monitoring health systems through a primary health care lens, https://iris.who.int/handle/10665/352205.

Note

← 1.  Colombia, Hungary, Mexico, New Zealand, the Slovak Republic and Türkiye are not included in this analysis as there are missing values for the indicators used to cluster health systems.