An ounce of prevention for a pound of cure: Basic health care and efficiency in health systems

Bancalari, Antonella; Bernal, Pedro; Celhay, Pablo A.; Martinez, Sebastian; Sánchez, Maria Deni Working Paper An ounce of prevention for a pound of cure: Basic health care and efficiency in health systems IDB Working Paper Series, No. IDB-WP-1231 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Bancalari, Antonella; Bernal, Pedro; Celhay, Pablo A.; Martinez, Sebastian; Sánchez, Maria Deni (2024) : An ounce of prevention for a pound of cure: Basic health care and efficiency in health systems, IDB Working Paper Series, No. IDB-WP-1231, Inter-American Development Bank (IDB), Washington, DC, https://doi.org/10.18235/0005669 This Version is available at: https://hdl.handle.net/10419/299428 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/3.0/igo/ A n ounce of prevention for a pound of cure: Basic health care and efficiency in health systems A ntonella Bancalari Pedro Bernal Pablo Celhay Sebastian Martinez Maria Deni Sánchez WORKING PAPER No IDB-WP-1231 InterA merican Development Bank Social Protection and Health Division January 2024 A n ounce of prevention for a pound of cure: Basic health care and efficiency in health systems A ntonella Bancalari Pedro Bernal Pablo Celhay Sebastian Martinez Maria Deni Sánchez InterA merican Development Bank Social Protection and Health Division January 2024 Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library An ounce of prevention for a pound of cure: basic health care and efficiency in health systems / Antonella Bancalari, Pedro Bernal, Pablo Celhay, Sebastián Martínez, Maria Deni Sánchez. p. cm. — (IDB Working Paper Series ; 1231) Includes bibliographic references. 1. Community health services-El Salvador. 2. Preventive health services-El Salvador. 3. Primary health care-El Salvador. I. Bancalari, Antonella. II. Bernal, Pedro. III. Celhay, Pablo. IV. Martínez, Sebastián. V. Sánchez, Maria Deni. VI. Inter-American Development Bank. Social Protection and Health Division. VII. Series. IDB-WP-1231 http://www.iadb.org Copyright © 2024 Inter-American Development Bank ("IDB"). This work is subject to a Creative Commons license CC BY 3.0 IGO (https://creativecommons.org/licenses/by/3.0/igo/legalcode). The terms and conditions indicated in the URL link must be met and the respective recognition must be granted to the IDB. Further to section 8 of the above license, any mediation relating to disputes arising under such license shall be conducted in accordance with the WIPO Mediation Rules. Any dispute related to the use of the works of the IDB that cannot be settled amicably shall be submitted to arbitration pursuant to the United Nations Commission on International Trade Law (UNCITRAL) rules. The use of the IDB's name for any purpose other than for attribution, and the use of IDB's logo shall be subject to a separate written license agreement between the IDB and the user and is not authorized as part of this license. Note that the URL link includes terms and conditions that are an integral part of this license. The opinions expressed in this work are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. [email protected] www.iadb.org/SocialProtection An ounce of prevention for a pound of cure: Basic health care and efficiency in health systems * Antonella Bancalari, Pedro Bernal, Pablo Celhay, Sebastian Martinez and Maria Deni Sánchez Abstract We examine the efficiency gains in health systems generated after the national roll out of basic healthcare in El Salvador between 2010 and 2013. Using data from over 120 million consultations and five million hospitalizations, we demonstrate that the expansion of community health teams, comprising less-specialized health workers, increases preventive care and decreases curative care and preventable hospitalizations. We also estimate coverage improvements for previously unattended chronic conditions amenable to effective primary care. These results suggest that decentralization of tasks to less-specialized health workers improves efficiency, maintaining quality of care. JEL: I15, I18, H21, H51. Keywords: community-based healthcare, efficiency, coverage. *Bancalari: Institute for Fiscal Studies and IZA Institute of Labor Economics (e-mail: an- [email protected]); Bernal: Inter-American Development Bank (e-mail: [email protected]); Celhay (corresponding author): Escuela de Gobierno and Instituto de Economía, Pontificia Universidad Católica de Chile (e-mail: [email protected]); Martinez: 3ie (e-mail: [email protected]); Sanchez: Inter-American Development Bank (e-mail:[email protected]). We thank the Ministry of Health in El Salvador and the Salud Mesoamerica Initiative for assistance and funding for this study. Bancalari gratefully acknowledges financial support from the RSE–Fulbright Visitor Scholar Fellowship, and Celhay from ANID, FONDECYT Regular 1221461, ANID, and PIA/PUENTE AFB220003. We are grateful to María Fernanda García Agudelo and Matias Muñoz for outstanding research assistance, and participants at the LACEA-LAMES Annual Meeting for useful comments. All opinions in this paper are those of the authors and do not necessarily represent the views of the Government of El Salvador, or the Inter-American Development Bank, its Executive Directors, or the governments they represent. 1 Introduction The substantial growth in healthcare spending in recent decades has brought the efficiency of health systems into the spotlight (Hall and Jones,2007;Garber and Skinner,2008;Christopoulos and Eleftheriou,2020). A notable source of inefficiency is the underutilization of primary care, leading to an overreliance on hospital services for illnesses that could be more effectively prevented or managed through primary care (Dafny and Gruber,2005;Garber and Skinner,2008; Alexander et al.,2019;Pinchbeck,2019). In lowand middle-income countries (LMICs), hospital care accounts for approximately 60% of government healthcare expenditure (Pinto et al., 2018). In developed countries, wasteful spending in health has become a significant source of inefficiency (OECD,2017). To this end, the reorganization and/or expansion of basic healthcare coverage as an alternative have garnered significant interest (Einav and Finkelstein,2023). We study efficiency gains from a supply-side expansion of basic healthcare through a nationwide reform in El Salvador, which established Community Health Teams (CHTs). These teams, comprising physicians, nurses, and community health workers, offer a range of preventive health services, including outpatient consultations, home visits, and community outreach activities. While it is well understood that community-based healthcare can improve health outcomes, less is known about how it can improve efficiency in healthcare provision in low-income contexts. We first document the effects of the reform on the available inputs for healthcare delivery (i.e., health units and healthcare workers) and the supply of preventive care services. To study efficiency, we then estimate the effects of CHTs on the production of two types of services: (i) curative consultations for conditions that can be effectively managed, treated, or prevented through primary care, referred to as amenable curative consultations; and (ii) hospitalizations for conditions that can be effectively managed, treated, or prevented through primary healthcare and, if not properly addressed, could lead to unnecessary hospitalizations or complications, referred to as preventable hospitalizations. A key and novel aspect of our study involves categorizing outpatient care production into preventive and curative components. In our context, most outpatient care is provided in a primary care setting, with roughly 95% of preventive consultations and 80% of curative consultations happening at this level. As CHTs complement traditional primary care unit services with home visits and community outreach, understanding how the production of outpatient care services changes with the promotion of CHTs provides a more comprehensive assessment of the economic value of basic care coverage. Our dataset enables this analysis, marking our study as one of the pioneers in exploring heterogeneity in the production of outpatient care based on the type of care provided. Additionally, we can precisely differentiate between curative care for communicable diseases 2 (CDs) and non-communicable diseases (NCDs) in outpatient visits by utilizing ICD-10 diagnosis codes. This precision is valuable as basic healthcare has the potential to prevent CDs and expedite the timely diagnosis and treatment of NCDs, thereby reducing the likelihood of complications, costly hospitalizations, and adverse health outcomes. We employ an event-study strategy exploiting the staggered roll-out of the CHT system across municipalities between 2010 and 2013.1We use a panel dataset of 254 municipalities, combining various sources of detailed, high-quality administrative records and census data. We construct this dataset aggregating over 120 million individual consultation records that took place between 2009 and 2018, and almost five million inpatient hospital records spanning 2005–2018, allowing us to measure preventive and curative consultations and hospitalizations per 1,000 inhabitants for each municipality and year. The duration of our data enables us to analyze dynamics for a period of up to 5 years following the implementation of the reform. We use the method proposed by Borusyak et al. (2024) to explore dynamics of treatment effects when policies are rolled out at different periods in different areas. Briefly, the method defines groups of municipalities according to the period that they were treated and estimates counterfactuals for each treated group using imputation procedures and relying on not-yet treated and never-treated units as controls at each point of time. Our findings show that CHTs improve efficiency in health systems through a task-shifting model, wherein certain tasks are appropriately delegated to skilled yet less-specialized health workers, such as nurses instead of doctors.2CHTs not only expanded preventive care but also reduced amenable curative consultations for CDs, while enhancing management for NCDs, ultimately decreasing preventable hospitalizations. Specifically, we find that CHTs increased preventive consultations by 36.9% compared to initial levels. The greater supply of preventive healthcare services and outreach efforts were sufficient to overcome demand-side barriers characteristic of LMICs (Dupas,2011). The effect of CHTs on curative healthcare is a priori ambiguous. On one hand, increased preventive care could lead to fewer amenable curative consultations, thereby saving resources, as the latter tend to be more expensive (efficiency effect). On the other hand, there could be an increase in curative visits, especially in the short term, if there is a high demand for healthcare for unattended conditions (coverage effect) (Hennessy,2008;Glazer and McGuire,2012). Our data allows to disentangle both effects by dividing curative outpatient consultations according to type. We find that CHTs decreased amenable curative consultations for CDs by 8.5%, while increasing coverage for amenable curative consultations for NCDs by 17.9%, compared to initial levels. 1Municipalities are the lowest jurisdictional level in El Salvador. An average municipality had 22,000 inhabitants by the 2007 Census. 2This is in line with the World Health Organization’s definition of ‘task-shifting’ (Campbell and Scott,2011). 3 Such improvements in case management within primary care units led to a 10.8% reduction in preventable hospitalizations following the creation of the CHTs. These reductions were primarily driven by a 13.6% drop in preventable hospitalizations for diagnosis related to CDs requiring inpatient treatment, compared to initial levels. We provide evidence that the rollout of the CHT system across municipalities is uncorrelated with potential confounding factors that would invalidate our empirical strategy to identify treatment effects. Although the Ministry of Health prioritized implementing CHTs in poorer municipalities, the actual start of their activities depended on the speed at which teams registered the population, which could have been slower in targeted poorer municipalities. We demonstrate that the results remain robust when controlling for initial municipal poverty by year fixed effects and also present evidence supporting the parallel trends assumption. Furthermore, we show that the results withstand several sensitivity checks and the use of alternative two-way fixed effects estimators. Finally, while we lack data to study causal effects on health outcomes, we find that the expansion of the CHT system is associated with a reduction in the mortality rate for CDs that can be effectively managed, treated, or prevented through outpatient care, referred to amenable mortality. We contribute to three main literature streams. First, this paper contributes to studies on the effectiveness of basic healthcare provided through a model of community health. This model, with its longstanding history as a supply-side alternative for delivering primary care, has recently seen a resurgence in LMICs (see, for example, Angwenyi et al. (2018), Kok et al. (2015), Mor et al. (2023)). While much of the existing literature focuses on estimating health effects, our study aims to understand the mechanisms behind these effects by examining efficiency in healthcare delivery within a national health system. One of the pioneering works in this domain is by Goldman and Grossman (1988), who demonstrated that community health centers in the USA are associated with a reduction in infant mortality rates.3Since then, a substantial body of literature, focusing primarily on maternal and child health, has emerged. For instance, Kose et al. (2022) in the USA, Das et al. (2013) reviewed studies in Asia and Africa, Brazil’s Family Health Program studied by Macinko et al. (2006), Aquino et al. (2009), Rocha and Soares (2010), and Herrera-Almanza and Rosales-Rueda (2023) in Madagascar, to name a few. Several other studies have examined how communitybased healthcare improves reproductive healthcare. For instance, Arends-Kuenning (2001), Barham (2012) and Joshi and Schultz (2013) in Bangladesh, Salehi-Isfahani et al. (2010) in Iran, and Herrera-Almanza and Rosales-Rueda (2020) in Madagascar. Our study contributes to this literature stream by providing evidence that one mechanism for 3Further insights into the history and evolution of community health programs can be found in Singh and Sachs (2013) and Perry et al. (2014). 4 the success of CHTs as an alternative in organizing primary healthcare is a task-shifting model. In this model, certain tasks are delegated to skilled yet less-specialized health workers without compromising quality. Our evidence is relevant not only to LMICs, which have been implementing such primary care models, but also to developed countries contemplating strategies to address operational waste in healthcare spending, aiming to ensure patients receive similar benefits of care while utilizing fewer expensive resources (Bentley et al.,2008;OECD,2017; Shrank et al.,2019). Second, our study complements the literature stream on health system efficiency. Previous work, primarily from advanced economies, has shown that expanding primary care improves overall health system efficiency by reducing emergency room use and hospitalizations for avoidable NCDs (e.g., Dafny and Gruber,2005;Kolstad and Kowalski,2012;Miller,2012;Dolton and Pathania,2016;Whittaker et al.,2016;Alexander et al.,2019;Pinchbeck,2019;Ding et al., 2021;Gruber et al.,2022). Notably, in Brazil, Macinko et al. (2010) found that a major expansion of primary care decreased unnecessary hospitalizations, and Bhalotra et al. (2020) observed that urgent care centers reduced hospital outpatient procedures and admissions, and that this is associated with improved hospital performance, indicated by a decline in inpatient mortality. While efficiency gains from avoiding emergency hospitalizations are well-suited for high-income and upper-middle-income countries, in lower-income contexts, the efficiency margins within the production of primary care services remain large. Hospitals or emergency rooms in LMICs are inaccessible to a portion of the population, as their infrastructure is often located in urban centers (Thornton,2008;Kremer and Glennerster,2011;Adhvaryu and Nyshadham,2015). In lower-income settings like ours, both preventive and outpatient curative care predominantly occur at the primary care level and share limited resources. Curative care is typically more resource-intensive than preventive consultations, requiring medication and significant medical staff hours. Moreover, curative consultations are usually unscheduled, disrupting physicians’ daily schedules in health units (Hey and Patel,1983;Courbage and Rey,2006;Williams et al., 2006;Nuscheler and Roeder,2016;Wang,2018;Peter,2021). Health centers often operate under tight capacity constraints, making the opportunity costs of using resources for otherwise preventable conditions significant. We contribute to this literature by focusing on efficiency gains in a low-income context, examining how CHTs can shift resources within outpatient healthcare that mainly takes place at the primary level. A significant advantage of our data is that it enables us to analyze the utilization of care for conditions that are amenable to effective primary care. By separately focusing on care for CDs, which are easily preventable, and NCDs, which are more resource-intensive, we identify two types of gains from the CHT system: one where more preventive care shifts resources towards less expensive curative care (efficiency effect), and another that reallocates resources from curative care for CDs to curative care for NCDs, thereby increasing the capacity 5 Here the set of 1[Kjt =h]are the lead and lag treatment indicator variables tracking the number of years Kjt =t−Ejsince the year of the CHTs creation for the municipality, Ej, a ≥0 and b ≥ 0 are the numbers of included leads and lags of the event indicator, respectively, and µjt is the error term. bis chosen such that all possible lags in the sample are covered. This specification also includes yearly pre-trends coefficients, i.e. a= 3. Absent pre-trends, the coefficients on the lags are interpreted as the dynamic path of causal effects: at h= 0, ..., b years after the creation of CHTs. τhcaptures treatment effect dynamics with respect to length of exposure to the treatment, i.e. the creation of CHTs. For each timing group treated at period k, never-treated, not-yet-treated, and already-treated serve as the control group. It has been well documented that traditional two-way fixed effects (TWFE) estimators, leveraging staggered roll-out, are subject to ‘negative weights’ because they use already-treated units act as the control group, and treatment effects may vary over time (Goodman-Bacon,2021). To address this concern, we use the imputation estimator proposed by Borusyak et al. (2024). Briefly, Borusyak et al. (2024)’s method defines groups of municipalities according to the period that they were treated and estimates counterfactual outcomes for each treated group. Potential control outcomes Yjt(0) are derived from municipalities that were never treated (27% of total municipalities), and those that were treated later on in each year. The counterfactuals are estimated using imputation procedures at each point of time, which are robust and efficient under heteroskedasticity. When calculating group-specific average treatment effects by time, we end up with many treatment effect parameters in a “fully dynamic” specification. For ease of interpretation, we take the mean over all point estimates using a linear combination, as suggested by Cunningham (2021). 4.2 Internal validity A Cox hazard model reveals that the timing of the creation of CHTs was unrelated to initial demographic characteristics of the municipality, the initial availability of inputs for healthcare production, as well as the initial level of healthcare services (Table B2, Column (2)). Although the MoH implemented CHTs giving priority to poorer municipalities (as discussed in Section 2), the actual start of CHTs activities (e.g. registering families in the CHT system) was likely slower in poorer municipalities. Thus, the net effect of poverty on the timing of the start of CHTs’ activities in null. In line with the MoH mandate, the initial percentage of the population living in poverty was higher in the treatment group than in the never-treated group of municipalities (see Table B1 in the appendix). This difference persists even when controlling for initial levels of other municipal characteristics, healthcare inputs, and outcomes (see Table B2, Column (1)). Treated districts 12 also had a larger share of rural population and were generally smaller in terms of population size, resulting in them having greater healthcare inputs per 1,000 inhabitants (see Table B1 in the appendix). However, these initial imbalances are not problematic, as the differences in levels are effectively controlled for by municipality fixed effects. Furthermore, as demonstrated in Section 5.5, the results remain robust to controlling for these initial municipality characteristics interacted with year dummies. Additionally, we show in Section 5that the parallel trends assumption in outcomes hold. An advantage of the dynamic event study is that it allows to visually assess the pattern of treatment effects relative to the creation of CHTs. In our main results we present up to three-year pretrends. While we can test for the significance of t−1for all the treated municipalities, and of t−2for 58% of the treated municipalities, a limitation is that we can only test for the significance of t−3for 6% of the treated municipalities when using the consultations data. Municipalities that implemented CHTs post-2012 only have sufficient data from 2009 onwards for the pre-trends analysis. To overcome this limitation, we additionally test for five-period pre-trends using data at the half-year level. Furthermore, exploiting the availability of hospital records from 2005 we are able to test for up to five-period pre-trends for all treated municipalities. We find no evidence of pre-trends using these alternatives, as discussed in Section 5.5. 5 The Effects of Community Health Teams 5.1 Changes in Inputs for Healthcare We start by evaluating how the reform in El Salvador affected the availability of inputs for the production of healthcare in municipalities. Using three rounds of data, 2009, 2010 and 2015, we estimate a static DiD model using Equation 1. The reform improved access to primary care services by increasing the number of primary care units and human resources, in particular nurses and support workers. Table 1, Panel A, shows that on average primary care units increased by 0.06 units per 1,000 inhabitants in treated municipalities after the creation of CHTs, equivalent to a 4.1% increase from the 2009 mean of never-treated municipalities. Furthermore, the total number of health staff in municipalities increased by 0.21 per 1,000 inhabitants on average (10.7%). This overall increase is mostly driven by an increase in the number of nurses and support workers (0.07 per 1,000 inhabitants, 18.4% and 17.5% respectively compared to the 2009 mean). Although also positive, the effects on the number of doctors and CHWs are not precisely estimated. Panel B shows how the reform expanded primary care services provided by larger multi-disciplinary teams, rather than relying on physicians alone. The composition of human resources changed, with an increase by 1.0 and 2.0 percentage points (ppts) in the share of nurses and support 13 workers, respectively, and a decrease by 2.0 ppts in the share of CHWs out of the total human resources, on average. The reform expanded the inputs used in the production of healthcare and changed the composition of healthcare workers in municipalities. The findings suggest that CHTs were based on a model that focused on skilled but less-specialized health workers and support personnel, which could help lower costs while improving health outcomes. 5.2 Expansion of Preventive Care As explained in Section 2, CHTs kick-started their activities with outreach efforts when registering individuals, coupled with proactive follow-up in scheduling preventive appointments. As such, we next evaluate the effect of CHTs on the number of preventive consultations. Before delving into the analysis, we test for the presence of pre-trends. Figure 2, Panel A, shows that the pre-trend coefficients are close to zero and are not statistically significant within conventional levels. Table 2showing the linear combination of all the coefficients estimated for the years prior to the creation of the CHTs confirms that the effect in the pre-treatment years is insignificant (column (1)). The creation of CHTs dramatically increased preventive healthcare in municipalities. The impact becomes significant after a one-year period, consistent with the timeline during which CHTs conducted household visits as part of their establishment. These household visits may have functioned as a substitute to preventive consultations at health centers. The effect jumps from 37 to 170 additional consultations per 1,000 inhabitants between t+ 1 and t+ 2, and it peaks in t+ 5 at 277 consultations per 1,000 inhabitants. The effect remains high even eight years after the creation of CHTs. Table 3summarizes the dynamic effects in a single coefficient capturing the average treatment effect for every year after the creation of the CHTs, based on Equation 2. Column (1) Panel A shows that, on average, the creation of CHTs increased preventive consultations by 187.6 per 1,000 inhabitants. This effect is equivalent to a 36.9% increase in preventive consultations with respect to the pre-treatment mean, and it is significant at the 1% level. We additionally estimate the static DiD effects following Equation 1. We find that the creation of CHTs increased by 72.3 preventive consultations per 1,000 inhabitants during the post-reform period (see Table 4). The lower magnitude in the coefficients of the static estimation compared with those from the imputation estimator is consistent with ‘negative weights’ introduced in traditional TWFE models, as discussed in Section 4. 14 5.3 Efficiency and Coverage Gains in Curative Care We now investigate the effect of the creation of CHTs on the number of amenable curative consultations. As explained in Section 3, these include visits to restore health due to conditions for which effective management and treatment can be achieved in a primary care setting, potentially avoiding the need for specialized or tertiary care. The absence of statistically significant pre-trends in Figure 2, Panel B, and Figure 3, Panels A and B, bolster our confidence to interpret the imputation estimations as causal effects of the arrival of CHTs on amenable curative care. We confirm this in Table 2(column (2)), where the average effect in years prior to the creation of CHTs is insignificant. We estimate no significant effects on amenable curative consultations (see Figure 2, Panel B). The average dynamic effect for up to eight years after the creation of CHTs shows a statistically insignificant decrease in amenable curative consultations (see Table 3, Panel A, column 2). Figure B1 in the appendix also shows an insignificant effect on total curative consultations (Panel A). We next explore whether the null effect on amenable curative consultations is the result of a coverage effect offsetting an efficiency effect. To do this, we classify amenable curative consultations as due to either CDs or NCDs. Focusing on amenable curative consultations due to CDs allows us to investigate potential efficiency gains. The initial outreach efforts by CHTs, along with their proactive follow-up in scheduling preventive appointments (as evidenced in Section 5.2), might have prevented the spread of infections and other CDs. CDs are often more easily preventable compared to NCDs, over which healthcare providers have less control. NCDs typically require lifestyle modifications and long-term management strategies, aspects that largely fall under the patient’s control. Panel A in Figure 3reveal that the creation of CHTs decreased curative consultations for CDs. The effect is immediate, a drop by 36 consultations per 1,000 inhabitants, consistent with the initial outreach activities of CHWs helping prevent the need for curative care for CDs in health units. The negative effects strengthen over time, following with the increase in preventive consultations in health units (as estimated in Section 5.2). The magnitude of this negative effect increased to 79 consultations in the fourth year and to 91 consultations after eight years (Figure 3, Panel A). On average, the creation of CHTs decreased amenable curative consultations due to CDs by 73.1 per 1,000 inhabitants, equivalent to a 8.5% drop with respect to the pretreatment mean (Table 3, column (2)). The effect is significant at the 1% level. This result serves as evidence of an efficiency effect from greater preventive care. Next, focusing on amenable curative consultations due to NCDs allows us to investigate potential coverage gains from CHTs for three reasons. First, there was a greater need for CDs, evidenced by lower coverage of curative care for these diseases prior to the creation of the 15 CHTs (855 CDs vs. 248 NCDs curative consultations per 1,000 inhabitants). Second, outreach activities by CHWs generated referrals to health units to treat illnesses and chronic conditions. Third, it is more resource-intensive to identify and follow-up on chronic conditions, like diabetes and asthma (Williams et al.,2006;Wang,2018), and hence lower amenable curative consultations due to preventable CDs might have released resources that could be allocated for NCDs treatment. In the year CHTs were created, curative consultations due to NCDs increased by 14 per 1,000 inhabitants, by 35 consultations in the fourth year and to 101 consultations after eight years (Figure 3, Panel B). The creation of CHTs increased curative consultations due to NCDs, on average, by 44.5 per 1,000 inhabitants (17.9%; Table 3, column (2)). This effect is also significant at the 1% level. This result serves as evidence of a coverage effect due to more resources available to manage previously unattended chronic conditions. We additionally estimate the static effect following Equation 1. We show in Table 4that the effect on amenable curative consultations is -25.0 per 1,000 inhabitants with a p-value of 0.16, and when split by disease type, the effect is -31.4 per 1,000 inhabitants for CDs and 6.3 per 1,000 inhabitants, though the latter is not significant. Overall, these findings suggest that the absolute gain in efficiency was greater in magnitude than the gain in coverage. 5.4 Efficiency Gains in Hospitalizations Did the expansion of community-based healthcare translate into efficiency gains in the system? To answer this question, we focus on preventable hospitalizations. As explained in Section 3, these include conditions that can be effectively managed, treated, or prevented in a primary care or outpatient setting, and those that if not appropriately addressed, could lead to unnecessary hospitalizations or complications. Figure 2Panel C confirms the absence of significant pre-trend estimates. Column (3) in Table 2 also confirms that the average effects on preventable hospitalizations were statistically insignificant before the creation of CHTs. Because we have data on hospitalizations since 2005, we also present estimates for up to five-year pre-trends in Section 5.5. It is important to note that, while the negative point estimate of t−1could be concerning (mostly for preventable hospitalizations for CDs), it gets closer to zero and even positive when imputing 5-year pre-trends and when using half-year data (see Figures C5 to C8 in the appendix). When omitting this first lead for normalizations, alternative TWFE estimators clearly show no pre-trends between t−5and t−2and a significant drop after t(see Figures C4 and C5 in the appendix). The creation of CHTs decreased the number of preventable hospitalizations. The effect is an immediate drop by 0.72 hospitalizations per 1,000 inhabitants, which peaks three years later at -0.9 and again eight years later at -1.5 hospitalizations. Figure B1 in the appendix shows that 16 the creation of CHTs had no effect on total hospitalizations (Panel C). The average dynamic effect after the creation of the CHTs is presented in Table 3(Panel A, column 3). We find that CHTs decreased preventable hospitalizations by 0.8 per 1,000 inhabitants –a drop equivalent to 10.8% with respect to the pre-treatment mean and statistically significant at the 1% level. Admissions were reduced due to preventable conditions as extreme cases were avoided through better case management through outpatient care. We evaluate efficiency as done for curative care. Figure 3, Panels C and D, reveal that after the introduction of CHTs preventable hospitalizations dropped for CDs and NCDs, though the effects on the latter are only precisely estimated in tand t+ 8. In the year CHTs were created, preventable hospitalizations for CDs dropped by 0.4 per 1,000 inhabitants and for NCDs they dropped by 0.3 per 1,000 inhabitants. The magnitude of the negative effect on preventable hospitalizations due to CDs increased to 0.7 hospitalizations four years later and to 0.9 hospitalizations after eight years (Panel C). The average dynamic effect for preventable hospitalizations by CDs is -0.6, equivalent to a drop by 13.6% with respect to the pre-treatment mean. This effect is statistically significant at the 1% level. The average effect on preventable hospitalizations for NCDs is -0.3, though it is not statistically significant at conventional levels (see Table 3, column 3). Consistent with the large increase in preventive care and curative care for amenable NCDs, more of these cases seem to have been resolved through outpatient care rather than requiring hospitalization. We additionally estimate the static effect following Equation 1, presented in Table 4. The effect on preventable hospitalizations is -0.8 per 1,000 inhabitants with a p-value below 0.01. When split by disease type, the effect is -0.5 per 1,000 inhabitants for CDs and -0.3 per 1,000 inhabitants (12.3% and 6.7% compared to the pre-treatment mean, respectively). All estimated effects are significant at conventional levels. As a placebo test, we additionally estimate the dynamic effect of the creation of CHTs on hospitalizations caused by external factors, such as injury, poisoning, accidents, assaults and selfharm. Community healthcare should not affect admissions by these unforeseen conditions that require specialized care. In line, we find no statistically significant effect on hospitalizations due to external causes (see Figure 2Panel D). 5.5 Robustness Checks Sensitivity Checks. All our estimated effects on preventive and curative consultations and hospitalizations are robust to sensitivity checks in which we set the year of treatment as the one in which a CHTs registered 10%, 15% and 20% of the municipality’s population (see Table C1 in the appendix). Throughout the different specifications, the estimates effects remain highly 17 significant. The magnitude of the effects on consultations, if anything, increases slightly when the creation year is set when a higher percentage of the population was registered, suggesting that our main estimates are conservative. Robustness for Effects on Inputs. To test for pre-trends before the introduction of CHTs, we conduct a placebo test using data from the years 2009 and 2010. We drop municipalities treated in 2010 (T2010) and we estimate a static DiD with an indicator variable that equals to one in 2010 for municipalities treated after 2010. Before the creation of CHTs, we find no significant difference in inputs for healthcare production, neither in counts nor in shares, across later-treated (after 2010) and never-treated municipalities (see Table C2 in the appendix). We replicate the estimations in Table 1for the sample of municipalities included in the placebo test, and we find that the results remain robust and even slightly higher in magnitude (see Table C3 in the appendix). The latter table alleviates concerns that a different composition of municipalities drives the null effects in the placebo test, as we exclude 41% of the treated municipalities in this test. Alternative estimators. We compare the results obtained with the imputation estimator of Borusyak et al. (2024) to the alternative estimators of De Chaisemartin and d’Haultfoeuille (2020) (DCHF), Sun and Abraham (2021) (SA), and Callaway and Sant’Anna (2021) (CS) that are also robust to treatment cohort heterogeneity (see Figures C1,C2 and C3 in the appendix). The results validate the main findings based on the imputation estimator, as the point estimates of the effects of CHTs are very similar. These estimations alleviate further the concerns related to ‘negative weights’ in traditional OLS estimations and bias introduced by the different composition of treatment cohorts. Additional pre-trends tests. Due to the availability of more rounds of hospital discharge records before the creation of CHTs (from 2005 onwards), we are able to test for pre-trends in these outcomes for additional periods for all treated municipalities. Figures C4 and C5 in the appendix show that there are no significant pre-trends when estimating up to five-year pre-trends and when using the alternative estimators of DCHF, SA and CS. For preventable hospitalizations by NCDs, the imputation estimator yields significant positive coefficients for t−5,t−3and t−2. Yet, these pre-trend coefficients are in the in the opposite direction of the treatment effects and the DCHF, SA and CS estimators yield pre-trend coefficients that are close to zero and insignificant for this outcome. To further study the pre-treatment patterns, we also utilize the availability of healthcare utilization data at the semi-annual level. The advantage of using semi-annual data is that it provides more variation in the rollout of the CHTs and allows us to test for longer pre-trends. When analyzing semi-annual data, we assume that the treatment begins in the semester before 5% of the municipality’s population is registered in the CHT system. This is because outcomes are now measured over a shorter time span, and the initial activities of the CHTs can already be reflected in these outcomes. With this conservative approach, 41.9% of treated districts were 18 treated in the first half of 2010, 43.0% in the second half of 2010, 9.2% in the first half of 2011, 1.6% in the second half of 2011, 2.7% in the first half of 2012, and 1.6% in the second half of 2012 (Figure C6 in the appendix shows the hazard plots for the event ”Creation of CHTs” comparing half-yearly and yearly data). We can now test for up to two-year pre-trends in the first cohort (41.9%) and for up to three-year pre-trends in the second cohort (43.0%), which together constitute the majority of the treatment group. The negative side of using data at the semi-annual level is that the data is more noisy, mostly for rare events like hospitalizations. Figures C7 and C8 in the appendix replicate Figures 2and 3, respectively, using half-year data. We are able to rule out up to five-year pre-trends for all outcomes. We only find a significant positive coefficient in t−1and t−4for curative consultations due to CDs, but this imbalance is in the opposite direction of the treatment effect. The absence of pre-trends is confirmed in Table C4 showing insignificant average pre-trend coefficients for all outcomes when using the semi-annual data. Moreover, the average dynamic treatment effects remain in the same direction and, as expected, are half the magnitude when using semi-annual data compared to annual data (see Table C5 in the appendix). Notably, in Figure C7 we observe a drop in t+ 1 for preventive consultations, exactly in the period when CHTs were more intensively visiting households to register them in the CHTs system, crowding-out preventive visits in health units. Alternative definition of preventable hospitalizations. We conduct robustness checks by employing alternative classifications of preventable hospitalizations. Figure C9 in the appendix shows that the estimated effects on hospitalizations remain robust (and are slightly larger in magnitude during the first years) when using only Kruk et al. (2018)’s list of conditions amenable to effective primary care, the same classification used for amenable curative consultations, as well as when using only Rodriguez Abrego (2012)’s list of ambulatory-care sensitive conditions (ACSC). Adding controls. We address concerns regarding initial municipality characteristics influencing the trends in our outcomes of interest. Considering that the MoH prioritized poorer municipalities, and that we observe this in the data when comparing treatment and control municipalities, we control for initial poverty interacted with year dummies. We measure initial poverty as the share of the population in a municipality that fall below the poverty line, which was measured in the 2007 census. Figure C10 in the appendix shows that the results remain robust. Notably, some point-wise confidence intervals become narrower, particularly for the effects on preventable hospitalizations due to NCDs. Our results also remain robust to controlling for initial population and the share of rural population interacted with year dummies. Other health shocks. We address concerns about other health policies and health shocks, such as epidemics, climate change, gangs and restrictions to mobility, affecting differently the regions where CHTs were deployed. Firstly, no other health reform was introduced in the targeted regions. Secondly, there is no specific geographical difference across treated and non-treated municipalities. Figure 1shows that CHTs were deployed throughout the national territory without 19 a specific spatial pattern. Thirdly, Figure C1, Panel (D), alleviates concerns about other health shocks due to violence being correlated with CHTs creation, as there are no significant effects on hospitalizations due to external causes. Finally, if other health shocks in treated municipalities, correlated with the timing of the creation of CHTs, were driving the results, we would expect to see effects in total curative consultations and total hospitalizations. However, Figure C1 in the appendix demonstrates that this is not the case. 5.6 Amenable Mortality Finally, we anticipate that CHTs have improved health outcomes due to the increase in preventive care and the expanded coverage in curative care. To explore this, we utilize data on mortality rates available from 2011 to 2018 and estimate a static DiD model in accordance with Equation 1. We use data on cause of death to compute mortality rates for conditions that are amenable to effective management and treatment that can be achieved in a primary care setting, potentially avoiding the need for specialized or tertiary care, following Kruk et al. (2018)’s classification. We further split amenable mortality rates by CDs and NCDs at the municipality level. Consistent with our previous results, Table D1 in the appendix shows that mortality caused by CDs amenable to healthcare decreased by 10.7 deaths per 1,000 inhabitants after the creation of CHTs (Panel A). The point estimate is equivalent to a drop of 24.7% with respect to the 2011 mean of never-treated municipalities and statistically significant at the 5% level. The estimated association with mortality caused by NCDs amenable to healthcare and with mortality not amenable to healthcare is also negative, but not statistically significant. These latter results are encouraging as CHTs are expected to decrease mortality caused by diseases amenable to effective primary healthcare, and that are easy to prevent. As only 6% of the treated municipalities implemented CHTs after 2011, we are unable to take advantage of the staggered treatment roll-out and test for pre-trends in mortality rates. Hence, we interpret the resulting coefficients with caution. As additional evidence, we compare amenable mortality rates between later-treated (after 2011) and never-treated municipalities in 2011, dropping T2010 and T2011 municipalities. In 2011, before being treated, we find no significant difference in mortality rates across later-treated and never-treated municipalities (Panel B). Additionally, we estimate a DiD static model dropping T2010 and T2011 municipalities in the sample, and we find that the results remain robust. Mortality caused by CDs amenable to healthcare decreased by 12 deaths per 1,000 inhabitants after the creation of CHTs, while there is no significant effect on mortality rates caused by NCDs and diseases and complications that are not amenable to primary care (Panel C). 20 5.7 Cost-effectiveness of the Reform In this section, we discuss the cost-effectiveness of introducing CHTs, for which we undertake some back-of-the-envelope calculations. We focus on monetizing preventable hospitalization gains because the overall effect on amenable curative care is zero, as coverage and efficiency gains offset each other. For this, we use data on hospitalization costs from the MoH of El Salvador and reports from the CHTs implementation. We first calculate how much an average municipality saved from the reduction in preventable hospitalizations per year. Using the coefficient of the effect on these hospitalizations of -0.8 per 1,000 inhabitants and per year per municipality (Table 3, Panel A, column 3), and considering the cost per hospitalization of USD 772.70, we estimate a saving per year and municipality equivalent to USD 615,841.90.7 Next we identify how costly are CHTs for an average municipality per year. The cost of running a CHT per year is USD 45,654.47. Using the post-treatment population forecasts and guidelines described in Section 2, the median number of CHTs in a treated municipality with an entirely rural population is 3.1, and in a municipality with an entirely urban population is 1.0. Hence, the total cost to run CHTs ranges between USD 45,654.5 and USD 141,528.9 per year per municipality. This calculation suggests that the introduction of CHTs in Salvador was highly cost-effective. Per USD 1 invested in CHTs, El Salvador saved roughly between USD 4.4 and USD 13.5 in expenditures for preventable hospitalization. This calculation is a lower bound, as we are not monetizing gains in the health status of the population (as suggested by Table D1 in the appendix). 6 Conclusions Our study investigates the efficiency gains resulting from a supply-side expansion of primary care through a nationwide reform in El Salvador, which introduced Community Health Teams (CHTs). These multidisciplinary teams, comprising physicians, nurses, and community health workers, offer a variety of preventive health services, including outpatient consultations, home visits, and community outreach. While the benefits of community-based healthcare on health outcomes are acknowledged, less is known about its impact on efficiency in healthcare provision, particularly in low-income contexts. Our empirical strategy leverages the staggered rollout of the CHT system across municipalities between 2010 and 2013, constructing a comprehensive dataset covering 254 munici7The estimate of cost per hospitalization in El Salvador is obtained from (Ministry of Health of El Salvador,2015). 21 Whittaker, W., L. Anselmi, S. R. Kristensen, Y.-S. Lau, S. Bailey, P. Bower, K. Checkland, R. Elvey, K. Rothwell, J. Stokes, et al. (2016). Associations between extending access to primary care and emergency department visits: a difference-in-differences analysis. PLoS medicine 13(9), e1002113. Williams, A., A. Lloyd, L. Watson, and K. Rabe (2006). Cost of scheduled and unscheduled asthma management in seven European Union countries. European Respiratory Review 15(98), 4–9. 28 Figure 1: Spatial Distribution of CHTs’ creation in El Salvador Never treated 2010 2011 2012 2013 Notes: This map shows the date in which CHTs were created, proxied by the year in which municipalities registered at least 5% of its population. 29 Table 1: Inputs for Healthcare Production (1) (2) (3) (4) (5) (6) Primary units Human resources Total Doctors Nurses CHWs Support Panel A: Count CHTs creation 0.06 0.21 0.03 0.07 0.05 0.07 (0.01) (0.08) (0.02) (0.02) (0.03) (0.02) [0.00] [0.01] [0.20] [0.00] [0.13] [0.00] Pre-treatment mean 1.47 1.96 0.31 0.38 0.54 0.40 Panel B: Share CHTs creation 0.00 0.01 -0.02 0.02 (0.01) (0.00) (0.01) (0.00) [0.72] [0.00] [0.01] [0.00] Pre-treatment mean 0.14 0.19 0.29 0.19 Muni-year 729 729 729 729 729 729 Municipality 250 250 250 250 250 250 Notes: Estimated coefficients from an linear regressions of the dependent variable on a binary treatment indicator that takes values equal to one for treated municipalities, after the creation of CHTs (i.e. registered at least 5% of its population), and zero otherwise, following Equation 1. Dependent variables by panel: (A) absolute numbers, and (B) share out of total human resources. Dependent variables in Panel (A) are defined per 1,000 inhabitants. We include municipality and year fixed effects in all estimations. Only three time periods are included in the data: 2009, 2010 and 2015. Outcome values are missing for four municipalities included in the main analysis. Standard errors clustered by municipality in parentheses and p-values in brackets. 30 Figure 2: Consultations and Hospitalizations A. Preventive consultations B. Amenable curative consultations -100 0 100 200 300 400 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -100 -50 0 50 100 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs C. Preventable hospitalizations D. Hospitalizations due to external causes -3 -2 -1 0 1 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -10 0 10 20 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Notes: Coefficients from the fully dynamic specification following Equation 2and estimated using the imputation estimator developed by Borusyak et al. (2024). The y-axis shows the average treatment effects and the x-axis the year relative to the creation of the CHTs. Dependent variables by panel: (A) Preventive consultations: total consultations for preventive care; (B) Amenable curative consultations: total curative consultations due to conditions amenable to effective primary healthcare; (C) Preventable hospitalizations: total hospital discharges due to conditions that can be effectively managed, treated, or prevented in a primary care or outpatient setting, and that if not appropriately addressed, could lead to unnecessary hospitalizations or complications; and (D) Hospitalizations due to external causes: total hospital discharges due to accidents and circumstances as the cause of environmental events and circumstances as the cause of injury, poisoning and other adverse effects. All outcomes are measured per 1,000 inhabitants. Confidence intervals at the 95% level. 31 Figure 3: Amenable Curative Consultations and Preventable Hospitalizations, by Disease Type A. Amenable curative consultations, CDs B. Amenable curative consultations, NCDs -200 -100 0 100 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -50 0 50 100 150 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs C. Preventable hospitalizations, CDs D. Preventable hospitalizations, NCDs -1.5 -1 -.5 0 .5 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -1 -.5 0 .5 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Notes: Coefficients from the fully dynamic specification following Equation 2and estimated using the imputation estimator developed by Borusyak et al. (2024). The y-axis shows the average treatment effects and the x-axis the year relative to the creation of the CHTs. Panels (A) and (B) correspond to Amenable curative consultations, total curative consultations due to conditions amenable to effective primary healthcare, split by communicable (CDs) and non-communicable diseases (NCDs), respectively. Panels (C) and (D) correspond to Preventable hospitalizations, total hospital discharges due to conditions that can be effectively managed, treated, or prevented in a primary care or outpatient setting, and that if not appropriately addressed, could lead to unnecessary hospitalizations or complications, split by CDs and NCDs, respectively. All outcomes are measured per 1,000 inhabitants. Confidence intervals at the 95% level. 32 Table 2: Pre-treatment Effects on Consultations and Hospitalizations (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 7.17 18.80 0.08 (23.90) (34.88) (0.24) [0.76] [0.59] [0.76] Pre-treatment mean 508.03 1103.59 7.41 Panel B. Communicable diseases CHTs creation 27.25 -0.13 (19.15) (0.15) [0.15] [0.39] Pre-treatment mean 855.16 4.03 Panel C. Non-communicable diseases CHTs creation -8.46 0.21 (18.17) (0.14) [0.64] [0.14] Pre-treatment mean 248.43 3.38 Muni-year 2540 2540 3556 Municipality 254 254 254 Notes: Estimates correspond to a linear combination of the pre-trend coefficients estimates in Figures 2 and 3for each corresponding outcome following Equation 2and using Borusyak et al. (2024)’s methodology. Dependent variables by column: (1) Preventive consultations: total consultations for preventive care; (2) Amenable curative consultations: total curative consultations for conditions amenable to effective primary healthcare; and (3) Preventable hospitalizations: total hospital discharges due to conditions that can be effectively managed, treated, or prevented in a primary care or outpatient setting, and that if not appropriately addressed, could lead to unnecessary hospitalizations or complications. These three outcomes are split by communicable diseases in Panel B and non-communicable diseases in Panel C. All outcomes are measured per 1,000 inhabitants. 33 Table 3: Effects on Consultations and Hospitalizations (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 187.56 -28.60 -0.80 (24.75) (22.90) (0.31) [0.00] [0.21] [0.01] Pre-treatment mean 508.03 1103.59 7.41 Panel B. Communicable diseases CHTs creation -73.06 -0.55 (21.28) (0.18) [0.00] [0.00] Pre-treatment mean 855.16 4.03 Panel C. Non-communicable diseases CHTs creation 44.46 -0.25 (8.56) (0.17) [0.00] [0.14] Pre-treatment mean 248.43 3.38 Muni-year 2540 2540 3556 Municipality 254 254 254 Notes: Estimates correspond to a linear combination of the average treatment effects estimates in Figures 2and 3for each corresponding outcome following Equation 2and using Borusyak et al. (2024)’s imputation estimator. Dependent variables as presented in Table 2, all measured per 1,000 inhabitants. Standard errors clustered by municipality in parentheses and p-values in brackets. 34 Table 4: Static DiD - Consultations and Hospitalizations (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 72.27 -25.02 -0.75 (21.88) (17.90) (0.28) [0.00] [0.16] [0.01] Pre-treatment mean 543.68 778.06 7.88 Panel B. Communicable diseases CHTs creation -31.36 -0.49 (14.64) (0.17) [0.03] [0.00] Pre-treatment mean 551.08 3.99 Panel C. Non-communicable diseases CHTs creation 6.34 -0.26 (7.27) (0.15) [0.38] [0.09] Pre-treatment mean 226.98 3.90 Observations 2540 2540 3556 Municipalities 254 254 254 Notes: Same notes as Table 3. Coefficients correspond to estimates of the effect of “CHTs creation” using equation (1). “CHTs creation” is an indicator variable that equals to one from the first year in which CHTs start operations in a municipality. 35 APPENDIX This online appendix provides additional information on the data, methods, and robustness checks. 1 A Additional material for Section 3, Data Table A1: ICD-10 Codes for Curative Consultations and Hospitalizations Amenable to Healthcare Ambulatory Care ICD-10 Description Communicable Noncommunicable Sensitive Conditions A00 Cholera x x A01 Typhoid and paratyphoid fevers x x A02 Other salmonella infections x x A03 Shigellosis x x A04 Other bacterial intestinal infections x x A05 Other bacterial foodborne intoxications, not elsewhere classified x x A06 Amoebiasis x x A07 Other protozoal intestinal diseases x x A08 Viral and other specified intestinal infections xx A09 Other gastroenteritis and colitis of infectious and unspecified origin x x A15 Respiratory tuberculosis, bacteriologically and histologically confirmed x x A16 Respiratory tuberculosis, not confirmed bacteriologically or histologically x x A17 Tuberculosis of nervous system x x A18 Tuberculosis of other organs x x A19 Miliary tuberculosis x A20 Plague x A21 Tularaemia x A22 Anthrax x A23 Brucellosis x A24 Glanders and melioidosis x A25 Rat-bite fevers x A26 Erysipeloid x A27 Leptospirosis x A28 Other zoonotic bacterial diseases, not elsewhere classified x A30 Leprosy [Hansen disease] x A31 Infection due to other mycobacteria x A32 Listeriosis x A33 Tetanus neonatorum x x A34 Obstetrical tetanus x A35 Other tetanus x x A36 Diphtheria x x A37 Whooping cough x x A38 Scarlet fever x A39 Meningococcal infection x A40 Streptococcal sepsis x A41 Other sepsis x A42 Actinomycosis x A43 Nocardiosis x A44 Bartonellosis x A46 Erysipelas x x A48 Other bacterial diseases, not elsewhere classified x Continued on next page 2 Table A1 – Continued from previous page ICD-10 Description AHC Communicable AHC Noncommunicable ACSC Q22 Congenital malformations of pulmonary and tricuspid valves x Q23 Congenital malformations of aortic and mitral valves x Q24 Other congenital malformations of heart x Q25 Congenital malformations of great arteries x Q26 Congenital malformations of great veins x Q27 Other congenital malformations of peripheral vascular system x Q28 Other congenital malformations of circulatory system x Z72 Problems related to lifestyle x Z73 Problems related to life-management difficulty x Z74 Problems related to care-provider dependency x Z75 Problems related to medical facilities and other health care x Z76 Persons encountering health services in other circumstances x Z80 Family history of malignant neoplasm x Z81 Family history of mental and behavioural disorders x Z82 Family history of certain disabilities and chronic diseases leading to disablement x Z83 Family history of other specific disorders x Z84 Family history of other conditions x Z85 Personal history of malignant neoplasm x Z86 Personal history of certain other diseases x Z87 Personal history of other diseases and conditions x Z88 Personal history of allergy to drugs, medicaments and biological substances x Z89 Acquired absence of limb x Z90 Acquired absence of organs, not elsewhere classified x Z91 Personal history of risk-factors, not elsewhere classified x Z92 Personal history of medical treatment x Z93 Artificial opening status x Z94 Transplanted organ and tissue status x Z95 Presence of cardiac and vascular implants and grafts x Z96 Presence of other functional implants x Z97 Presence of other devices x Z98 Other postsurgical states x Z99 Dependence on enabling machines and devices, not elsewhere classified x * Partially coded as Ambulatory Care Sensitive Conditions 9 Table A2: Descriptive Statistics: Most common ICD-10 codes Amenable curative consultations Preventable hospitalizations (1) (2) (3) (4) ICD-10 Description Pre-treatment mean % Pre-treatment mean % Communicable diseases A04 Other bacterial intestinal infections 0.14 1.83 A06 Amoebiasis 19.64 1.92 0.24 3.31 A08 Viral and other specified intestinal infections 0.20 2.68 A09 Other gastroenteritis and colitis of infectious and unspecified origin 48.80 4.76 1.92 25.90 B35 Dermatophytosis 22.67 2.21 B82 Unspecified intestinal parasitism 43.47 4.24 J00 Acute nasopharyngitis [common cold] 233.78 22.81 0.08 1.15 J02 Acute pharyngitis 133.40 13.01 J06 Acute upper respiratory infections of multiple and unspecified sites 143.68 14.02 0.10 1.37 J15 Bacterial pneumonia, not elsewhere classified 0.18 2.41 J18 Pneumonia, organism unspecified 14.57 1.42 0.07 0.94 J20 Acute bronchitis 24.75 2.41 0.33 4.43 J21 Acute bronchiolitis 0.32 4.36 J30 Vasomotor and allergic rhinitis 18.69 1.82 Non-communicable diseases E11 Type 2 diabetes mellitus 24.62 2.40 0.90 12.20 E14 Unspecified diabetes mellitus 18.02 1.76 0.15 2.01 G40 Epilepsy 11.55 1.13 0.30 4.05 I10 Essential (primary) hypertension 121.14 11.82 0.49 6.66 I11 Hypertensive heart disease 0.06 0.87 I15 Secondary hypertension 4.33 0.42 I64 Stroke, not specified as haemorrhage or infarction 0.09 1.24 I67 Other cerebrovascular diseases 0.12 1.61 J40 Bronchitis, not specified as acute or chronic 5.58 0.54 J44 Other chronic obstructive pulmonary disease 3.60 0.35 0.36 4.85 J45 Asthma 18.32 1.79 0.63 8.52 J46 Status asthmaticus 0.07 0.91 K40 Inguinal hernia 2.96 0.29 K80 Cholelithiasis 2.90 0.28 Note: Columns (1) and (3) report the municipality average by the ten most common ICD-10 codes in the pre-treatment period across municipalities. Columns (1) and (3) show the mean for each condition and (2) and (4) show the mean as a percentage of the overall mean for each outcome. 10 Table A3: Descriptive Statistics: Consultations and Hospitalizations by Type All Treated (1) (2) (3) (4) Pre-treatment Post-treatment Pre-treatment Post-treatment Panel A: Consultations Total consultations 2398.66 2621.09 2582.37 2862.98 Preventive consultations 465.77 830.28 508.03 917.97 % of total consultations 20.75 32.31 21.41 32.96 Curative consultations 1932.88 1790.80 2074.34 1945.02 % of total consultations 79.25 67.69 78.59 67.04 Curative amenable consultations 1024.71 844.16 1103.59 916.14 % curative consultations 53.23 47.29 53.50 47.30 Curative amenable consultations - CDs 791.60 551.78 855.16 598.05 % amenable curative consultations 77.32 65.11 77.73 65.24 Curative amenable consultations - NCDs 233.11 292.38 248.43 318.09 % amenable curative consultations 22.68 34.89 22.27 34.76 Panel A: Hospitalizations Total hospitalizations 54.35 67.40 54.89 67.98 Preventable hospitalizations 7.32 8.64 7.41 8.58 % of total hospitalizations 13.14 12.44 13.17 12.29 Preventable hospitalizations - CDs 3.89 4.12 4.03 4.14 % preventable hospitalizations 52.96 48.02 54.26 48.55 Preventable hospitalizations - NCDs 3.43 4.52 3.38 4.44 % preventable hospitalizations 47.04 51.98 45.74 51.45 Note: This table reports the mean absolute numbers and percentage across municipalities for each type of care and condition. Pre-treatment period for curative and preventive consultations is 2009, and 2005-2009 for hospitalizations. Post treatment period is 2010-2018. 11 B Additional material for Section 5, The Effects of CHTs Table B1: Balance in initial characteristics, by ‘pure’ control and treatment groups Control Treatment Difference (1) (2) (3) Municipal characteristics: Total population 33542.54 18203.39 -15339.16** [49880.75] [33273.57] (6504.22) % Rural population 0.49 0.65 0.16*** [0.28] [0.21] (0.04) % Pop in poverty 0.38 0.48 0.09*** [0.09] [0.10] (0.01) Inputs per 1,000 inhabitants: Primary units 0.11 0.20 0.09*** [0.11] [0.23] (0.02) Total HR 1.44 2.18 0.73*** [0.84] [1.26] (0.14) Doctors 0.18 0.34 0.17*** [0.15] [0.39] (0.03) Nurses 0.29 0.42 0.13*** [0.21] [0.35] (0.04) CHWs 0.41 0.61 0.20*** [0.27] [0.37] (0.04) Admin 0.28 0.45 0.16*** [0.19] [0.41] (0.04) Note. Municipal characteristics are from the 2007 census, and inputs for primary healthcare production are from 2009. Columns 1 and 2 report sample mean with standard deviation in brackets for the control and for the treatment group, respectively. Column 3 reports the difference between the ‘pure’ control group (never treated) and the ‘pure’ treatment group (treated at some point), estimated using OLS, with robust standard errors reported in parentheses. Statistical significance denoted by *** p<0.01, ** p<0.05, * p<0.1. 12 Table B2: Treatment status and timing to start CHTs Status Timing OLS Cox hazard model (1) (2) (3) (4) Municipal characteristics: Total population 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) % Rural population 0.05 0.00 0.27 0.30 (0.18) (0.19) (0.49) (0.55) % Pop in poverty 0.99*** 1.04*** 1.21 1.32 (0.34) (0.35) (1.06) (1.09) Outcomes Preventive consultations 0.00*** 0.00** 0.00 0.00 (0.00) (0.00) (0.00) (0.00) Curative consultations CDs -0.00 -0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) Curative consultations NCDs 0.00 0.00 -0.00 -0.00 (0.00) (0.00) (0.00) (0.00) Hospitalizations CDs 0.01 0.01 -0.05 -0.05 (0.02) (0.02) (0.06) (0.06) Hospitalizations NCDs -0.03 -0.03 0.00 0.02 (0.03) (0.03) (0.07) (0.08) Inputs per 1,000 inhabitants: Primary units -0.21 0.20 (0.25) (0.72) Total HR 0.07 0.01 (0.11) (0.29) Doctors 0.10 0.27 (0.15) (0.51) Nurses -0.19 0.11 (0.14) (0.46) CHWs 0.14 -0.32 (0.14) (0.39) Admin -0.07 -0.34 (0.20) (0.59) Observations 254 250 186 184 Note. Municipal characteristics are from the 2007 census, inputs for primary healthcare production are from 2009, and outcomes are the average of half-year observations in 2009 for consultations and 2005-2009 for hospitalizations. Column (1) shows coefficients of an OLS regression of being treated over initial characteristics. Column (2) shows coefficients of a Cox regression of timing until the start of CHTs. Standard errors are reported in parentheses. Statistical significance denoted by *** p<0.01, ** p<0.05, * p<0.1. 13 Figure B1: Effects on Curative Consultations and Hospitalizations A. Total curative consultations -100 0 100 200 300 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs B. Total hospitalizations -4 -2 0 2 4 6 Average causal effect -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Notes: Same notes as Figure 2. Dependent variables by panel: (A) Curative consultations: total consultations for curative care; (B) Total hospitalizations: total hospital discharges. 14 C Additional material for Section 5.5, Additional tests Table C1: Sensitivity Analysis by Year of CHTS creation (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 5% 187.56 -28.60 -0.80 (24.75) (22.90) (0.31) [0.00] [0.21] [0.01] Pre-treatment mean 5% 508.03 1103.59 7.41 Municipalities treated 5% 186 186 186 CHTs creation 10% 199.96 -28.11 -0.65 (24.78) (22.77) (0.30) [0.00] [0.22] [0.03] Pre-treatment mean 10% 512.95 1108.11 7.35 Municipalities treated 10% 180 180 180 CHTs creation 15% 216.14 -31.78 -0.58 (24.60) (22.54) (0.29) [0.00] [0.16] [0.05] Pre-treatment mean 15% 520.02 1120.24 7.26 Municipalities treated 15% 172 172 172 CHTs creation 20% 230.59 -31.95 -0.55 (24.56) (22.93) (0.28) [0.00] [0.16] [0.05] Pre-treatment mean 20% 527.82 1137.37 7.33 Municipalities treated 20% 163 163 163 Panel B. Communicable diseases CHTs creation 5% -73.06 -0.55 (21.28) (0.18) [0.00] [0.00] Pre-treatment mean 5% 855.16 4.03 Municipalities treated 5% 186 186 CHTs creation 10% -74.58 -0.47 (21.32) (0.17) [0.00] [0.01] Pre-treatment mean 10% 859.70 4.02 Municipalities treated 10% 180 180 CHTs creation 15% -78.40 -0.47 (21.24) (0.17) [0.00] [0.01] Pre-treatment mean 15% 870.22 3.98 Municipalities treated 15% 172 172 CHTs creation 20% -80.56 -0.46 (21.66) (0.16) [0.00] [0.01] Pre-treatment mean 20% 883.95 4.03 Municipalities treated 20% 163 163 Panel C. Non-communicable diseases CHTs creation 5% 44.46 -0.25 (8.56) (0.17) [0.00] [0.14] Pre-treatment mean 5% 248.43 3.38 Municipalities treated 5% 186 186 CHTs creation 10% 46.47 -0.18 (8.41) (0.17) [0.00] [0.27] Pre-treatment mean 10% 248.41 3.33 Municipalities treated 10% 180 180 CHTs creation 15% 46.63 -0.12 (8.36) (0.16) [0.00] [0.46] Pre-treatment mean 15% 250.03 3.28 Municipalities treated 15% 172 172 CHTs creation 20% 48.61 -0.10 (8.38) (0.15) [0.00] [0.53] Pre-treatment mean 20% 253.43 3.30 Municipalities treated 20% 163 163 Notes: Same notes as Table 3. In each specification, the year of the creation of CHTs varies depending on the percentage of the municipality’s population that health teams enrolled. 15 Table C2: Placebo for Inputs for Primary Healthcare Production using Later-Treated (1) (2) (3) (4) (5) (6) Primary units Human resources Total Doctors Nurses CHWs Support Panel A: Count CHTs creation 0.00 -0.01 -0.00 -0.00 -0.00 -0.00 (0.00) (0.02) (0.00) (0.00) (0.00) (0.00) [0.95] [0.75] [0.43] [0.75] [0.58] [0.97] Pre-treatment mean 1.561 1.764 0.233 0.323 0.539 0.340 Panel B: Share CHTs creation -0.00 0.00 -0.00 0.00 (0.00) (0.00) (0.00) (0.00) [0.319] [0.353] [0.319] [0.319] Pre-treatment mean 0.13 0.18 0.31 0.18 Muni-year 329 329 329 329 329 329 Municipality 165 165 165 165 165 165 Notes: Same notes as Table 1. Coefficients correspond to estimates of the effect of ‘CHTs creation’ using equation (1). ‘CHTs creation’ is an indicator variable that equals to one in 2010 for municipalities treated after 2010 as a placebo test. Analysis excludes the cohort of municipalities treated in 2010 (T2010). Sample includes the years 2009 and 2010. Table C3: Inputs for Primary Healthcare Production, Excluding T2010 (1) (2) (3) (4) (5) (6) Primary units Human resources Total Doctors Nurses CHWs Support Panel A: Count CHTs creation 0.06 0.21 0.03 0.07 0.05 0.07 (0.01) (0.08) (0.02) (0.02) (0.03) (0.02) [0.00] [0.01] [0.20] [0.00] [0.13] [0.00] Pre-treatment mean 1.47 1.96 0.31 0.38 0.54 0.40 Panel B: Share CHTs creation 0.00 0.01 -0.02 0.02 (0.01) (0.00) (0.01) (0.00) [0.72] [0.00] [0.01] [0.00] Pre-treatment mean 0.14 0.19 0.29 0.19 Muni-year 719 719 719 719 719 719 Municipality 240 240 240 240 240 240 Notes: Same Notes as Table 1. Analysis excludes the cohort of municipalities treated in 2010 (T2010). Sample includes the years 2009, 2010 and 2015. 16 Figure C1: Consultations and Hospitalizations, Alternative TWFE Estimators A. Preventive consultations -100 0 100 200 300 400 Treatment effects -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Borusyak et al. de Chaisemartin-D'Haultfoeuille Callaway-Sant'Anna Sun-Abraham C. Preventable hospitalizations -2 -1 0 1 Treatment effects -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Borusyak et al. de Chaisemartin-D'Haultfoeuille Callaway-Sant'Anna Sun-Abraham Notes: Same notes as Figure 2. In addition to the imputation estimator of Borusyak et al. (2024), we use three robust estimators: De Chaisemartin and d’Haultfoeuille (2020), Sun and Abraham (2021), and Callaway and Sant’Anna (2021). 17 Figure C2: Amenable Curative Consultations by Disease Type, Alternative TWFE Estimators A. Amenable curative consultations, CDs -300 -200 -100 0 100 Treatment effects -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Borusyak et al. de Chaisemartin-D'Haultfoeuille Callaway-Sant'Anna Sun-Abraham B. Amenable curative consultations, NCDs -50 0 50 100 150 Treatment effects -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Borusyak et al. de Chaisemartin-D'Haultfoeuille Callaway-Sant'Anna Sun-Abraham Notes: Same notes as Figure 3. In addition to the imputation estimator of Borusyak et al. (2024), we use three robust estimators: De Chaisemartin and d’Haultfoeuille (2020), Sun and Abraham (2021), and Callaway and Sant’Anna (2021). 18 Table C4: Pre-treatment Effects on Consultations and Hospitalizations, Half-Year Data (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 0.37 11.97 0.04 (10.69) (14.63) (0.12) [0.97] [0.41] [0.72] Pre-treatment mean 254.01 551.79 3.70 Panel B. Communicable diseases CHTs creation 14.30 -0.03 (9.36) (0.08) [0.13] [0.70] Pre-treatment mean 427.58 2.01 Panel C. Non-communicable diseases CHTs creation -2.33 0.07 (10.25) (0.07) [0.82] [0.28] Pre-treatment mean 124.22 1.69 Muni-year 5080 5080 7112 Municipality 254 254 254 Notes: Same Notes as Table 2. Data at the half-year level. 25 Table C5: Consultations and Hospitalizations, Total and by Type, Half-Year Data (1) (2) (3) Preventive consultations Amenable curative consultations Preventable hospitalizations Panel A: Total CHTs creation 60.47 -15.31 -0.38 (8.68) (9.82) (0.16) [0.00] [0.12] [0.02] Pre-treatment mean 254.01 551.79 3.70 Panel B. Communicable diseases CHTs creation -28.73 -0.28 (8.80) (0.09) [0.00] [0.00] Pre-treatment mean 427.58 2.01 Panel C. Non-communicable diseases CHTs creation 13.42 -0.09 (3.14) (0.09) [0.00] [0.31] Pre-treatment mean 124.22 1.69 Muni-year 5080 5080 7112 Municipality 254 254 254 Notes: Same Notes as Table 3. Data at the half-year level. 26 Figure C9: Effects on Amenable Hospitalizations and ACSC Hospitaizations A. Amenable hospitalizations, CDs B. Amenable hospitalizations, NCDs -2 -1 0 1 Average causal effect -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -1 -.5 0 .5 1 1.5 Average causal effect -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs C. ACSC hospitalizations, CDs D. ACSC hospitalizations, NCDs -1 -.5 0 .5 Average causal effect -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -3 -2 -1 0 1 2 Average causal effect -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Notes: Same notes as Figure 3. Dependent variables are: in Panels (A) and (B) the total hospital discharges due to conditions amenable to effective primary care following Kruk et al. (2018)’s classification (the same one used for amenable curative care), split by communicable (CDs) and non-communicable diseases (NCDs), respectively; in Panels (C) and (D) the total hospital discharges due to ambulatory care sensitive conditions, following Rodriguez Abrego (2012)’s classification, split by communicable (CDs) and non-communicable diseases (NCDs), respectively. 27 Figure C10: Consultations and Hospitalizations, Poverty x Year Dummies A. Preventive consultations C. Preventable hospitalizations -100 0 100 200 300 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -4 -3 -2 -1 0 1 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs A. Amenable curative consultations, CDs B. Amenable curative consultations, NCDs -200 -100 0 100 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -50 0 50 100 150 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs C. Preventable hospitalizations, CDs D. Preventable hospitalizations, NCDs -2 -1.5 -1 -.5 0 .5 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs -2 -1.5 -1 -.5 0 .5 Average causal effect -3 -2 -1 0 1 2 3 4 5 6 7 8 Years since the creation of CHTs Notes: Same notes as Figures 2and 3. All specifications control for the initial share of poor population (from the 2007 census) interacted with year dummies. 28 D Additional material for Section 5.6, Amenable Mortality Table D1: Effects on Mortality Rates (1) (2) (3) Amenable MR Non-amenable MR Communicable Non-communicable Total Panel A: Static effect CHTs creation -10.68 -16.77 -5.35 (5.07) (19.88) (22.99) [0.04] [0.40] [0.82] Municipality-year 2032 2032 2032 Municipality 254 254 254 Panel B: Placebo using later treated Treated municipalities 1.16 3.17 23.33 (9.01) (23.72) (33.49) [0.90] [0.89] [0.49] Municipalities 79 79 79 Panel C: Static effect using early treated CHTs creation -11.73 -14.98 -4.06 (5.29) (20.59) (24.31) [0.03] [0.47] [0.87] 2011 control mean 42.33 123.37 331.03 Municipality-year 632 632 632 Municipality 79 79 79 Note: This table reports the results of the effect estimated from an linear regressions of the dependent variable on a binary treatment indicator that takes values equal to one for treated municipalities, after the creation of CHTs (i.e. enrolled at least 5% of its population), and zero otherwise, following Equation 1. Dependent variables by column: (1) Communicable: amenable mortality rate caused by communicable diseases; (2) Non-communicable: amenable mortality rate caused by non-communicable diseases; (3) No AMR: mortality rate by diseases not amenable to healthcare. Amenable mortality are deaths avoidable through access to quality healthcare, which we classify following the definition by Kruk et al. (2018). All outcomes are measured per 1,000 inhabitants. Panel data of mortality rates is available yearly between 2011 and 2018. Panels A and C present coefficients of a static difference-in-difference estimation following Equation 1. Panel B presents coefficients of a cross-sectional OLS estimation of being treated later (after 2011) as opposed to never treated using 2011 data. Panels B and C drop from the sample of analysis municipalities that were treated before or in 2011 (T2010 and T2011). We include municipality and year fixed effects in all estimations in Panel A and C. Standard errors clustered by municipality in parenthesis and p-values in brackets. 29