Proximity Matters: The Impact of Urgent Care Centers on Emergency Department Arrivals

2.1. Addressing ED Overuse: Access to Alternative Facilities

To reduce unnecessary ED visits, many healthcare providers and policymakers have turned to expanding access to lower cost, alternative care options, including primary care, retail clinics, and UCCs. However, the evidence on whether such alternatives reduce ED demand has been mixed.

Retail Clinics and Primary Care. Retail clinics (RCs), often located in pharmacies or supermarkets and typically staffed by nurse practitioners, offer a limited clinical scope. Martsolf et al. (2017) show that RC penetration has no meaningful reduction in the number of low-acuity ED visits, except for privately insured patients. Bavafa et al. (2022) show that a negative shock to a primary care provider’s availability increases nonemergent ED visits, and similarly, Yoon et al. (2015) show that increasing same-day access to primary care reduces nonemergent ED visits. However, Oterino de la Fuente et al. (2007) shows a simultaneous increase in primary care and ED visits.

Urgent Care Centers. UCCs have received more attention due to their broader scope of services. Yet the direction and magnitude of their impact remains varied. Allen et al. (2021) find that having an open UCC in a ZIP code area reduces the number of ED visits for less emergent conditions by 17.2%. Pacheco et al. (2019) compare monthly ED visits per 1,000 people in areas with and without UCC openings and find a 2.69% reduction. Wang et al. (2021) conduct a ZIP-code-level analysis and estimate that an increase of 1,000 UCC visits is associated with only 27 fewer low-acuity ED visits. Other studies report no consistent impact: Yakobi (2017) finds no change in annual ED census across four hospitals in the same catchment area; and Merritt et al. (2000) observe inconclusive patterns in follow-up ED visits after UCC use.

These studies rely on geographically aggregated data and define UCC “exposure” at the ZIP code, catchment area, or regional level. This approach treats all UCC–ED pairs within an area as equivalent, regardless of their actual spatial relationship. Such aggregation risks masking important heterogeneity in patient behavior and facility substitution patterns, especially since, as discussed below, locations are known drivers of patients’ facility choices. This calls for more granular investigation into how specific spatial configurations shape ED demand and facility choices.

One exception is Carlson et al. (2020), which examines whether proximity to a UCC, defined as living within 1 mile, affects a patient’s likelihood of visiting the ED for a low-acuity condition. The outcome is binary: whether an ED visit is low-acuity, and the independent variable is also binary: whether the patient making the visit lives within 1 mile of any UCC. Using encounter-level data in two EDs and a logistic regression model, the authors find mixed results: a lower odds of low-acuity ED visits among patients living within 1 mile of a UCC at one medical center, and no significant effect at another. This design essentially estimates whether proximity affects the type of ED visits patients make, but not whether total ED demand changes; a shift in odds does not necessarily imply a reduction in volume. In line with Carlson et al. (2020), we contribute to the literature on location heterogeneity in UCC effects, but our research question differs. We estimate the heterogeneous causal impact of UCC on total ED arrivals by ED–UCC proximity during UCC operating hours, especially among lower acuity patients.

More importantly, examining UCC impact through patient–UCC proximity or regional-level UCC availability, as in Carlson et al. (2020) and the aggregate-level studies, addresses a different set of questions and yields different practical implications than our focus on ED–UCC proximity. Patient–UCC proximity speaks to how long-term care-seeking behavior is shaped by the greater cost-effectiveness and geographic accessibility of a UCC relative to an ED. In contrast, our ED–UCC proximity analysis holds constant these inherent cost and convenience advantages of UCCs and asks whether UCCs’ spatial configuration relative to the ED has additional effects on ED demand.

Our study complements long-term access-focused analyses by showing that facility siting itself can influence patient choices between EDs and UCCs in ways that the general features of UCCs and patient–UCC proximity—such as lower cost, shorter waits, and no appointment requirements—cannot fully explain. From a policy and planning perspective, our work offers new insight for facility location design: two UCCs may serve the same population and provide similar long-term access benefits, yet their impact on ED demand can differ markedly depending on their spatial configuration relative to the ED. Particularly, a collocated setup may unlock additional UCC impact by removing frictions that hinder switching between facilities.

2.2. Collocating Facilities and Spatial Configuration

The importance of facility location in shaping service outcomes has been well-documented. Stahl (1982) finds that collocated facilities and sellers reduce customers’ search costs. The benefit of collocation is considered to be particularly salient when service requires visual inspection and when service heterogeneity is high (McCann and Folta, 2008). These findings suggest that physical proximity may enable real-time decision-making—behavioral patterns that are unlikely to arise when facilities are spatially separated.

Several studies have examined the impact of physically collocating healthcare facilities. Collocating a psychiatric emergency center and a general ED reduces wait time for psychiatric patients (Frank et al., 2005); collocating mental and physical health resources reduces patient length of stay, readmission rate, and suicidal behavior (Ahuja et al., 2026). In the context of the British National Health Service, collocating walk-in centers ¹ alongside EDs does not seem to bring in evident changes in ED patient volume in one study (Salisbury et al., 2007). Later work in a different city, however, does find an 8% reduction in nonurgent ED attendances (Arain et al., 2015).

Only two studies, to the best of our knowledge, have directly examined the collocated UCC–ED configuration. Raidla et al. (2020) show that managing one triage line across the collocated UCC and ED leads to a decrease in length of stay and door-to-doctor time of the patients triaged to the UCC. In our setting, patients choose the facility for themselves as opposed to being directed through a triage process. Agarwal et al. (2012) conducted an interview focusing on why patients first present to the ED despite being eligible for UCC care. Neither of the two studies estimates the causal impact of UCC proximity on ED volume.

Compared to the extensive literature examining how UCCs and other alternative facilities affect ED demand at an aggregate level, there is limited empirical work on whether the impact of UCCs varies by their spatial proximity to EDs. Yet this potential heterogeneity is important: understanding how physical proximity shapes patient choice has direct implications for system-level decisions on where to place new UCCs and how to configure them relative to existing EDs. We address this gap by comparing the effects of collocated and non-collocated UCCs on ED demand during UCC operating hours, offering new insight into how location-specific design choices influence patient facility choice outcomes.

2.3. Research Question: Proximity and Patient Facility Choice

Whether ED–UCC proximity leads to different patient behavior is not clear a priori. If patients make care decisions entirely before arrival and based solely on static criteria such as clinical condition, monetary costs, and geographic convenience to patient location—then the spatial relationship between the ED and UCC may have little impact. However, if patients are also responsive to factors such as perceived capability of facility types or (perceived) wait time (Agarwal et al., 2012; Durand et al., 2012; Turnbull et al., 2019: p. 25), then ED–UCC proximity may meaningfully shape patients’ facility choices. Prior studies offer evidence for both types of considerations in patient decision-making.

On one hand, if all current ED patients believe their medical needs cannot be met by UCCs and if clinical need is the sole driver of facility choice, then neither a newly opened collocated nor non-collocated UCC would serve as a substitute. In this case, UCC openings would have no effect on ED demand, regardless of proximity. Patient interviews suggest that proximity to home or work is a major factor in choosing the ED, even when patients acknowledge that their condition does not require emergency care (Agarwal et al., 2012). When a UCC, whether collocated with the ED or not, opens near these convenience-driven patients, offering an even more geographically accessible and lower cost option, they may opt for the UCC instead of the ED. In this scenario, both collocated and non-collocated UCCs could reduce ED volume by serving as more convenient and cost-effective alternatives, but in a similar manner, yielding no differential impact by proximity. Thus, in either case, ED–UCC proximity would not result in heterogeneous outcomes.

On the other hand, if patients make real-time decisions upon arrival to the ED, ED–UCC proximity may lead to heterogeneous impacts on ED volume. Multiple interview studies (Agarwal et al., 2012; Durand et al., 2012; Turnbull et al., 2019: p. 25) indicate that many less severe patients choose the ED not because of clinical need, but because of uncertainty—they are reassured by the availability of diagnostic tools in EDs (e.g., X-rays) and the opportunity to receive care in one place. For these patients, the ED serves as a safe default when they are unsure whether their condition warrants emergency treatment (Turnbull et al., 2019: p. 53). When such patients arrive at an ED with a collocated UCC, they retain the reassurance that their condition will be addressed at the site, but face lower switching costs between the two facilities. This flexibility allows them to postpone their final choice of facility until they can assess the situation on-site. As Vice President of Intuitive Health—a Texas-based firm whose business model centers on collocating EDs and UCCs—Mason Deal explains, collocated models spare patients from the burden of deciding where to go for appropriate care before seeking treatment (Mugford, 2020). In this way, the collocated UCC effectively gets a second look, serving as a fallback that can draw patients away from the ED once they arrive.

We expect that ED congestion is likely to prompt such final facility decisions. In particular, patients who initially enter the ED may observe a crowded waiting room and reconsider their choice. While our partner’s EDs do not communicate expected wait times, patients can infer congestion from the visible census. At a collocated site, this may prompt them to switch to the adjacent UCC, especially when the perceived benefits of ED care no longer justify the cost of delay. In contrast, such dynamic substitution is unlikely at non-collocated sites, where switching to a UCC requires nontrivial travel. Therefore, the added benefit of collocation arises precisely from its ability to accommodate uncertain, delay-sensitive patients who are not fully committed to the ED and are responsive to real-time conditions.

This leads to our primary research question: How does the proximity of a UCC to an ED affect patient substitution between the facilities during UCC operating hours? Further, if there is an effect, what mechanism leads to the substitution? We are among the first to examine how ED–UCC proximity shapes patients’ facility choices. Whereas prior studies treat UCC openings as exogenous shocks and focus solely on ED volume changes using encounter data from the ED alone, we go further: analyzing both ED and UCC arrivals and uncovering the mechanism by which UCC proximity affects ED use.

3.1. Overview of the Hospital-based EDs

At our partner medical system, EDs operate 24 hours a day. Upon arrival to the ED, patients check in and wait to be called to the triage desk, where a triage nurse records their chief complaints and vital signs. The triage nurse then assigns an acuity level to the patient following the Emergency Severity Index (ESI) algorithm. Acuity levels range from 1 to 5, corresponding to immediate, emergent, urgent, less urgent, and nonurgent, respectively. Lower scores generally indicate more severe cases. Acuity levels 1 and 2 reflect distinctly higher urgency than levels 3, 4, and 5. Levels 3, 4, and 5 are distinguished mostly by the number of types of resources (e.g., labs, X-rays, electrocardiogram, etc.) required to make a diagnosis (Gilboy et al., 2005). Lower scores involve more required resources. Level 5 requires no resource, while Levels 4 and 3 require at least one type of resource.

After triage, patients return to the waiting area and wait to be called to the treatment area. No wait time information is disclosed to the patients. The waiting and treatment areas are physically separated; importantly, the treatment area is not visible to patients from the waiting area. The triage nurses manage the patient flow from the waiting area to the treatment area and assign patients to care teams. Once inside the treatment area, the patient occupies an ED bed and is treated by providers affiliated with the assigned team.

The care team is responsible for evaluating the patient and ultimately making a diagnosis and disposition decision. The physician’s disposition decision determines whether the patient will be discharged, admitted (into the inpatient unit for follow-up care), or transferred (to other healthcare facilities). Patients who are discharged—those requiring no further care—typically leave the ED shortly after the disposition decision is made. In contrast, patients who require follow-up care (i.e., will be admitted or transferred) remain in the ED until an inpatient bed becomes available or transportation to the receiving facility is arranged. During the time between the physician’s admission or transfer decision (i.e., bed request) and the patient’s actual departure from the ED, the patient will be boarding in the treatment area.

3.2. Overview of the UCCs

UCCs are not new to the residents who live in and around the metropolitan area we study. Across healthcare systems, UCCs in the region offer similar ranges of services and maintain comparable operating hours.

UCCs operated by our partner medical system are physically separate from EDs. They operate between 8:00 a.m. and 7:45 p.m. or 8:00 p.m., seven days a week, offering a limited range of services. In particular, UCCs treat nonurgent to urgent (acuity levels 3 to 5) encounters such as minor illness (e.g., cold, flu, sore throat, diarrhea), minor injuries (e.g., cuts, bites, sprains), and common issues like pink eye, urinary tract infections, and skin irritations. To support diagnosis and treatment, UCCs are equipped with some resources such as labs and X-rays. However, UCCs do not offer any observation units or in-hospital services. If a physician deems a patient’s condition too severe to be managed at the UCC, the patient will be redirected to an ED. The operating hours, service scope, and test capabilities reflect nationwide standards for UCCs, as summarized by the American Academy of Urgent Care Medicine (AACUM, 2022) and referenced in Section 1. Conversely, ED care providers cannot divert ED patients to a UCC.

Information about UCCs mentioned so far is readily available online. Physicians have also contributed to raising awareness of UCC service capabilities, limitations, and their relative time and cost advantages. As a result, patients should be familiar with the hours and scope of UCC services.

3.3. UCC Openings and ED–UCC Pairs

We observe three UCC openings in our focal medical system between 2013 and 2017: November 2013, June 2016, and November 2017. ² For each UCC, we identify the ED in the medical system that is geographically closest to it. As we will elaborate further in Section 3.4 and Figure EC.1, we show that each UCC opening has limited (indeed, for most cases, negligible) impact on EDs other than the closest ED. We therefore define each newly opened UCC and its nearest ED as a pair. The focus of our study is how differences in the relative location of each ED–UCC pair lead to heterogeneity in the UCC’s impact on ED demand.

The distances between the UCCs and their paired EDs (in chronological order of their openings) are 0.7 miles, 5.7 miles, and 0.1 miles, respectively, based on the driving distance according to Google Maps. For the third ED–UCC pair, the 0.1-mile distance amounts to a 3-minute walk from door to door, effectively negligible. For the other two pairs, transportation between the facilities imposes a non-negligible cost. We refer to the third pair as collocated, and the remaining pairs as non-collocated.

To publicize a UCC opening, the medical system announces it through local news outlets, social media (e.g., Facebook), and highly visible banners placed at the facility. Note that all UCCs in our setting, whether collocated or non-collocated, are operated within the same system and therefore have identical monetary costs, all lower than ED costs. Insurance coverage is the same across all UCC–ED pairs, as both UCCs and EDs are part of the same medical system. Against this backdrop, two questions arise: Do patients switch away from EDs when a new UCC opens? And does the relative proximity of the two facilities influence patients’ facility choices? To address these questions, we obtain encounter-level data surrounding the three UCC openings from our partner medical system.

3.4. Data Description

Our partner medical system collects data through its electronic health record (EHR) systems. We focus on encounters made by adult patients at both EDs and UCCs.

Observation Windows. Our data are collected at the encounter (i.e., patient-visit) level across three observation windows, each centered on one UCC opening. Within each window, we observe encounters at the newly opened UCC as well as at all EDs in the system. Each observation window spans approximately one year on either side of the UCC opening event. For example, observation window 1 covers November 2012 to November 2014, covering the UCC opening in November 2013. We number the three windows in chronological order of the openings. Our data come as three separate datasets, one collected in each observation window. We call each dataset a sample, the UCC opened during the sample the sample UCC, and its nearest ED the focal ED. We see 732, 732, and 663 days in the three samples, respectively. We refer to Section EC.1 for details on how we verify that no other major UCC openings, to the best of our knowledge, occurred near the focal EDs during the same period as our sample UCCs’ openings. The average number of daily encounters (including all acuity levels) at the focal EDs is 234.2, 56.6, and 233.1, and at the sample UCCs is 48.4, 8.9, 49.8, respectively in the three samples. In total, the dataset includes 367,379 encounters at the focal EDs and 30,927 encounters at the sample UCCs. Although Samples 2 and 3 partially overlap in calendar time, the focal EDs differ across the two samples.

Encounter Timestamps. For each encounter (each row), we observe detailed timestamps (hour, minute, and second) that track a patient’s journey through a facility. For ED encounters, the recorded timestamps include the patient’s arrival at the door, triage, first contact with a care provider, roomed time, bed request and assignment time if admitted, the care provider’s disposition decision time, and lastly departure from the ED. For UCC encounters, we observe the patient’s arrival at the door, roomed time, first contact with a care provider, and lastly departure from the UCC facility.

To comply with our partner medical system’s data privacy policy, all encounter dates were deidentified prior to analysis. Specifically, within each sample, each unique date is assigned a binary indicator for whether it falls before or after the sample UCC opening date. Dates are then randomly shuffled within each sample, with each date retaining all of its encounters and the binary label. For example, if two encounters in our data fall on Date 1 and Date 2, the visit dates alone do not reveal which one actually occurred first, unless one was before the paired UCC opened and the other after.

Although the chronological ordering of dates is obscured, two key sources of variation are fully preserved. First, within-day arrival patterns remain intact: all encounters on a given date do show the actual time of the day and therefore retain their relative ordering within the day. Second, the binary indicator identifying whether each date falls before or after the sample UCC opening date is retained. Our empirical strategy (Section 4.1) exploits these two preserved features of the data.

Patient Identifiers and Facility. In addition to timestamps and the shuffled dates, in each encounter, we observe a unique patient identifier and the facility at which the encounter occurred. The patient identifier is consistent across all facilities and over time in each sample, allowing us to track the same individual across encounters. In total, we observe 264,494 unique patients who visited either the focal ED or the sample UCC across three samples. Of these, only 2.0% visited more than one ED, suggesting that patients typically rely on a specific ED for their emergency care needs. Patients make an average of 1.67 visits to either the focal ED or sample UCC ( s . d . = 2.69 ) in any given observation window (i.e., over a roughly 2-year span). Note that we have no data about patient visits within our partner’s network but not at an ED or sample UCC. For example, we do not know whether an ED visit led to a follow-up primary care appointment.

Combining the unique patient identifier with the preserved within-day timestamps enables us to track a patient’s movements across facilities within the same day. For instance, if a patient visits an ED and a UCC on the same day, we can observe the timing of their arrival at each facility. Among dates following UCC opening, 49.2% had at least one patient who visited both the sample UCC and the collocated ED on the same day; the corresponding figure for the non-collocated ED was 41.1%. This provides evidence of patient flow between UCCs and EDs, suggesting that UCC openings can plausibly influence ED demand. We further observe each patient’s ESI acuity level, admission source, and disposition decision.

Pairing UCC with Nearest ED. We now validate our earlier claim that a UCC opening has limited, and in most cases negligible, impact on EDs other than the one nearest to it, that is, the focal ED. To investigate how the relative location of UCCs and EDs drives heterogeneity in impact, we pair each newly opened UCC with its focal ED. This approach rests on the assumption that, if UCCs affect ED utilization, the nearest ED is likely to experience the strongest effect. This assumption is supported by our data.

Particularly, for each window and its corresponding sample, we first identify all patients who visited the sample UCC inside the window. We then examine the ED encounters made by these patients and calculate the share of those encounters that occurred at each ED in the medical system. We label the EDs as “Nearest,” “Second Nearest” to the sample UCC ranked by the UCC–ED relative distance. All other EDs are labeled as “Others.” We find that sample UCC patients almost exclusively flow to the nearest ED, that is, focal ED by our definition, in samples 1 and 3 ( ∼ 98% of patients). In sample 2, the nearest ED was still the dominant destination, capturing the majority ( ∼ 66%) of visits. Note that the focal EDs in samples 1 and 3 serve more densely populated areas, whereas the focal ED in sample 2 is a suburban center hospital. By contrast, other EDs show minimal overlap in their patient populations with the sample UCCs and appear largely unaffected by UCC openings. We illustrate this finding using a 100% stacked bar chart for each observation window in Section EC.2, Figure EC.1 of the E-companion. These patterns support our analytic choice to study the impact of each UCC on its closest ED; patients make local decisions when choosing their care facilities.

4.1. Identification Strategy

The discussion thus far highlights key features of our data: (1) we observe three separate samples, each centered on a different UCC opening and including encounters from both the sample UCC and all EDs in our partner medical system; (2) within each sample, we observe granular ED arrival timestamps for all encounters before and after the UCC opening, although the actual calendar dates are obscured, with only a binary indicator denoting before- versus after-opening. Our identification strategy builds on these data features. We start by describing our identification strategy within each sample; our main specification then pools across the three samples for estimation.

A Within-day Difference-in-Differences Design. As mentioned in Section 3.4, randomly shuffled dates in our data obscure chronological ordering of dates and therefore do not permit the construction of relative-time measures such as “days since UCC opening.” This limitation precludes us from implementing the canonical multi-ED difference-in-differences design, which would compare treated and control EDs over calendar time, to estimate the average treatment effect of UCC openings on treated EDs.

Instead, we adopt a within-day difference-in-differences (DiD) framework to estimate the causal effect of UCC openings on ED patient arrivals during UCC operating hours. In each sample, there is one focal ED (the nearest to UCC), as discussed. We exploit two sources of variation in this focal ED: (1) UCC operating hours (8 a.m. to 8 p.m., which we refer to as daytime) versus non-operating hours (8 p.m. to 8 a.m., which we refer to as nighttime), and (2) dates before versus after UCC opening. For the bulk of our analysis, we work with 15-minute intervals. Our unit of analysis is ED × date × time-of-day interval. Note that our data preserves within-day patterns and permits within-day analysis. Drawing an analogy with canonical DiD terminology, our design follows a framework in which dates serve as the unit dimension and time-of-day intervals serve as the time dimension. Dates before UCC opening constitute control units, and dates after constitute treated units. The treatment is whether the UCC is operating at a given time of day. Non-operating hours constitute the pre-treatment period, and operating hours constitute the post-treatment period.

The Within-day DiD Estimator. The causal effect of UCC operations is then identified by comparing within-day arrival patterns on dates before and after the UCC opening. Our DiD estimator summarizes this comparison by contrasting the treated–control (dates) difference during operating hours with the treated–control (dates) difference during non-operating hours, equivalently, the change in the day–night gap from dates before to dates after the UCC opening. This estimator captures the causal impact of UCC opening on EDs during UCC operating hours. Our key identifying assumption is that, absent UCC operations, the difference in ED arrivals between dates after opening (treated) and dates before opening (control) would be stable across time-of-day intervals. Control dates provide a counterfactual for the treated dates over the course of a day: what the ED arrival pattern would have been had the focal UCC not been operating.

Exploiting an Institutional Feature: UCC Operating Hours. Our design exploits a source of exogenous within-day variation that supports the plausibility of our key identifying assumption and also addresses key endogeneity concerns. First, UCC operating hours in our partner medical system are standardized, fixed, and determined by system-wide policy: all UCCs operate from 8 a.m. to 8 p.m., seven days a week, consistent with regional and national norms (AACUM, 2022); these hours were established prior to individual UCC openings and remain unchanged throughout the study period. Consequently, on dates after UCC opening, UCC availability mechanically switches “on” and “off” at predetermined hours that are unrelated to contemporaneous local ED demand. That is, the treatment switch, when UCC operations begin each day, is not only determined by external convention but also set prior to local demand realization, making it exogenous in both who sets it and when it is set.

Our estimator leverages this policy-driven temporal discontinuity by comparing ED arrivals during UCC operating hours with arrivals during non-operating hours, within the same date. Because the operating schedule is not chosen in response to short-run demand fluctuations, there is no inherent reason for untreated potential outcomes to exhibit a discontinuous change at the operating-hour boundary. That is, absent UCC operations, the treated–control difference in ED arrivals should not shift specifically at the operating-hour boundary. Therefore, any divergence between treated and control dates that emerges specifically at the boundary is unlikely to reflect other coincident changes, and instead points to UCC availability. This institutional feature ensures that the treatment switch itself is exogenous to local ED demand conditions.

Second, the comparison across dates eliminates confounders that affect overall ED demand but are unrelated to the within-day pattern. As we compare each ED to itself before and after UCC opening, time-invariant ED characteristics are differenced out. Moreover, by focusing on a 1-year window before and after each UCC opening, we limit exposure to slow-moving changes in, for instance, patient composition and behavior. Over a relatively short horizon, it is less likely that features of the ED or its underlying patient base change sharply enough to generate differential time-of-day patterns unrelated to UCC operations.

Unlike a multi-ED DiD design that compares different EDs, our identification does not exploit variations across EDs. UCC placement across EDs may reflect strategic planning considerations and therefore raise concerns about endogenous treatment at the ED level, potentially biasing cross-ED estimates. By instead exploiting within-ED variation in UCC operating status, holding the ED fixed, and using a predetermined on/off schedule, our design avoids this source of selection bias.

Identifying Assumptions. Our key identifying assumption, that control dates provide a valid counterfactual time-of-day pattern for treated dates, requires that no unrelated changes alter ED within-day arrival patterns following UCC opening. The institutional feature discussed above supports the plausibility of this assumption: there is no inherent reason for the within-day arrival pattern to shift at the operating-hour boundary absent UCC operations. For this assumption to be violated, a confounding change would need to coincide with both the UCC opening date and the specific operating-hour boundary. Hence, the plausibility of our key assumption rests on additional conditions: (1) No differential pre-trends: the arrival patterns of treated and control dates do not diverge during non-operating hours; (2) No demand-side shifts at the operating-hour boundary: patient behavior and acuity mix do not change in ways that alter ED within-day arrival patterns following UCC opening, including the possibility of timing substitution, whereby patients defer early-morning ED visits until the UCC is open; (3) No supply-side shifts at the operating-hour boundary: staffing, triage, or workflow do not change in ways that alter ED within-day arrival patterns following UCC opening. We assess in detail the above conditions in Section 6.

4.2. Variable Construction

4.3. Model-free Evidence

We first present model-free evidence in Figure 1, which shows one subfigure for each sample. The facet labels show the distance (in miles) between the focal ED and UCC in the sample. The horizontal axis plots the hour of day. The vertical dashed line at 8:00 a.m. marks the time when UCC operations begin, and treatment is considered “switched on.” For each hour, we calculate the average number of less severe encounters during that hour across 1) control dates (dates before UCC opening), and 2) treated dates (dates after UCC opening). We then apply a log transformation to each of the two averages and compute their differences: the logged average on treated dates minus the logged average on control dates. The vertical axis plots the resulting differences. This difference in logs is consistent with a Poisson model—a log link function applied to a conditional mean—which we adopt to model counts in our main specification.

OPEN IN VIEWER

Figure 1.

Model free evidence: ED encounters before vs. after UCC opening.

Figure 1 shows that less severe encounters decrease in the ED that is collocated with a UCC (Distance 0.1). The reduction in patients’ initial ED visits ( N u m I n i E n c ) appears particularly notable; although smaller in magnitude, the total number of ED encounters ( N u m E n c ) also declines. In contrast, for the two non-collocated UCCs (Distance 0.7 and 5.7), ED encounters do not appear to decrease and may have increased slightly. The same patterns hold when using level differences: instead of computing the difference in logged averages, we compute the difference in raw averages. These model-free patterns provide suggestive evidence that the impact of UCC on ED demand during UCC operating hours may vary with ED–UCC proximity. In particular, a reduction in ED demand appears to occur during UCC operating hours only when the UCC is collocated with the focal ED. We formally test and quantify these observed patterns using the empirical model presented in the next section.

In Table 1, we report means and standard deviations for the outcome variables, stratified by operating status (operating or non-operating hours), date types (before or after a UCC opening), and UCC types (collocated or non-collocated). We also provide the mean and standard deviations of the daily share of encounters by acuity and the daily number of encounters at the sample UCCs. ED arrivals during non-operating hours (nighttime) are highly stable before and after UCC opening: for collocated EDs, nighttime encounters average changed from 44.0 to 44.4 per day; for non-collocated EDs, it changed from 27.9 to 27.8. In contrast, differences appear primarily during operating hours (daytime): when a collocated UCC opens, daytime ED arrivals decline from 87.0 to 84.6, while non-collocated case exhibits a modest increase.

OPEN IN VIEWER

Table 1.

Summary statistics.

		Collocated				Non-collocated
		Before Opening		After Opening		Before Opening		After Opening
Variable	Time of Day	Mean	SD	Mean	SD	Mean	SD	Mean	SD
NumIniEnc	Non-operating	44.0	8.0	44.4	8.3	27.9	17.0	27.8	16.5
NumIniEnc	Operating	87.0	12.1	84.6	11.4	63.8	34.5	64.7	34.0
NumEnc	Non-operating	44.0	8.0	44.5	8.3	27.9	17.0	27.9	16.6
NumEnc	Operating	87.0	12.1	85.4	11.3	63.8	34.5	65.5	34.6
Share of Urgent (%)	All day	41.1	4.0	41.3	3.9	48.4	9.1	50.0	8.9
Share of Less Urgent (%)	All day	13.7	2.8	14.5	3.2	20.4	5.4	19.8	5.1
Share of Nonurgent (%)	All day	1.4	0.9	1.9	1.1	1.5	1.3	1.2	1.1
# Encounters (Sample UCC)	Operating	–	–	49.8	16.2	–	–	28.8	22.9

To construct the table, we conduct calculations as follows. For instance, 44.0 in (Non-operating, before opening, collocated) cell is calculated by counting the total encounters during all non-operating hours for each date before the collocated UCC opening, and then take the average across all dates. 27.9 in (Non-operating, before opening, non-collocated) cell is calculated by counting the total encounters during all non-operating hours for each date in each of the two non-collocated samples, and then take the average across all sample-dates. ³

Formally, we stack all dates observed across the three samples and estimate the following specification:

E [ y s d t ∣ X s d t ] = exp ( β 1 c o l U C C s d × U C C O p e r a t e t + β 2 n o n c o l U C C s d × U C C O p e r a t e t + λ s d × η s + γ t × η s + X s d t ′ φ 4 ) .

(1)

The outcome variable y s d t denotes our outcomes of interest at the focal ED in sample s , date d , and time-of-day interval t . c o l U C C s d , n o n c o l U C C s d , and U C C O p e r a t e t are as defined in Section 4.2.1. In particular, we let t represent a 15-minute time interval, a common unit of analysis in literature (Lu et al., 2013); we refer to this as the time-of-day throughout. In the main analysis, we define U C C O p e r a t e t = 1 from 7:45 a.m. to 8:00 p.m., as UCC arrival timestamps across samples show entries between 7:45 and 8:00 a.m., indicating that UCC staff regularly admitted patients who arrived shortly before the official opening. We include a full set of fixed effects: date fixed effects ( λ s d ), time-of-day fixed effects ( γ t ), and their interactions with sample fixed effects η s (that is, λ s d × η s , γ t × η s ). These two-way interactions involving η s ensure that the identification is driven solely by within-sample variation. This specification controls for any unobserved shocks specific to a date or time-of-day. For example, seasonal variation on the date level or routine spikes in demand around lunch time would not bias our estimates.

We also include additional covariates in X s d t . First, neither date nor time-of-day fixed effects control for systematic variation in the within-day distribution across days. For instance, a flu season may not only increase total daily volume but also concentrate arrivals disproportionately in morning hours. To address this, we include interactions G s d × γ t , where G s d indexes the total number of urgent to nonurgent encounters on date d at the focal ED in sample s and γ t is the time-of-day fixed effects. Total daily volume plausibly reflects underlying factors that drive differences in diurnal profiles, while remaining coarser than the date × time-of-day level at which our treatment effect is identified. This yields a separate time-of-day profile for each observed level of daily volume, nonparametrically absorbing volume-dependent diurnal arrival patterns. Our results are also robust to excluding these temporal (within-day) controls (see EC.4.4.1).

Second, we control for system-wide demand fluctuations using two covariates measured at the ED × date × time-of-day level: 1) N u m E n c O t h e r s d t , number of encounters at a distant, in-system ED (outside any focal ED–UCC pair), and 2) N u m E n c H i g h s d t , number of immediate and emergent (levels 1 and 2) encounters at the focal ED. These controls absorb variation in ED demand due to system-wide congestion or shocks not captured by our fixed effects, isolating our treatment effect from concurrent system-wide demand fluctuations. Additionally, because the actual calendar dates are de-identified, we cannot directly control for calendar-based patterns such as weekdays or weekends. Fluctuations in encounters at a distant ED or among the high-acuity patients likely reflect these temporal patterns and therefore serve as a proxy for calendar-based seasonality in the absence of calendar date identifiers. We present the estimates from our main specification model (1) in the next section. Our main findings are also robust to excluding X s d t entirely (see EC.4.4.1).

To model the number of arrivals in 15-minute intervals, we employ a count model and obtain a Poisson quasi-maximum likelihood estimator (QMLE). Using Poisson regression to model counts is common in the operations literature (Li et al., 2022). Our coefficients of interest are β 1 , β 2 , which capture the effect of UCC operations on ED demand, with separate coefficients for collocated and noncollocated UCCs. Specifically, exp ( β 1 ) − 1 consistently estimates a scale-invariant average proportional treatment effect of the collocated UCC, E [ y ( k ) | c o l U C C = 1 , P o s t = 1 ] − E [ y ( 0 ) | c o l U C C = 1 , P o s t = 1 ] E [ y ( 0 ) | c o l U C C = 1 , P o s t = 1 ] ; exp ( β 2 ) − 1 is the analogous estimate for non-collocated UCCs. Both estimates have a natural percentage interpretation—the percentage change in the average outcome at the focal ED on treated dates during UCC operating hours.

Conditional on the fixed effects and covariates described above, our identifying assumption, as discussed in Section 4.1, is that there are no shocks that coincide with dates of UCC openings (treated versus control dates) and alter the within-day distribution of ED arrivals, in particular, the arrival pattern over operating versus non-operating hours. This is an analogue of the parallel-trends assumption in canonical DiD designs, which assumes that there are no shocks coinciding with treatment timing that differentially affect treated and control units. Although not directly testable, we conduct extensive plausibility checks of our identifying assumption, including pre-trends tests on non-operating hours in Section 6.1, relaxing the strict requirement of a stable within-day arrival pattern absent UCC operations by implementing the generalized synthetic control method (Xu, 2017) in Section 6.2, and addressing alternative demand- and supply-side shifts that might alter within-day arrival patterns and confound our estimate in Section 6.3.

5.1. The Impact of UCCs on ED Demand

We begin by investigating percentage changes in ED arrivals for patients with acuity levels 3 to 5 during UCC operating hours. The results are reported in Table 2.

For overall ED patient arrivals N u m E n c , including both initial visits and same-day reroutes from the newly opened UCC, we observe a 3.0% (column (2), β = -0.030, p = 0.055) reduction in nonurgent to urgent encounters at the ED during the operating hours of the newly opened collocated UCC. This effect is negative and significant. When we exclude the same-day reroutes from the new UCC and focus only on the initial ED encounters, N u m I n i E n c , the reduction becomes more pronounced—a 3.9% decrease (column (1), β = -0.039, p = 0.014) in ED encounters during the UCC’s operating hours. These results suggest that a conveniently collocated UCC effectively switches less severe patients away from choosing the ED as their first point of care and therefore mitigates ED overuse. Even after accounting for rerouted cases, the substitution effect remains statistically significant. To put these effect magnitudes into context, prior research finds that an 80% increase in ED attendance fees in Hong Kong SAR reduces less urgent ED visits by 8.9% (Lee et al., 2025)–a somewhat larger reduction than we observe, but one that also raises the cost of access. A study by Enard and Ganelin (2013) estimates that roughly 13% - 27% of ED visits in the U.S. could be managed in primary care or UCC settings. The reduction we observe from the opening of a collocated UCC captures a meaningful share of this potentially divertible volume.

In contrast, for the non-collocated UCC, we find no significant decrease in total ED arrivals; in fact, the coefficient for N u m E n c , β = 0.022, is positive and significant (column (2), p = 0.085), suggesting an increase in ED volume during UCC operating hours. This increase appears to be driven by same-day rerouted patients: when we exclude these cases and focus only on initial encounters, we observe no significant impact on ED arrivals during UCC operating hours (column (1), β = 0.013, p = 0.300). These findings suggest that when the UCC is non-collocated, there is no strong evidence of switching from ED to UCC. However, some patients are directed from the non-collocated UCC back to ED, resulting in a marginal increase in total ED demand during UCC operating hours.

Our goal is to assess whether collocated and non-collocated UCCs have significantly different impacts on ED demand. To this end, we compare the coefficients on the interaction terms c o l U C C × U C C O p e r a t e and n o n c o l U C C × U C C O p e r a t e . We find that the difference in coefficients is -0.052 for total ED encounters ( N u m E n c , p = 0.010) and -0.052 for initial encounters ( N u m I n i E n c , p = 0.011), indicating a significantly larger substitution effect for collocated UCCs. Such heterogeneity is robust to alternative sets of covariates, time binning, and functional forms (see EC.4.4).

The heterogeneous impact of UCCs on ED demand remains robust when examined by distance. When the UCC is located 0.7 miles away from the ED, we observe an increase in total ED arrivals and no effect in the initial ED arrivals (see coefficients of D ( 0.7 ) U C C × U C C O p e r a t e in columns (4) and (3) of Table 2). When the distance increases to 5.7 miles, there is no significant effect on either total or initial ED arrivals (see coefficients of D ( 5.7 ) U C C × U C C O p e r a t e in columns (4) and (3) of Table 2). These patterns stand in contrast to the reduction in ED volume observed during the operating hours of a collocated UCC. To further validate our findings, we also estimate ordinary least squares (OLS) models (columns (5) and (6) of Table 2), replacing the log link in equation (1) with an identity link. The results are consistent with those from the Poisson regression specification.

ED Encounters by Acuity Levels. We further examine the impact of collocated and non-collocated UCCs on ED arrivals by acuity level to decompose the observed aggregate effects and understand which patient types are shifting between facilities. Recall that acuity levels 3 to 5 represent urgent, less urgent, and nonurgent encounters, respectively. We report the Poisson regression estimates in Table 3.

OPEN IN VIEWER

Table 3.

Difference-in-differences estimates: Impact of UCC operations on ED arrivals (by acuity).

	NumIniEnc			NumEnc
	(1)	(2)	(3)	(4)	(5)	(6)
	Urgent	Less Urgent	Nonurgent	Urgent	Less Urgent	Nonurgent
c o l U C C × U C C O p e r a t e	−0.031 †	0.034	−0.617 * * *	−0.021	0.039	−0.616 * * *
	(0.019)	(0.027)	(0.081)	(0.018)	(0.027)	(0.081)
n o n c o l U C C × U C C O p e r a t e	0.036 *	−0.035	−0.032	0.048 * *	−0.032	−0.032
	(0.016)	(0.026)	(0.076)	(0.017)	(0.026)	(0.076)
Observations	204,192	204,192	204,192	204,192	204,192	204,192
Fixed effects: s d , s t	Yes	Yes	Yes	Yes	Yes	Yes

We also include controls X s d t in all columns. Standard errors are clustered at the sample-hour. ***: p < 0.001 , **: p < 0.01 , *: p < 0.05 , † : p < 0.1 .

Table 3 column (3) shows that the opening of a collocated UCC leads to 46.1% fewer nonurgent encounters per 15-minute interval during UCC operating hours. The effect is consistent for urgent encounters as well. Table 3 column (1) shows a 3.1% reduction per 15-minute interval. These effects correspond to approximately 2 fewer urgent and 1 fewer nonurgent patient per day during UCC operating hours.

To contextualize the impact, we conduct a back-of-the-envelope annual ED cost savings analysis. The observed reductions imply roughly 3 fewer patients would choose ED as their initial care facility each day during UCC operating hours. ⁴ At an average cost of $2,256 per ED visit and an average cost of $220 per UCC visit among lower-acuity cases (Burke, 2026), this yields approximately $2.2 million in annual cost savings. When same-day rerouted visits are included, the results remain directionally consistent, though attenuated. For non-collocated UCC, we observe statistically insignificant changes in less urgent and nonurgent cases (columns (2) and (3) of Table 3), but a significant and positive coefficient for urgent cases (column (1) Table 3). The urgent cases increase by 3.7% after a non-collocated UCC opening during UCC operating hours. Overall, collocated UCCs demonstrate greater success in reducing ED volume compared to non-collocated UCCs. Importantly, however, the absence of ED volume reduction should not be interpreted as evidence that non-collocated UCCs have limited value. The positive coefficients for urgent encounters are consistent with the possibility that non-collocated UCCs play a different role. They may expand access by bringing patients into the healthcare system who would otherwise have forgone care, some of whom may subsequently require ED follow-up.

We next examine the mechanisms that explain why collocated UCCs successfully facilitate switching encounters away from the ED, whereas non-collocated UCCs do not.

Sampling the ED Queue with Collocated UCC. Collocating a UCC with an ED entails negligible transportation cost between the two facilities, making it feasible for patients to reconsider their choice of care facility after arriving at the ED. This setup creates the possibility for patients to sample the ED queue before deciding whether to remain or switch to the collocated UCC. In particular, nonurgent to urgent encounters may opt for the UCC if they observe substantial congestion in the ED waiting area, given the minimal transportation cost of moving between the two facilities and the comparable service capability they provide for less severe encounters. In simple terms, patients switch from ED to UCC only when the ED waiting room looks crowded and UCC is next door.

Collocation Setup Enables Queue Sampling. Formally, we test the queue-sampling mechanism which may lead to a reduction in ED encounters by estimating whether UCC arrival volume increases in response to ED congestion during hours when the collocated UCC is operational. We operationalize the real-time ED congestion as the number of patients waiting in the ED waiting area at time t . We focus on time t during UCC operating hours ( t ∈ { t | U C C O p e r a t e t = 1 } ) on dates after UCC opening ( d ∈ { d | c o l U C C s d = 1 } ) in the sample that covers the collocated UCC opening ( s = 3 ). We estimate the following Poisson model:

E [ U C C A r r i v a l s d t ∣ M s d t ] = exp ( β q E D W a i t R o o m C e n s u s s d t − 1 + m s d t ′ φ q + γ t + λ s d ) .

(2)

Our variable m s d t controls for common demand shocks and autocorrelation, and includes: (a) total hourly number of patients arriving to the ED–UCC pair, D e m a n d s d τ , where τ is the hour containing interval t ; and (b) lagged number of UCC arrivals, U C C A r r i v a l s d t − 1 . We also include time-of-day fixed effects γ t and date fixed effects λ s d as defined in equation (1). This specification addresses simultaneity concerns. D e m a n d s d τ controls for the local demand shocks such as flu outbreaks or tourist seasons in the ED–UCC pair. Aggregating to the hour interval captures the demand trend fairly and also avoids high collinearity. In addition, our fixed effects absorb demand shocks varying across dates, and varying across time of day. We use lagged waiting room census, E D W a i t R o o m C e n s u s s d t − 1 as the key regressor to mitigate simultaneity and reverse causality concerns—UCC arrivals in time interval t do not affect ED arrivals that have already occurred by interval t − 1 , and therefore do not affect the E D W a i t R o o m C e n s u s measured in interval t − 1 . Lastly, we control for serial correlation in the outcome by including its lagged value U C C A r r i v a l s d t − 1 following literature (Bai et al., 2020).

The first column in Table 4 reports the Poisson regression results. We find that the coefficient of lagged EDWaitRoomCensus is positive and significant ( β q = 0.010 , p < 0.001 ). That is, the more patients observed in the ED waiting room, the more encounters arrive to the collocated UCC. In particular, a 10% increase in the number of patients in the ED waiting room (roughly 1.7 additional patients) leads to a 1.7% increase in the UCC arrival counts. This finding suggests that patients respond to real-time ED congestion by switching to the collocated UCC. When we further examine how UCC census affects ED arrivals, we find no statistically significant effect, indicating that queue-sampling is uni-directional: there is no statistical evidence that patients checking the UCC waiting room census leads to higher ED arrivals.

OPEN IN VIEWER

Table 4.

Poisson QMLE and 2SRI estimates for queue sampling mechanism analysis.

	UCC Arrivals
	(1)	(2)	(3)
	QMLE	2SRI	QMLE
Lagged EDWaitRoomCensus	0.010 * * *	0.062 *	0.003
	(0.003)	(0.030)	(0.003)
Z_resid		− 0.053 †
		(0.029)
Lagged EDWaitRoomCensus × c o l U C C			0.007 †
			(0.004)
Observations	9,154	9,154	43,558
Fixed effects	d , t	d , t	s d , s t

Fixed effects include date and time, d , t . All columns contain controls Demand and Lagged UCC Arrivals. Standard errors are clustered at the sample-hour. ***: p < 0.001 , **: p < 0.01 , *: p < 0.05 , † : p < 0.1 .

ED Wait Room Census Endogeneity. While lagged E D W a i t R o o m C e n s u s helps mitigate simultaneity, unobserved demand shocks may still jointly affect ED congestion and UCC arrivals. For instance, a flu outbreak might increase elderly patients at the ED and younger patients at the UCC, generating positively correlated shocks that bias the estimated effect upward. A snowstorm in the neighborhood may increase ED visits from weather-related injuries while discouraging patients from traveling to the UCC, leading to negatively correlated shocks that bias the estimated effect downward. To further address potential endogeneity of the lagged E D W a i t R o o m C e n s u s , we implement a two-stage residual inclusion (2SRI) IV approach.

We use lagged E D B o a r d i n g C e n s u s as our instrument for lagged E D W a i t R o o m C e n s u s . The instrument, E D B o a r d i n g C e n s u s , is defined analogously to E D W a i t R o o m C e n s u s as the number of patients boarding in the ED treatment area in every 15-minute time interval (see Section 5.2). The time when a patient’s bed is assigned is marked as the beginning of a patient’s boarding time, while the patient’s departure time (from the ED) marks the end of a patient’s boarding time. E D B o a r d i n g C e n s u s has a mean of 5.2 with s.d. 6.5 patients per interval during the operating hours of UCC.

For this IV to be valid, it must satisfy the relevance and exclusion restriction conditions. The E D B o a r d i n g C e n s u s variable meets the relevance condition as prior research has documented that increases in the number of boarded patients lead to higher waiting room congestion in the ED (Smalley et al., 2020). Our first-stage regression (3) further confirms this relevance. The coefficient of E D B o a r d i n g C e n s u s IV is positive and statistically significant ( β f 1 = 0.18, p = 0.003), indicating that more boarded patients are associated with more patients in the ED waiting area, consistent with prior findings. Furthermore, the first stage Cragg-Donald F-statistic is 73.7, above the rule of thumb threshold 10 (Stock and Yogo, 2005), thereby rejecting the null hypothesis of a weak instrument.

For the exclusion restriction to hold, the instrument must be uncorrelated with unobserved drivers of UCC arrivals conditional on our controls and fixed effects, and it must affect UCC arrivals only through its impact on ED waiting-room census. We next address the primary channels through which this condition could be violated and explain why the exclusion restriction is plausible in our setting.

One concern is that unobserved demand shocks may simultaneously drive both ED boarding counts and UCC arrivals, thereby violating the exogeneity of our instrument. Our specification mitigates this concern at multiple time levels. First, we include date fixed effects, which absorb all date-level demand fluctuations, including flu outbreaks, weather shocks, and seasonal patterns that may simultaneously influence ED and UCC arrivals. Consequently, any remaining unobserved demand variation must occur within a given day. To address such within-day shocks, we further control for total hourly arrivals to the ED–UCC pair and the time-of-day fixed effects, which capture short-run fluctuations in aggregate demand.

At still finer time intervals, events might cause sudden, localized demand shocks that simultaneously increase ED boarding and UCC arrivals. The lag structure of our instrument mitigates this concern. Specifically, ED boarding census at time t − 1 counts patients who have already been triaged, evaluated, and admitted, and whose boarding status typically accumulates over several preceding hours. A sudden demand shock at time t may immediately affect UCC arrivals at time t , but it cannot retroactively alter boarding levels that accumulated prior to t . Moreover, prior literature shows that ED boarding is primarily driven by inpatient bed capacity constraints (Song et al., 2015), a supply-side factor that is unlikely to be systematically correlated with short-run demand fluctuations among lower-acuity patients eligible for UCC treatment.

Finally, the ED waiting area and the boarding area are physically separated in our setting, and arriving patients in the waiting area do not observe the number of boarded patients. Therefore, substitution to the UCC is unlikely to be directly influenced by boarding census. These features, namely, controls and fixed effects, the lag structure, and the physical separation between the waiting and boarding areas, collectively support the plausibility of the exclusion restriction in our setting. We acknowledge, however, that the IV estimate still rests on the remaining identifying assumption that, conditional on date fixed effects, hourly demand controls, and time-of-day fixed effects, lagged ED boarding census affects current UCC arrivals only through lagged ED waiting-room census.

Using this IV, we estimate a two-stage residual inclusion model. The 2SRI approach is preferred over the traditional two-stage least squares (2SLS) when the second-stage model is nonlinear, as it produces consistent estimates (Terza et al., 2008). This method suits our context, as our second-stage model is a Poisson regression for a count outcome. We write the first stage regression in equation (3), and the second stage regression in equation (4).

E D W a i t R o o m C e n s u s s d t − 1 = β f 1 E D B o a r d i n g C e n s u s s d t − 1 + m s d t ′ φ q + γ t − 1 + λ s d + ε s d t − 1 ,

(3)

E [ U C C A r r i v a l s d t ∣ M s d t ] = exp ( β q I V E D W a i t R o o m C e n s u s s d t − 1 + β 2 I V Z _ r e s i d s d t − 1 + m s d t ′ φ q + γ t + λ s d ) ,

(4)

Consistent with the model without an IV, the coefficient for E D W a i t R o o m C e n s u s from the 2SRI model is positive and significant ( β q I V = 0.062 , p = 0.036 , Table 4, column 2). More patients in the ED waiting room leads to more patient arrivals to the collocated UCC.

Nonlinear Impact of Waiting Room Census on ED-collocated UCC pair Queue Sampling. Thus far, we have treated the number of patients in the ED waiting area as a continuous variable. However, if patients switching to the UCC is indeed driven by perceived ED congestion, the response is unlikely to be linear. For example, one additional patient in a nearly empty waiting room may not trigger a change in patients’ care facility choice, whereas one additional patient in an already crowded room may prompt patients to switch.

To test for nonlinear effects, we modify model (2) and replace the continuous measure of E D W a i t R o o m C e n s u s with a factor variable E D W a i t R o o m C e n s u s _ F c t which takes values from Q1 through Q5 to represent the quintiles of E D W a i t R o o m C e n s u s . The cutoffs of the quintiles are 6, 13, 20, 28 patients in the collocated sample. That is, if E D W a i t R o o m C e n s u s ∈ [ 0 , 6 ] , then E D W a i t R o o m C e n s u s _ F c t = Q 1 ; if E D W a i t R o o m C e n s u s ∈ ( 6 , 13 ] , then E D W a i t R o o m C e n s u s _ F c t = Q 2 . The remaining quintiles are defined analogously based on the cutoff values 20 and 28. We thus divide the waiting room census into five bins (quintiles), with the first quintile serving as the reference group.

The coefficients of E D W a i t R o o m C e n s u s _ F c t from estimating the Poisson regression are shown in Figure EC.2. The figure shows that, UCC arrivals indeed respond significantly (compared to the baseline level) to ED congestion only when the ED waiting room is truly “crowded” when the ED waiting room exceeds 28 patients in the last quintile. This pattern suggests a convex relationship: switching to the collocated UCC occurs only when ED crowding reaches visibly high levels.

Limited Queue Sampling When UCC is Non-collocated. A key feature of the collocation setup is the minimal switching cost between the ED and UCC. In contrast, for non-collocated UCCs, the physical separation introduces frictions that may hinder real-time queue sampling. If switching (of facilities) behavior is indeed partially driven by observed ED congestion, we would not expect to see a comparable effect when the UCC is not conveniently collocated.

To this end, we estimate model (2) on all three ED–UCC pairs. In particular, we modify model (2) by interacting E D W a i t R o o m C e n s u s s d , t − 1 with the indicator of whether the UCC is collocated (i.e., I { c o l U C C s d = 1 } ). This interaction specification allows us to compare the responsiveness of UCC arrivals to ED congestion across collocated and non-collocated settings. The results are reported in Table 4, column (3). We find that the coefficient on the interaction term E D W a i t R o o m C e n s u s s d h t − 1 × I { c o l U C C s d = 1 } is positive and significant ( β = 0.007 , p = 0.052 ), while the coefficient of E D W a i t R o o m C e n s u s s d h t − 1 is insignificant ( β = 0.003 , p = 0.338 ). This suggests that there is no significant signal of customers conducting queue sampling when the UCC is non-collocated with the focal ED, whereas there is evidence supporting the existence of patients’ queue sampling behavior when the UCC is collocated with ED.

Our findings carry two important implications for practitioners and policymakers. First, we document that the effectiveness of UCCs in reducing ED demand is not uniform—it depends on physical proximity to the ED. While previous work and public discourse often focus on placing UCCs based on regional population needs or ZIP code-level access gaps, our findings highlight that relative positioning between UCCs and EDs is also important. If the goal is to relieve overuse at a particular ED, placing a UCC collocated with that ED is far more effective than adding a UCC elsewhere in the region. Our findings speak to an ongoing and unsettled debate over whether expanding access to UCCs meaningfully reduces ED demand. Our study offers new insight into when and how UCCs can serve as an effective lever to reduce ED demand during UCC operating hours.

Second, we identify queue sampling—patients observing ED crowding and switching to the collocated UCC in real time—as a plausible mechanism driving the heterogeneous treatment effects by ED–UCC proximity. Importantly, patients switch from ED to UCC only when the ED is visibly congested, precisely when reducing ED demand is most needed, and when the UCC is collocated. An additional analysis shows that census count is a highly significant positive predictor of door-to-provider wait time (see Section EC.4.7), suggesting that patients responding to census congestion are, in effect, responding to a wait time signal. This points to a practical implication: posting real-time wait times in the ED waiting room could lower the cognitive burden of queue sampling and encourage more real-time switching to a collocated UCC. Because queue sampling reflects a general behavioral trade-off between waiting time costs and switching costs among other factors, this mechanism and the resulting substitution across facilities in a collocated setup are likely generalizable beyond our specific healthcare facility.

These insights have clear operational and policy relevance. For hospital managers, collocated UCCs should be viewed not merely as standalone facilities but as flexible capacity extensions that can absorb less severe, delay-sensitive patients during periods of high ED demand. Our findings offer one empirical lens on the ongoing debate over whether UCCs should be collocated with the ED or distributed geographically: we empirically show that when the goal is to relieve demand at a specific ED, proximity matters. Furthermore, the local markets where our ED–UCC pairs are located already had UCCs in operation prior to the new UCC openings. Our estimates therefore reflect the incremental impact of an additional UCC. Notably, the collocated UCC was introduced last and in the setting with the highest baseline UCC density (see Section EC.5), suggesting that any diminishing marginal benefit of UCC access would bias our estimate downward. This makes our estimated reduction conservative: even in an area already served by several UCCs, collocation still proves effective in alleviating ED demand during its operating hours. This setting mirrors today’s healthcare landscape, where new UCCs almost always enter regions with existing facilities. Our findings therefore speak directly to the practical question of whether adding a collocated UCC continues to deliver meaningful relief for ED demand when the UCC operates.

Whereas prior work has emphasized long-term effects of UCC availability on future ED use, our study demonstrates that ED–UCC proximity can further shape real-time patient decisions under congestion. By identifying queue sampling as a facility choice mechanism enabled by collocation, we clarify not just whether UCCs reduce ED demand, but when and why they do, providing actionable guidance for both facility planning and daily ED operations. As the UCC market continues to grow, identifying locations that support seamless, real-time switching of patients from the ED may unlock the full potential of UCCs as tools for ED congestion relief during UCC operating hours.

6.1. Pre-treatment Parallel Trends and Within-day Comparability

No Divergence During Non-operating Hours. As standard in DiD frameworks, we test our identifying assumption first by checking the trends during the pre-treatment period—the arrival patterns of treated and control dates do not diverge during non-operating hours. This provides the first piece of evidence that the control dates provide a plausibly valid counterfactual for the treated dates. Following standard literature, we implement two approaches: a leads-and-lags model and a differential pre-treatment trend test. These tests detect diurnal pattern changes that would manifest in non-operating hours. We show that across both collocated and non-collocated setups, pre-treatment coefficients are largely statistically indistinguishable from zero, with only one marginally significant at the 10% level. Further, we find indistinguishable time trends across the treated and control dates during non-operating hours from the differential pre-treatment trend tests. We relegate the details of both analyses to Section EC.4.1.

Both pieces of evidence suggest that the arrival pattern of treated and control dates has a stable difference during non-operating hours. Particularly, in the collocated sample, leads-and-lags coefficients show that the divergence between treated and control dates emerges specifically from when UCC starts operating (EC.4.1), is sustained during UCC operating hours, and is absent during non-operating hours. This is consistent with UCC operations driving the change in arrival patterns.

Comparability of Operating and Non-operating Hours. Our identifying assumption implicitly requires no structural breaks between operating and non-operating hours, absent UCC operations. Control dates provide one source of assurance: to the extent that time-of-day structural breaks are comparable across treated and control dates, they are differenced out. Pre-trend tests support this comparability. We further provide a second source of assurance: such breaks are unlikely because the underlying patient composition is similar across the two time periods. Section EC.4.6, Table EC.8 shows that the acuity mix across operating and non-operating hours before UCC opening is indeed comparable in both collocated and non-collocated samples, suggesting that an intrinsic break in patient composition, and consequently in underlying demand patterns, at the operating-hour boundary is unlikely absent UCC operations.

6.2. Generalized Synthetic Control: Relaxing the Parallel Trends Assumption

Although our DiD estimator passes the pre-treatment trend tests in Section 6.1, those tests only detect confounders that meaningfully shift non-operating hour arrival patterns. A confounder that gradually alters within-day patterns may not be strong enough to be detected by standard pre-trend tests. Furthermore, these pre-trend tests cannot inform the pattern during operating hours. To address both of these concerns, we implement the generalized synthetic control (GSC) proposed by Xu (2017). We use the GSC framework to flexibly model unobserved confounders that vary across time of day and alter within-day arrival patterns.

In particular, the GSC framework accounts for these confounders through an interactive fixed effects (IFE) model with latent factor components. The IFE model is estimated using only the control dates, recovering the latent factors and their factor loadings, and therefore learning the within-day arrival pattern (trajectory) on control dates. For each treated date, the factor loadings are estimated by projecting the non-operating hour outcomes onto the latent factor space identified from control dates. The counterfactual operating-hour pattern is then imputed using the estimated IFE model together with the estimated date-specific factor loadings. This approach is valid as long as non-operating hours are informative of operating hours on the same date. That is, the within-day arrival pattern is a coherent whole, such that a treated date’s non-operating hour pattern is mostly sufficient to locate it within the latent factor space learned from control dates. Such within-day coherence is plausible given the comparable patient composition across operating and non-operating hours shown in Section 6.1.

Formally, let y d t denote the outcome at time t on treated date d . The GSC estimator for the treatment effect is then calculated as β d t = y d t ( 1 ) − y ^ d t ( 0 ) . It is the difference between the actual demand outcome and its (untreated) demand counterfactual estimated from the IFE model. Our main quantity of interest—the average treatment effect on the treated at time t , A T T ^ t —can then be obtained by averaging across all heterogeneous treatment effects β d t for all treated dates d . We provide further details of the GSC analysis in Section EC.4.2. Table EC.1 shows that the GSC estimates are consistent with those from our DiD specification: reductions in ED volume during UCC operating hours occur only in the presence of a collocated UCC. In contrast, we find no significant reduction effect of non-collocated UCCs on ED demand.

6.3. Examining Alternative Mechanisms

So far, Section 6.1 tests for confounders that would manifest in non-operating hours; Section 6.2 further addresses confounders whose effects on within-day arrival patterns can be captured by the latent factors in control dates. These checks progressively narrow the space of viable confounders. Combined with the controls in model (1), any remaining confounding variation would need to be an ED-specific, volume-independent shock that coincides with dates after UCC opening and the specific 8 a.m. operating-hour boundary. In this section, we examine specific mechanisms that could plausibly generate such a variation.

Extended Effective UCC Operating Hours. One concern is that the opening of a UCC may induce changes in patient behavior once patients learn about its operating hours. For instance, lower-acuity patients, whose conditions could be treated at either an ED or a UCC and whose symptoms allow some waiting, may delay an early-morning ED visit (e.g., 6 a.m.) until UCC opens at 8  a.m. Such timing substitution, where non-operating-hour ED demand shifts into operating hours and potentially diverts to the UCC, would coincide with UCC opening day and operating hours and alter the diurnal pattern of ED arrivals. This could mean that our operating-hours estimand does not fully reflect UCC’s impact on total daily ED demand. For collocated UCCs, timing substitution could only strengthen our negative coefficient on ED demand, while for non-collocated UCCs, it could bias our null estimate in either direction relative to total daily demand. We assess the extent of this concern through two analyses.

We return to the discussion in Section 4.3. Non-operating-hour ED demand remains highly stable before and after UCC opening, whereas changes occur primarily during operating hours. This pattern suggests that our estimated effects are driven by changes during UCC operating hours, with limited evidence of substantial substitution from non-operating hours. To further address this concern, we redefine our indicator U C C O p e r a t e t to equal 1 if t falls within an extended UCC operating window (6 a.m.–10 p.m.), and 0 otherwise. The extended window explicitly incorporates the hours during which timing substitution is more likely to occur. We present the results of this analysis in Section EC.4.3, Table EC.3. Our main findings remain robust in both magnitude and statistical significance. This alleviates concerns that our estimates are biased by changes in patient arrival timing following UCC opening. Our estimates also remain stable across alternative extended UCC operating hours (see Figure EC.5), further suggesting limited timing substitution behavior. The stability of non-operating-hour demand and the robustness and stability of our estimates under extended operating hours suggest that our operating-hour estimand is directionally consistent with how collocated and non-collocated UCC openings may affect daily ED demand: a reduction following collocated UCC openings but not non-collocated ones.

Staffing, Workflow, and Triage. Another concern is that staffing, triage, or workflow changes coinciding with UCC opening could affect diurnal ED arrival patterns and thereby confound our estimates. We argue that this is unlikely for several reasons. Our outcome is ED arrivals, which reflect patients’ decisions made prior to any interaction with ED internal processes. Such changes operate downstream of arrival and cannot directly confound the arrival outcome unless they feed back into patients’ visit-timing decisions—unlikely in our setting since our ED does not publicly post wait times, leaving patients without the real-time information needed to respond to supply-side changes. Furthermore, while triage involves some human discretion in assigning exact acuity, the distinction between levels 3–5, our focus, and levels 1–2 is relatively coarse based on ESI algorithm and unlikely to shift meaningfully around UCC opening.

We can further assess this concern through door-to-provider times (the time from arriving at the ED waiting room to the time leaving the ED waiting room), using them as a proxy for staffing and workflow conditions. Using the same specification as our main analysis with average door-to-provider time as the outcome, we find that door-to-provider times decrease during UCC operating hours at the collocated ED ( β = − 17.5 , p < 0.001), consistent with our main findings of fewer arrivals. If an independent staffing or workflow change were making the ED faster during operating hours, this would be expected to attract additional patients—the opposite of our findings. Conversely, if these changes slowed the ED, we would observe longer door-to-provider times, not shorter. The joint pattern of fewer arrivals and faster door-to-provider times is more internally consistent with patients switching to the UCC and inconsistent with a staffing or workflow confounder in either direction. Additionally, nurses and other staff at the collocated ED are unionized, making short-term staffing adjustments difficult to implement around UCC opening. Further reassuringly, we do not find a significant change in door-to-provider time at the non-collocated ED ( β = − 4.1, p = 0.69), consistent with our null effect in non-collocated ED arrival outcomes.

Acuity Mix and Patient Behavior. Apart from supply-side staffing changes, demand-side changes in acuity mix and patient behavior across day and night may be another concern. We first provide direct evidence that patient mix is stable over time. Section EC.4.8, Table EC.9 shows that the acuity mix during non-operating hours is stable before and after UCC opening. This is not surprising given that our study spans a 2-year window around UCC opening—it is unlikely to see disruptive changes in acuity mix or patient behavior over a relatively short period. Note that our data does not cover the COVID-19 period.

Given the stability during non-operating hours, any confounding variation in patient mix or patient behavior that differentially affects day versus night would need to coincide precisely with both dates after UCC opening and the specific 8 a.m. operating-hour boundary. Such a confounding event is unlikely in our setting, except for patient behavior shifts directly induced by UCC availability—patients’ timing substitution around the operating-hour boundaries, which we addressed above. Finally, we note that this assumption is standard in difference-in-differences designs. In a canonical cross-ED DiD, patient mix stability must hold along the calendar-time dimension and can only be assessed through pre-treatment trends. Our design in fact imposes a stricter requirement: any confounding change must coincide with not only the calendar date of UCC opening but also the specific time-of-day window of UCC operation.