Pathways to reduced emergency department and urgent care center use: Lessons from the comprehensive primary care initiative

Abstract

Objective

To determine the association between a large‐scale, multi‐payer primary care redesign—the Comprehensive Primary Care (CPC) Initiative—on outpatient emergency department (ED) and urgent care center (UCC) use and to identify the types of visits that drive the overall trends observed.

Data Sources

Medicare claims data capturing characteristics and outcomes of 565 674 Medicare fee‐for‐service (FFS) beneficiaries attributed to 497 CPC practices and 1 165 284 beneficiaries attributed to 908 comparison practices.

Study Design

We used an adjusted difference‐in‐differences framework to test the association between CPC and beneficiaries’ ED and UCC use from October 2012 through December 2016. Regression models controlled for baseline practice and patient characteristics and practice‐level clustering of standard errors. Our key outcomes were all‐cause and primary care substitutable (PC substitutable) outpatient ED and UCC visits, and potentially primary care preventable (PPC preventable) ED visits, categorized by the New York University Emergency Department Algorithm. We used a propensity score‐matched comparison group of practices that were similar to CPC practices before CPC on multiple dimensions. Both groups of practices had similar growth in ED and UCC visits in the two‐year period before CPC.

Principal Findings

Comprehensive Primary Care practices had 2% (P = .06) lower growth in all‐cause ED visits than comparison practices. They had 3% (P = .02) lower growth in PC substitutable ED visits, driven by lower growth in weekday PC substitutable visits (4%, P = .002). There was 3% (P = .04) lower growth in PPC preventable ED visits with no weekday/nonweekday differential. As expected, our falsification test showed no difference in ED visits for injuries. UCC visits had 9% lower growth for both all‐cause (P = .08) and PC substitutable visits (P = .07).

Conclusions

Our results suggest that greater access to the practice and more effective primary care both contributed to the lower growth in ED and UCC visits during the initiative.

What is Already Known on This Topic

• Prior research suggests that primary care delivery has the potential to shape ED and UCC use.

• CPC was associated with 2% lower growth in all‐cause ED visits.

• It is unknown which types of ED use drove this result, and the association between CPC and UCC visits has not been studied.

What This Study Adds

• The CPC Initiative reduced growth in outpatient ED and UCC visits, driven by slower growth in weekday PC substitutable ED and UCC visits and by PPC preventable ED visits.

• New approaches to payment for and the delivery of primary care in large‐scale primary care redesign initiatives can lead to meaningful reductions in ED and UCC visits.

1. INTRODUCTION

An estimated 13%‐27% of emergency department (ED) visits in the United States could be managed in physician offices, saving approximately $4.4 billion annually. ¹ With a surge in the number of urgent care centers (UCCs) in recent years (44% growth nationally between 2013 and 2018), ² many patients are increasingly turning to UCCs for the treatment of low‐acuity conditions, even overtaking the number of low‐acuity visits at EDs in some populations. ³ Using EDs and UCCs instead of primary care offices for nonemergent conditions could lead to fragmentation of care, inefficient resource use, and higher spending. ⁴ Further, a lack of longitudinal care management from primary care providers could lead to symptoms and exacerbations of chronic conditions requiring an ED visit that may have otherwise been prevented with effective primary care. Consequently, improved access to high‐quality primary care could reduce primary care substitutable (PC substitutable) visits—that is, those visits to EDs and UCCs for conditions that can be treated in a primary care setting—and could also reduce potentially primary care preventable (PPC preventable) visits—that is, those visits to EDs for conditions that require ED care but whose symptoms and exacerbations can be mitigated with effective primary care. ¹ , ³

In October 2012, the Centers for Medicare & Medicaid Services (CMS) launched the four‐year Comprehensive Primary Care (CPC) Initiative in collaboration with 39 private and public payers. CPC tested a new approach to payment for and delivery of primary care in nearly 500 diverse primary care practices in seven regions, with the goal of improving quality and reducing costs. Practices were required to meet annual milestones to improve: access to and continuity of care, planned care for preventive and chronic needs, risk‐stratified care management, engagement of patients and their caregivers, and coordination of care with patients’ other providers. ⁵ CPC supported practices with enhanced payments accounting for a median of 10%‐20% of practice revenue each program year, data feedback, and learning support. ⁶

We hypothesized CPC would reduce preventable ED and UCC use because previous research showed CPC strengthened primary care and was associated with lower growth in all‐cause outpatient ED visits by 2%, relative to comparison practices. ⁶ , ⁷ The biggest improvements were in risk‐stratified care management, expanded access to care, and activities related to data‐driven quality improvements. CPC practices also reported reducing wait times for appointments; adding staff to the practice (eg, care managers) who could serve as additional points of contact to the practice; improving after‐hours access to clinicians via email, telephone, or in person; and offering same‐ or next‐day appointments. Practices made smaller improvements in continuity of care, coordination of care (including transitional care after ED visits or hospitalizations), and planned care for chronic conditions.

This study paints a fuller picture of the association between CPC and ED and UCC visits among Medicare fee‐for‐service (FFS) beneficiaries by examining different types of these visits to identify the drivers of the lower growth. We expect the greatest changes to occur with those types of visits most amenable to improved primary care—PC substitutable visits and PPC preventable visits. We stratified visits by day of week to study whether changes were concentrated during regular business days (via increased same‐day/next‐day appointments and more responsiveness during weekdays by the care team) versus weekends and holidays. To our knowledge, this is the first study to examine whether a large‐scale primary care redesign is associated with a weekday or non‐weekday differential in ED visits. Unlike previous studies of CPC, we also examined UCCs, which can substitute for both primary care and ED use—with their surge in recent years, they are an increasingly important setting to study.

2. METHODS

2.1. Study design

We compared Medicare FFS beneficiaries attributed to the 497 practices participating at the end of CPC’s first quarter to those attributed to 908 comparison practices. We attributed beneficiaries to the practice that had delivered the largest share of their primary care visits over the prior 2 years, following the attribution approach taken by CMS for CPC (see Supplement 1). Using an intent‐to‐treat design, beneficiaries remained in the analysis once attributed.

Comparison practices were selected using propensity score matching of practice, market, and beneficiary characteristics. ⁶ , ⁷ , ⁸ , ⁹ By design, comparison and CPC practices had very similar observable characteristics before CPC, such as practice size, electronic health record use, attributed Medicare beneficiary demographics, and Medicare spending and service use (see appendix 4 of Peikes et al [2018a] ⁷ ).

2.2. Data and sample

We analyzed Medicare claims files from the CMS Virtual Research Data Center for 565 674 Medicare FFS beneficiaries attributed to the 497 CPC practices and 1 165 284 beneficiaries attributed to 908 comparison practices during CPC.

2.3. Outcome measures

In our claims‐based analysis, we studied two settings: (a) outpatient ED visits (ie, visits that did not result in hospital admissions), including observation stays; and (b) visits to UCCs (these typically are freestanding physicians’ offices with extended hours focused on delivering walk‐in care for acute conditions, excluding retail health clinics). In Supplement 3, we describe the identification of ED and UCC visits in the claims data and construction of the outcome measures.

Our outcome measures were the number of ED or UCC visits per 1000 beneficiaries during a 12‐month observation window. First, we analyzed the total number of all‐cause visits in each setting. Second, we examined the number of PC substitutable visits in each setting—those visits that can be treated with resources available in a primary care setting if timely access to a primary care practice was available. Third, we studied the number of PPC preventable visits in the ED setting—those emergent visits that required the resources of an ED, but that were potentially preventable with effective primary care. We constructed each measure for all days of the week combined and then separately for weekdays and nonweekdays.

We categorized ED and UCC visits using the New York University Emergency Department Algorithm (NYU EDA), the most widely used claims‐based tool for retrospectively assessing the probability that ED visits are nonemergent, preventable, or optimally treated in an ED. ¹⁰ , ¹¹ , ¹² , ¹³ The algorithm uses the patient's principal diagnosis code to assign each ED visit probabilities of falling into four categories: (a) nonemergent; (b) emergent, primary care treatable; (c) emergent (ED care needed) but preventable/avoidable; and (d) emergent (ED care needed), not preventable. For example, a principal diagnosis of “headache” has a 0.78 probability of being nonemergent; a 0.09 probability of being emergent, primary care treatable; and a 0.13 probability of being emergent (ED care needed), not preventable. In Table 1, we provide an overview of the NYU EDA and categorization of example diagnoses.

TABLE 1.

Overview of visit categorization using the New York University Emergency department algorithm

Visit Categorization using the NYU EDA Categories	Examples of NYU EDA visit categorization for different principal diagnosis codes
	ICD‐9 code of 784.0 (headache)	ICD‐9 code of 466.0 (acute bronchitis)	ICD‐9 code of 486 (pneumonia)	ICD‐9 code of 427.5 (cardiac arrest)	ICD‐9 code of 959.01 (head injury)
Primary care substitutable visits (1 + 2)	0.87	0.82	0.33	0	0
1. Non‐emergent: Immediate medical care not required within 12 h	0.78	0	0.09	0	0
2. Emergent/primary care treatable: Care required within 12 h, but care could have been provided in a primary care setting	0.09	0.82	0.24	0	0
Potentially primary care preventable (3 + 4)	0.13	0.18	0.67	1	0
3. Emergent/preventable: Emergency care was required, but visit could possibly have been prevented with timely and effective primary care	0	0.18	0.67	0	0
4. Emergent/not preventable: Emergency care was required, and primary care treatment could not have prevented the visit	0.13	0	0	1	0
Other: Injury, alcohol related, substance use related, mental health related, unclassified	0	0	0	0	1
Total sum of probability weights	1	1	1	1	1

The NYU EDA uses the primary ICD‐9 diagnosis code to assign each ED visit a probability of falling into one of four categories: (1) non‐emergent; (2) emergent, primary care treatable; (3) emergent (ED care needed) but preventable; and (4) emergent (ED care needed), not preventable. The NYU EDA separately assigns principal diagnosis codes to categories for: Injury (eg, open wounds, sprains); mental health; alcohol use; or substance use. These visits are excluded from the four categories above. Further, the NYU EDA does not classify some uncommon and new diagnosis codes. Since few diagnostic categories can be identified as emergent and/or preventable with certainty, the algorithm uses the patient's principal diagnosis code to assign probabilities for an ED visit being in each of Categories 1 through 4. The probabilities range from 0 to 1 in each category and sum to 1 across the four categories. Diagnoses in the excluded categories are assigned a probability weight of 0 for each of Categories 1 through 4, but a 1 in the relevant excluded category (eg, head injury gets a weight of 1 for the injury category; schizophrenia gets a weight of 1 for the mental health category). We used the “patch” version developed by Johnston et al (2017) ¹⁴ which takes into account all new ICD‐9 codes introduced since the algorithm's creation and also provides weights for ICD‐10 codes.

Abbreviation: ED, emergency department.

Source: Information comes from the New York University Emergency Department Algorithm (NYU EDA) derived by Billings, Parikh, and Mijanovich (2000). ¹⁰

We defined PC substitutable visits as the first two categories (categories 1 and 2) and PPC preventable ED visits as the last two categories (categories 3 and 4). The PPC preventable measure focuses exclusively on visits requiring ED resources, although some PC substitutable visits might have been prevented upstream through better primary care. To construct our measures, we summed the probability weights in the relevant NYU EDA categories across visits in the program year. In many cases, an individual diagnosis code maps to multiple categories (rather than a single category). In the example above, “headache” is not perfectly PC substitutable (ie, it has a probability of being PC substitutable, which is 0.87 in this case). We used the patch by Johnston et al ¹⁴ to assign NYU EDA probability weights to account for changes in diagnosis codes since the original algorithm was developed (such as ICD‐10 codes).

We expected improved access to effective primary care would be associated with reductions in the rate of PC substitutable visits. ¹⁰ We also expected there might be some reductions in PPC preventable rates from CPC: improved primary care from stronger care management, care coordination, and patient education could lead to fewer exacerbations of the patient's underlying illness and, consequently, reduce the incidence of emergent visits that require ED resources. While not all diagnoses included in PPC preventable ED visits are likely to be affected by CPC, we expected some types of emergent visits would be prevented and that the overall incidence of PPC preventable ED visits would decline over time.

As a falsification test, we examined ED visits for injuries because we expected strengthened primary care would have minimal effects on the most common types of ED visits for injuries observed in our dataset (head injuries, open wounds, and chest contusions). As such, a statistically significant change in the rate of ED injury visits could suggest underlying differences in the comparison group that are correlated with outcome measures (which would lead to biased estimates).

2.4. Statistical analysis

We used an adjusted difference‐in‐differences (DD) framework and compared changes in the rate of Medicare FFS beneficiary ED and UCC visits in CPC practices between the 12 months before CPC (baseline) and the combined four years of CPC to changes in the rate of visits in the comparison practices over the same period. The regression models controlled for characteristics of the beneficiaries, practices, and practices’ markets to net out prior observable differences between CPC and comparison beneficiaries not fully eliminated by matching (see Table S1 in Supplement 2 for a full list of control variables). All outcomes were annualized to account for partial year FFS eligibility of some beneficiaries. We applied weights to observations in the regression models to give more weight to those observed in Medicare FFS data for more of the period than those observed for less of the period. All regression models accounted for nonindependence across observations within the same practice using standard errors clustered at the practice level. All P values were 2‐sided and considered statistically significant at P < .1. There was no formal statistical adjustment of P values for having multiple comparisons; therefore, these results should be considered exploratory. Data analyses were performed using STATA version 15.1 (StataCorp).

3. RESULTS

3.1. ED and UCC use over time

The ED and UCC rates increased over our study period for beneficiaries in both CPC and comparison practices, which could reflect the aging cohort and/or secular trends (Figure 1). Although beneficiaries in comparison practices had higher rates of ED and UCC use than those in CPC practices, we saw similar growth in the period before CPC, which supports the credibility of the “parallel trends” assumption of the DD model.

FIGURE 1.

Regression‐adjusted Emergency department and urgent care center rates over time. Notes: Predicted means are regression adjusted for baseline patient characteristics (including Hierarchical Condition Category scores, which are a measure of risk for subsequent expenditures) and baseline characteristics of the practice and the practice's market. Non‐weekday refers to weekends and major holidays. Primary care substitutable visits comprise of the “non‐emergent” and “emergent, primary care treatable” categories of the New York University Emergency Department Algorithm. Abbreviations: CPC, Comprehensive Primary Care Initiative; ED, emergency department; UCC, urgent care center. Source: Authors’ analysis of Medicare claims data for October 2010‐December 2016[Color figure can be viewed at wileyonlinelibrary.com]

3.2. Composition of ED and UCC visits for beneficiaries attributed to CPC practices in the baseline period

PC substitutable visits made up 42% of ED visits (Table 2). To understand which types of diagnoses are associated with each category in each setting, we list the five most common diagnoses that had probability weights over 0.5 in the category (so more likely than not, the visit was emergent/nonemergent and/or preventable). Unspecified chest pain (typically noncardiac), dizziness, and headaches made up the most common diagnoses that were associated with PC substitutable visits. PPC preventable visits accounted for over 25% of all ED visits, with chest pain, syncope and collapse (fainting), and pneumonia being the most common diagnoses. Injuries accounted for 21% of all ED visits, with head injuries, open wounds, and contusions being the most common diagnoses. The breakdown across NYU categories at baseline for the CPC research sample using Medicare claims is similar to the 2012 national rates using all payers’ claims. ¹⁴

TABLE 2.

Classification of Emergency Department and urgent care center visits for Medicare fee‐for‐service beneficiaries in comprehensive primary care initiative practices in the baseline period (and five most common principal diagnoses)

	Percentage across all visits, unless noted
ED visits (NYU EDA categories)
Primary care substitutable ED visits (1 + 2)	41.7%
1. Non‐emergent: Dizziness and giddiness; Headache; Unspecified essential hypertension; Other malaise and fatigue; Lumbago	19.0%
2. Emergent/primary care treatable: Chest pain, not elsewhere classified (typically non‐cardiac); Abdominal pain, unspecified site; Acute bronchitis; Epistaxis; Other respiratory abnormalities	22.7%
Potentially primary care preventable ED visits (3 + 4)	25.4%
3. Emergent/preventable: Pneumonia, organism unspecified; Dehydration; Congestive heart failure, unspecified; Diabetes with other specified manifestations, type II or unspecified type, not stated as uncontrolled; Unspecified asthma with (acute) exacerbation	6.5%
4. Emergent/not preventable: Chest pain, unspecified; Syncope and collapse; Unspecified atrial fibrillation; Altered mental status, unspecified; Palpitations	18.9%
Injury: Head injury, unspecified; Contusion of face, scalp, and neck except eye(s); Open wound of finger; Open wound of scalp; Contusion of chest wall	21.1%
Alcohol, Drug, Psych	3.6%
Unclassified	8.2%
All ED visits	100.0%
Number of ED visits	193 879
UCC visits
Primary care substitutable UCC visits (1 + 2)	58.3%
1. Non‐emergent: Acute sinusitis, unspecified; Low back pain; Acute pharyngitis; Cough; Impacted cerumen	32.4%
2. Emergent/primary care treatable: Acute bronchitis; Acute upper respiratory infections of unspecified site; Cellulitis and abscess of leg, except foot; Bronchitis, not specified as acute or chronic; Abdominal pain, unspecified site	25.9%
Potentially primary care preventable UCC visits (3 + 4)	14.7%
3. Emergent/preventable: Pneumonia, organism unspecified; Asthma, unspecified type, unspecified; Asthma, unspecified type, with (acute) exacerbation; Congestive heart failure, unspecified; Dehydration	7.7%
4. Emergent/not preventable: Chest pain, unspecified; Shortness of breath; Sciatica; Hematuria, unspecified; Atrial fibrillation	7.0%
Injury: Open wound of finger; Open wound of hand; Open wound of knee, leg (except thigh), and ankle; Open wound of forearm; Contusion of chest wall	18.4%
Alcohol, Drug, Psych	0.6%
Unclassified	8.0%
All UCC visits	100.0%
Number of UCC visits	25 381
Number of beneficiaries	442 709

Baseline is the year before CPC started (October 2011–September 2012). This table shows the percentage of ED and UCC visits during the baseline period that fall into each of the New York University Emergency Department Algorithm (NYU EDA) categories for beneficiaries attributed to CPC practices. Note that a visit can have positive probability weights in multiple NYU EDA categories. The five most common principal diagnoses with a NYU EDA probability weight greater than 0.5 are shown for each category in each setting.

Abbreviations: CPC, Comprehensive Primary Care; ED, emergency department; NYU EDA, New York University Emergency Department Algorithm; UCC, urgent care center.

Source: Authors’ analysis of Medicare claims data for October 2011–September 2012.

As expected, we found that PC substitutable visits made up a larger proportion (58%) of UCC than ED visits (42%; Table 2). Diagnoses related to upper respiratory infections were the most common types of conditions associated with PC substitutable visits to UCCs.

Weekdays account for 5/7 of all days, and we found that roughly 5/7 of PC substitutable ED and UCC visits were on weekdays (68% and 69%, respectively). The 20 most common diagnoses for PC substitutable ED and UCC visits on weekdays and nonweekdays were nearly identical (data not shown). This suggests that the types of lower‐acuity conditions for which beneficiaries are treated in the ED or UCC do not vary substantially across the days of the week.

3.3. Comprehensive Primary Care and PC substitutable ED visits

Regression‐adjusted DD estimates show CPC practices had 3% lower growth in PC substitutable ED visits than comparison practices (P = .02; Table 3). Although the PC substitutable ED regression‐adjusted rates increased for CPC beneficiaries over the course of CPC—from 176 to 194 annual visits per 1000 beneficiaries—growth was lower than for beneficiaries in comparison practices. The year‐by‐year DD estimates for PC substitutable ED visits were larger and more favorable in the last two years of CPC (see Table S3 in Supplement 4).

TABLE 3.

Difference‐in‐differences estimates of the association between the comprehensive primary care initiative and primary care substitutable and potentially primary care preventable emergency department and urgent care center visits

Measure	Baseline		Years 1‐4		DD Estimate (SE)	DD Percentage	90% Confidence Interval	P value
	CPC Practices	Comparison Practices	CPC Practices	Comparison Practices
Emergency department visits (annualized rate per 1000 beneficiaries)
All‐cause ED visits	422	432	495	514	−9.6* (5.2)	−1.90%	(−18, −1)	.065
Primary care substitutable ED visits (all days of week)	176	180	194	204	−5.8** (2.5)	−2.88%	(−10, −2)	.021
Primary care substitutable ED visits (weekday)	119	120	132	139	−5.5*** (1.8)	−3.97%	(−8, −3)	.002
Primary care substitutable ED visits (non‐weekday)	58	60	62	64	−0.3 (1.1)	−0.46%	(−2, 2)	.801
Potentially primary care preventable ED visits (all days of week)	108	109	133	138	−3.5** (1.7)	−2.56%	(−6, −1)	.036
Potentially primary care preventable ED visits (weekday)	74	75	93	97	−2.4* (1.3)	−2.47%	(−5, 0)	.073
Potentially primary care preventable ED visits (non‐weekday)	34	34	40	41	−1.1* (0.6)	−2.77%	(−2, 0)	.074
Urgent care center visits (annualized rate per 1000 beneficiaries)
All‐cause UCC visits	61	67	78	92	−7.6* (4.3)	−8.84%	(−15, −1)	.078
Primary care substitutable UCC visits (all days of week)	36	40	46	55	−4.6* (2.5)	−9.16%	(−9, 0)	.069
Primary care substitutable UCC visits (weekday)	25	27	32	38	−3.8* (2.0)	−10.53%	(−7, −1)	.057
Primary care substitutable UCC visits (non‐weekday)	11	12	14	17	−0.9 (0.8)	−5.87%	(−2, 0)	.244

Baseline is the year before CPC started (October 2011–September 2012). Estimates and predicted means are regression adjusted for baseline patient characteristics (including Hierarchical Condition Category scores, which are a measure of risk for subsequent expenditures) and baseline characteristics of the practice and the practice's market. Estimated difference‐in‐differences (DD) reflect the difference in changes in the regression‐adjusted average outcomes between beneficiaries in CPC practices and those in comparison practices from baseline to the 51‐month intervention period (October 2012–December 2016). Regression‐adjusted baseline and combined Years 1‐4 estimates are based on regressions with a single post‐intervention indicator for the duration of CPC. The DD percentage is the absolute DD estimate as a percentage of what the CPC group mean would be in the absence of CPC (that is, the CPC practices’ unadjusted group mean minus the DD estimate). There were 565 674 beneficiaries during any quarter in CPC practices and 1 165 284 in the comparison practices. There were 6 575 258 beneficiary‐level observations. ED visits are outpatient and include observation stays. Non‐weekday refers to weekends and major holidays. Primary care substitutable visits refer to the”non‐emergent” and “emergent, primary care treatable” categories of the New York University Emergency Department Algorithm. Potentially primary care preventable refers to the “emergent, preventable” and the “emergent, not preventable” categories of the NYU algorithm.

Abbreviations: CPC, Comprehensive Primary Care; DD, difference‐in‐differences; ED, emergency department; SE, standard error; UCC, urgent care center.

^*/**/*** Significantly different from zero at the 0.10/0.05/0.01 level, 2‐tailed test.

Source: Authors’ analysis of Medicare claims data for October 2011‐December 2016.

PC substitutable ED visits declined more relative to comparison practices than all‐cause ED visits (3% versus 2%), consistent with our hypothesis that strengthened primary care achieved through CPC would lead to more change in PC substitutable than all‐cause ED visits. The difference in the percentage change of visits between PC substitutable and all‐cause ED visits is statistically significant (P = .06; see Table S2 in Supplement 4). While PC substitutable ED visits represented 42% of all‐cause ED visits before CPC, their lower growth accounted for 60% of the reduction in all‐cause ED visits (Figure 2).

FIGURE 2.

Estimated share of the lower growth in the all‐cause Emergency Department visit rate, by emergency department category. Notes: This bar graph shows the share of the estimated reduction in all‐cause ED visits that each category contributes to, based on the DD estimates. Primary care substitutable visits refer to the “non‐emergent” and “emergent, primary care treatable” categories of the New York University Emergency Department Algorithm. Potentially primary care preventable visits refer to the “emergent, preventable” and the “emergent, not preventable” categories of the algorithm. “Injuries + other” visits refer to the injury, alcohol, drug, psych, and unclassified categories of the algorithm. To provide context on the breakdown of visits at baseline, primary care substitutable visits accounted for 42% of all ED visits, potentially primary care preventable visits accounted for 25%, and injuries + other visits accounted for the remaining 33%. Abbreviations: DD, difference‐in‐differences; ED, emergency department. Source: Authors’ analysis of Medicare claims data for October 2010‐December 2016[Color figure can be viewed at wileyonlinelibrary.com]

When we stratified visits by day of week, we found CPC practices had 4% lower growth in PC substitutable ED weekday visits relative to comparison practices (P = .002) and no difference in PC substitutable ED nonweekday visits. The overall PC substitutable ED results (−5.8 visits per 1000 beneficiaries) were driven by the weekday estimate (−5.5 visits per 1000 beneficiaries in a year). If CPC had an equal association with weekday and nonweekday PC substitutable visits, we would expect the weekday DD estimate to be 68% of the overall PC substitutable DD estimate (its proportion of PC substitutable visits before CPC); instead, it accounted for 95% of the overall PC substitutable estimate. As expected, statistical testing of this result could not reject the null hypothesis that the DD estimate for weekday PC substitutable DD visits was 95% of the DD point estimate for total PC substitutable visits (P = 1.0; see Table S2 in Supplement 4).

3.4. CPC and PPC preventable ED visits

CPC practices had 3% lower growth in PPC preventable ED visits, relative to comparison practices (P = .04), with reductions for both weekdays (2%; P = .07) and nonweekdays (3%; P = .07; see Table 3). The point estimates indicate that the overall decline (−3.5 visits per 1000 beneficiaries) was driven by proportional declines in weekday visits (−2.4 per 1000 beneficiaries) and nonweekday visits (−1.1 per 1000 beneficiaries). In support of this finding, we could not reject the null hypothesis that both weekday and weekend PPC preventable ED visits contributed to the lower growth in total PPC preventable ED visits in proportion to their baseline rates (P = .93 for each test; see Table S2 in Supplement 4).

The year‐by‐year DD estimates for PPC preventable ED visits were larger and more favorable in the last two years of CPC (Table S4 in Supplement 4). Lower growth in PPC preventable ED visits also disproportionally contributed to the lower growth in all‐cause ED visits, accounting for 36% of the reduction, but for only 25% of all baseline ED visits (Figure 2).

3.5. Comprehensive Primary Care and UCC visits

Regression‐adjusted DD estimates also showed lower growth in all‐cause UCC visits and PC substitutable UCC visits among CPC practices relative to comparison practices (Table 3). For each measure, CPC practices had 9% lower growth compared with comparison practices (P < .1). The all‐cause UCC and PC substitutable UCC regression‐adjusted rates increased for CPC beneficiaries over the course of CPC—from 61 to 78 annual visits per 1000 beneficiaries for all UCC visits, and from 36 to 46 annual visits per 1000 beneficiaries for PC substitutable UCC visits, respectively. However, again growth was lower in CPC than comparison practices.

We found differential percentage changes for PC substitutable UCC visits on weekdays versus weekends, but the difference was less prominent than for ED visits. CPC practices experienced 11% lower growth in PC substitutable UCC weekday visits (P = .06) and a 6% lower growth in nonweekday visits. With a low incidence of UCC nonweekday visits, the estimates have high standard errors and are not statistically different from zero (P = .2).

3.6. Robustness analyses

We studied the sensitivity of our results to the exclusion of observation stays from our ED visit measures. There are many factors that determine whether an ED sends a patient home versus keeps a patient for an observation stay or inpatient admission, and some of these factors could reflect the availability of effective primary care. Consequently, CPC may have affected the decision to transfer patients from ED to observation stay. At baseline, observation stays accounted for 10% of all‐cause ED visits, 9% of PC substitutable ED visits, and 16% of PPC preventable ED visits—with nearly identical proportions for CPC and comparison practices. When we removed observation stays from the construction of ED visits, we found that the takeaways were unchanged from our primary ED measure that includes observation stays. The size of the DD estimates increased modestly in magnitude using this alternative measure, meaning CPC was associated with even lower growth in ED visits (see Table S6 in Supplement 4). This evidence suggests that our results are not being driven by changes in observation stays associated with CPC.

We also assessed the sensitivity of our findings to a more conservative PC substitutable measure, which only includes nonemergent visits (Category 1). We found that the DD estimates on PC substitutable ED and UCC visits, overall and by weekday/nonweekday, were similar to the main specification, but the estimated percentage impacts were slightly larger (meaning even lower growth in ED visits), as we expected (Table S7 in Supplement 4).

As we anticipated, CPC was not associated with changes in ED visits for injuries (DD estimate was less than 1% and not statistically significant, P = .54; Table S8 in Supplement 4).

4. DISCUSSION

Comprehensive Primary Care` practices had 2% lower growth in all‐cause ED visits and 3% lower growth in PC substitutable ED visits than comparison practices among Medicare FFS beneficiaries. CPC practices also had 3% lower growth in visits requiring ED resources that were PPC preventable. Lower growth in weekday visits (4%) among CPC practices, rather than nonweekday visits, drove the PC substitutable ED results, while reduced growth in both weekday and nonweekday ED visits contributed to the PPC preventable ED findings. For UCC visits, we found lower growth in the all‐cause and PC substitutable UCC rates (9% each) for CPC practices and that changes in weekday visits were important for reducing PC substitutable UCC visits. We found no change in ED visits for injuries, as expected, which suggests that our results are not an artifact of a nonexperimental design or differences in secular trends between CPC and comparison practices.

Our results provide evidence that strengthened primary care delivery among CPC practices is associated with reduced ED and UCC visits. Our findings suggest this arose through two mechanisms:

1. Better access to primary care once symptoms develop. Our results suggest improvements in access during business days played a central role in reducing PC substitutable visits to EDs and likely also to UCCs. The differential change for weekdays versus nonweekdays is consistent with the hypothesis that improved access to the primary care team on regular business days (such as through reduced office wait times, more same‐day/next‐day appointments, expanded teams, and enhanced communication between patients and teams) may have helped slow the growth in ED and UCC visits. Because we found that the most common diagnoses for PC substitutable ED and UCC visits were nearly identical for weekdays and nonweekdays, our findings do not appear to be the result of nonweekday visits being less amenable to treatment in a primary care setting because they are more serious.

2. Enhanced care delivery that prevents or lessens acute exacerbations that require an urgent visit. Our findings for PPC preventable ED visits are also consistent with more general improvements in care delivery, beyond access, slowing growth in ED rates. Given that there was no weekday versus nonweekday differential, our findings suggest that CPC improved patient health by preventing or lessening acute symptoms that would result in an emergent visit, regardless of day of the week. For example, CPC milestone requirements, such as self‐management education, reviewing quality improvement measures, and transitional care after hospitalizations, could all reduce the need for emergent ED care. However, since we did not formally test any mechanisms in this study, we cannot isolate which care delivery activities reduced growth.

We estimated that CPC was associated with 23 079 fewer all‐cause ED visits over the course of the initiative. In 2017, MedPAC estimated the average Medicare payment cost of an outpatient ED visit encounter involving nonurgent care was $3089, and the average cost at a UCC for nonurgent care was $88 (both excluding patient cost sharing); they did not report the average cost for a primary care visit in this study. Assuming that all avoided ED visits associated with CPC were diverted to the primary care office and that, on average, these office visits cost no more than the UCC visit (a savings of approximately $3000 per avoided ED visit), our back‐of‐the‐envelope calculations suggest that CPC was associated with lower Medicare FFS expenditures on the order of $69.2 million. If CPC findings were applied to all 67.7 million Medicare beneficiaries nationwide (and the DD estimates remained the same as in our study); then, we estimate the potential savings to Medicare would be approximately $1.92 billion annually, or 0.3 percent of the $700 billion in Medicare health care spending in 2017.

In Supplement 4, we provide more details of these calculations. Note that these calculations do not take into account savings to non‐Medicare payers nor the costs of the CPC model itself. Further, the average Medicare payment for avoided ED visits may be lower than the overall ED visit average if lower‐acuity visits are more likely to be avoided. Finally, the DD estimates may not generalize to another set of practices. While these estimates are rough, they suggest that savings to Medicare from fewer ED visits is not trivial, and that CPC could have meaningful policy implications. Future research should examine whether other, less costly interventions could generate similar or larger results.

4.1. Limitations

This study has some important limitations. First, practices were not randomly assigned to CPC versus comparison group status. Despite having similar observable characteristics, CPC and comparison practices could have differences in unobservable characteristics that influenced outcomes. In addition, we did not adjust the p values for having multiple comparisons. For these reasons, the results should be considered exploratory.

Second, we applied the NYU algorithm to the UCC setting, even though it was created with ED physician input using ED data. A similar approach using UCC data might lead to different allocations of probabilities to the four categories. However, our focus on the nonemergent and primary care treatable categories for our PC substitutable UCC analysis, and not on the categories necessitating ED resources, reduces this concern.

Finally, the generalizability of our findings may be limited because CPC was tested in the regions, payers, and practices that volunteered to participate and were selected by CMS. Our analysis only studies Medicare FFS beneficiaries. Furthermore, given the flexibility of the CPC model, another set of practices might have transformed care differently, leading to different results.

In conclusion, CPC improved primary care delivery and was associated with lower growth in ED and UCC visits, particularly for weekday PC substitutable visits, which make up a sizeable proportion of visits. Lower growth in PPC preventable conditions in the ED setting also played a key role in slowing the all‐cause rate. Providing timely and effective primary care to patients during the week, and expanding access on the weekend, could be key to reducing unnecessary use of ED and urgent care and to lowering costs. More research on the contributions of specific care delivery changes could further explain our findings.

Supporting information

Author Matrix

Supplementary Material

Timmins L, Peikes D, McCall N. Pathways to reduced emergency department and urgent care center use: Lessons from the comprehensive primary care initiative. Health Serv Res. 2020;55:1003–1012. 10.1111/1475-6773.13579