The Effects of Physician Vertical Integration on Referral Patterns, Patient Welfare, and Market Dynamics

Abstract

The growth of physician vertical integration raises concerns about distorted referral patterns, higher spending, and market foreclosure. Using 100% Medicare data, we combine reduced-form analysis with a discrete choice model to estimate the effects of physician vertical integration on patients’ provider choices and welfare for two common “downstream” surgical procedures. Physician-hospital integration results in an approximately 10% increase in referrals to higher-priced facilities instead of lower-priced providers. Our counterfactual analysis implies that if all primary care physicians become integrated, total Medicare spending will increase by $315 million.

1. Introduction

The practice environment for the approximately 1 million physicians in the United States is rapidly changing. Rather than practice in traditional models of independent or solo practices, many physicians are now employed by hospitals or health systems. From 2012 to 2018, the share of physician practices employed by a hospital or health system increased from 46.8% to 54.1% (Kane 2019). From a clinical perspective, vertical integration between physicians and hospitals is designed to enable care coordination, enhance investments in technology, and improve patient care. However, vertical integration also raises important market competition questions. Physicians and hospitals that vertically integrate may do so to increase bargaining leverage with private insurers. Internalizing the bargaining leverage of larger hospitals allows for smaller physicians to obtain higher reimbursement rates (Peters 2014). Consistent with this model, several existing studies find that vertical integration leads to increases in prices providers negotiate with commercial insurers (Capps et al. 2018; Baker et al. 2014b, 2020; Chernew et al. 2021), but without noticeable changes on patient health outcomes (Koch et al. 2021).

A potentially more consequential impact to market competition is the effect of vertical integration on patient demand and allocation of patients across providers. Unlike many other markets, physicians often act as gatekeepers for patients (Chernew et al. 2021). The role physicians play in directing patient care raises the potential that hospitals may acquire “upstream” physician practices in the hopes of steering patient demand for “downstream” services (e.g., hospital care) (Baker et al. (2016); Brot-Goldberg and de Vaan (2018)). This type of behavior follows common models of input foreclosure (e.g., Ordover et al. (1990)), but with few exceptions, how hospitals use vertical integration to acquire patient volume has not been fully examined. Related work examines the impact of vertical integration on self-referrals for skilled nursing facilities (Cutler et al. 2020) and, most closely related to this paper, the impact of hospital-surgeon vertical integration on outpatient surgical procedures (Richards et al. 2022).

The existing literature on health care market competition has primarily focused on horizontal competition between hospitals (e.g., Dafny et al. (2019)), but has not fully considered the allocation impacts of vertical integration. From a policy perspective, understanding this dynamic is important. Because physician vertical consolidation events typically occur among small physician practices, as opposed to a single hospital or health system merger event, they are typically outside the scope that would potentially merit intervention of federal and state regulators (Capps et al. 2017). This dynamic creates the potential that hospital-physician vertical integration may be a “backdoor” route to horizontal market consolidation that would have been opposed if it occurred through hospital mergers. Fully understanding the impacts of vertical integration on procedure allocation and referral patterns is especially important given the size of the U.S. health care market and how rapidly physician practices are vertically integrating.

In this paper, we use data over the 2013 to 2019 period from a 100% sample of the U.S. fee-for-service Medicare population to test how vertical integration between physician groups and hospitals changes patient demand for downstream hospital services. We focus on two procedures of particular interest—arthroscopy surgeries and colonoscopies. Both procedures are extremely common among the Medicare population—our data includes approximately one million arthroscopy and 10 million colonoscopy procedures. In addition, both procedures are commonly performed in two distinct provider organizations—hospital outpatient departments (HOPDs), which are a unit of a traditional hospital, and ambulatory surgical centers (ASCs), which are freestanding centers that are often independent from hospitals. Several other studies have compared ASCs and HOPDs, and generally found that ASCs are less costly and have equivalent or higher quality (e.g., Gardner et al. (2005); Grisel and Arjmand (2009); Munnich and Parente (2014); Munnich and Richards (2020)).

Examining how vertical integration impacts competition between ASCs and HOPDs is particularly relevant in the Medicare population, as Medicare imposes strict fee schedules that pay each provider organization different rates. Unlike private insurance markets, where prices are negotiated, vertical integration does not allow providers to negotiate higher prices with the Medicare system. However, Medicare reimburses ASCs at approximately one-half to two-thirds the reimbursement rate for HOPDs. For example, in a common procedure we examine, a colonoscopy with a biopsy, Medicare pays $805 for procedures performed in an ASC and $1,371 for procedures performed in HOPDs.¹ From a purely financial standpoint, performing the same procedure in an HOPD instead of an ASC creates an “arbitrage” opportunity to increase Medicare payment by re-allocating patient volume from ASCs to HOPDs (Chernew (2021)). If vertical integration leads to increases in the use of HOPDs, this site-of-care-based payment differential potentially serves as both a motivation for and a consequence of hospital acquisition of physician practices. Related work finds that increases in site-based payment differentials contribute to physician vertical integration (Post et al. (2021)) and reverses efficiencies of vertical integration (Lin et al. 2021). This paper extends the existing literature on hospital-physician vertical integration and referral patterns by formalizing how vertical integration can change physician referral incentives.

To address these questions, we leverage changes in primary care physician (PCP) affiliation to examine how patients choose between HOPDs and ASCs for “downstream” procedures. PCPs are a patient’s entry point into the health care system and are the focal point of care delivery. For the two procedures we examine, a patient selects a PCP, who then refers the patient to a specialist physician, who then schedules a surgery for the patient at either an ASC or HOPD. Notably, we use PCPs to measure a patient’s exposure to vertical integration, rather than specialist physicians (e.g., orthopedic surgeons for orthopedic surgeons and gastroenterologists for colonoscopies) due to the concern that health systems may influence patient referrals to these same specialists. Thus, our results represent a “intent-to-treat” measure of the impacts of integration on facility choice for outpatient surgery services. Over our sample period, the share of Medicare beneficiaries with an integrated PCP organization increased by 35%.

We first find reduced-form evidence that hospital employment of physician practices shifts patient demand to hospitals and away from ASCs. Following vertical integration, there is a 5.0 percentage point increased use of HOPDs instead of ASCs for artrhoscopy surgery, and a 6.8 percentage point increase for colonoscopies. These changes translate to 8.2% and a 9.3% relative increase for arthroscopy and colonoscopies, respectively. Due to Medicare payment rules, increased use of HOPDs leads to an approximately 2% increase in Medicare payments across all surgical procedures received by patients attributed to a newly-integrated primary care provider. We also find increases in the travel distance between the patient and the provider (4.8 log-miles for arthroscopies and 3.2 log-miles for colonoscopies). Importantly, we find no meaningful change in clinical quality measures.

To more fully understand how physician vertical integration can impact referral choices, and how these changes to patient allocaiton across providers can impact patient welfare, we develop and estimate a discrete choice model to describe the patient-physician joint decision of healthcare delivery organization for “downstream” procedures. Similar to our reduced form analysis, our estimates suggest that vertical integration leads to a 3.1 and 8.1 percentage points higher probability of choosing an HOPD rather than an ASC for arthroscopy and colonoscopy, respectively. At an aggregated level, our counterfactual estimates imply that for just these two services, changing from status quo to fully integrated relationships for all physicians will lead to a $315.4 million increase in Medicare spending and a $63.1 million increase in patients’ out-of-pocket payment costs.

Our work extends existing studies on the impacts of hospital-physician vertical integration on referral patterns. Most notably, this paper extends Richards et al. 2022, which examines the impact of hospital-physician vertical integration and use of ASCs vs. HOPDs for similar outpatient surgical procedures using all-payer data from Florida. Our paper adds a 100% sample of Medicare claims data, which allows us to track changes in Medicare payments, rather than ”chargemaster” prices. Our data also allows us to track post-procedure patient outcomes as a way of measuring quality impacts. Finally, our combination of reduced form and structural approaches allows us to both examine mechanisms and estimate counterfactuals. This paper is also closely related to Whaley et al. 2021, which uses the same Medicare 100% sample to examine changes in site of care and Medicare spending following vertical integration for common laboratory and imaging tests.

Our results have implications for the economics of understanding provider competition in an important market. Medicare and other insurers commonly pay health care providers different amounts for care performed in different settings. To the extent to which site-of-care payment differentials motivate vertical integration, our results show the impacts of allocation changes following vertical integration for both the Medicare system and patients. At the same time and independent of changes to payment policies, regulators have faced questions of how to address vertical integration.

2. Medicare claims data

We use data from a 100% sample of Medicare claims data that covers the 2013 to 2019 period, including all inpatient, outpatient, physician, and other types of care for Medicare enrollees who are enrolled in Traditional Medicare (e.g., fee-for-service Medicare). These data include approximately 38 million enrollees per year. The breadth of this data allows us to comprehensively examine the impacts of hospital-physician vertical integration on provider referral patterns in the US.

With the data, we identify two specific procedures of interest—joint arthroscopy, a minimally-invasive surgical procedure used to diagnose and repair minor joint damage, and colonoscopy, a procedure used to screen for and to treat early-stage colorectal cancer.² We restrict our sample to beneficiaries between ages 65 and 100 and who are continuously enrolled in Medicare for all 12 months in a year. We also exclude cases where the patient and provider are located more than 100 miles apart. We also limit the sample to Medicare beneficiaries attributed to a provider care organization (described below) and procedure claims with valid payments. After imposing those restrictions, we obtain two full samples— an arthroscopy sample with approximately 830,000 observations and a colonoscopy sample with approximately 11 million observations.

For each procedure, we identified procedures performed in outpatient settings—an Ambulatory Surgical Center (ASC) or a Hospital Outpatient Department (HOPD)—and excluded the small number of procedures performed in other settings, such as physician office, inpatient, and emergency department settings, using the site-of-care codes contained in the Medicare claims. Of these, 43% of arthroscopies and 52% of colonoscopies were performed in an ASC instead of an HOPD. As secondary outcomes, we also measured the total Medicare reimbursement amount for each procedure and patient payments, and the distance between the centroid of the patient’s zip code and the zip code as the provider.³ Finally, an advantage of these procedures is the existence of clinical complication measures. For each procedure, we identified both procedure and post-operative complications using the approach applied in previous studies (Whaley and Brown (2018), Aouad et al. (2019)). These complications, which are fully listed in Appendix Tables A1 and A2, include adverse outcomes like intestine perforations, cardiac events, and readmissions.

2.1. Group identification and integration status

In addition to information on procedure types, we also used the data to collect information on patient primary care. Existing studies on vertical integration and referrals have examined market-level trends in vertical integration (Scheffler et al. 2018) and integration status of referring physicians (e.g., Baker et al. (2016); Chernew et al. (2021)). A challenge with market-level vertical integration or referring-level measures is the potential that changes in patient allocation occur before a patient is referred to a specific provider. We instead examine the impact of changes in referral status of the patients’ PCP. Doing so allows us to measure the impacts of vertical integration on patient allocation using the patient’s initial contact with the health care delivery system. To construct patient vertical integration measures, we identify the physician group that provides the majority each patient’s primary care services. A full description of the attribution model used is described in Appendix B.

Consistent with existing studies, in our sample, we find an increasing trend in the share of providers that are vertically integrated. This trend, shown in Figure 1, is similar to results from surveys conducted by the American Medical Association (Kane 2019). Panel A shows that the number of physicians working in an integrated group increased from 46.1% to 53.9%. Panel B shows that the increase in physician group vertical integration is accompanied by a 35.4% increase (from 30.0% to 40.6%) in the number of Medicare beneficiaries who receive primary care from a vertically integrated group. This larger increase in the number of beneficiaries provides descriptive evidence that groups integrated during this period are larger than groups that did not integrate and that PCP market became more concentrated over the 2013 to 2019 period.

Figure 1:

Hospital-Physician Vertical Integration Trends

Notes: The two subplots show the trend of vertical integration from 2013 to 2019. The sample is the 100% Medicare Fee-For-Service sample. For the top figure, the integration is measured by the number of physicians working at integrated PCP groups. For the bottom figure, the integration rate is measured by the number of Medicare beneficiaries whose assigned PCP groups are integrated.

2.2. Market definition

Traditional fixed geographical market definitions do not change across years and prohibit patients from choosing from other regions, even though patients may live close to a geographical boundary. To relax these restrictions, we combined geographical regions with patient flow to define markets. We start with Hospital Referral Regions (HRRs), a fixed geographic market definition that has been constructed by the Dartmouth Atlas of Care to measure hospital market areas (Atlas et al. (2009)). For each HRR, we include all the options into the choice set for that HRR if they have ever been chosen by patients in that region and the distance to the patient is shorter than 100 miles. Using this market definition, we allow market boundaries to vary across years, and allow patients to travel across HRRs if the options are popular in their HRRs and not too far from their homes. Following this definition, the average market for arthroscopy has 365 patients and 26 providers (16 in HOPDs), and the average market for colonoscopy has 4482 patients and 76 providers (45 in HOPDs).

2.3. Choice model sample construction

We construct four separate analytic samples based on the procedure and the analytic approach—two for the reduced-form analysis that measures physician group level changes in referral patterns following vertical integration and the other two samples for patient choice model analysis. First, we exclude groups who churned in and out of systems between 2013 and 2019. After that, we aggregate visits to PCP group-month-procedure code level. We count the number of visits as the total volume and number of visits in hospital-based facilities as hospital-based volume. For the other outcome variables, we use the average value within each cell. The reduced-form sample has 872,382 observations of 50,393 PCP groups for joint arthroscopy, and 3,735,833 observations of 74,235 PCP groups for colonoscopy.

The other two samples are constructed for the choice model. We first apply the aforementioned market definition. Then, we restrict patients’ choices to popular choices. For each market, we drop providers whose number of visits are below median, keep the most popular five providers, and group the rest of them into two “others” options—one for HOPD and one for ASC⁴. To deal with repeated choices, we assume that each visit decision is made independently.⁵ The final arthroscopy choice model sample includes 373,185 beneficiaries, 22,982 PCP groups, and 4,046 arthroscopy providers. The colonoscopy sample has 5,719,086 beneficiaries, 45,176 PCP groups, and 7,550 colonoscopy providers. Due to computational limitations, we use a 25% random sample of the final choice model samples for estimation. Note that the most important contribution of a 100% sample is to have a complete choice set and accurate characteristics for each option in the choice set. In our 25% random sample, beneficiaries’ choice sets are constructed based on the 100% sample.

Descriptive characteristics of each sample are presented in Table A3. Panel A presents characteristics of group-level samples that used for the reduced form analysis and Panel B presents characteristics of patient-level samples that used for the patient choice analysis.

3. Reduced-form effects of vertical integration on facility allocation and patient outcomes

We first measure the reduced form effects of physician vertical integration on main outcome, site-of-care choice between ASCs and HOPDs, and our secondary outcomes. To measure the reduced-form effect of vertical integration, for a procedure $k$ received by patient $i$ from provider $j$ in time period $t$, we use a difference-in-differences approach and estimate a regression model of the form

$$Y_{ikjt}=\theta V{I}_{jt}+\beta X_{it}+{\gamma }_k+{\zeta }_j+{\delta }_t+{\epsilon }_{jt}$$ (1)

where $Y_{ikjt}$ is the outcome of interest. Our main outcome is an indicator that a procedure is performed in a HOPD owned by the acquiring hospital instead of in an ASC. As secondary outcomes, we also include the log-transformed Medicare payment, the log-transformed miles between the patient’s and provider’s zip code centroids, and an indicator for any procedural complications, which we use to measure the quality of care. $V{I}_{jt}$ denotes that provider group $j$ is vertically integrated with a hospital in year $t$. We control for patient characteristics (age, CMS hierarchical condition category risk score, race, and gender).⁶ Because within each procedure different surgeries have different intensity and thus reimbursement rates (e.g., a colonoscopy with vs. without a biopsy removal), we also include fixed effect for the specific type of surgery.⁷ The ${\delta }_t$ and ${\zeta }_j$ terms represent time and primary care provider group fixed effects, respectively. The $\theta$ coefficient measures changes in each patient outcome following vertical integration of that patient’s primary care provider organization, relative to existing trends among the control population. We estimate this model using OLS regressions and cluster standard errors at the primary care provider group $\left(j\right)$ level.

Our primary identification assumption is that provider organizations who select into vertical integration arraignments do so independently of other contemporaneous changes to the outcomes we measure. We include physician group fixed effects, which account for any time-invariant differences between groups (e.g., the referral pattern preferences). A potential threat to our identification approach is if PCP groups select into vertical integration due to factors that impact our outcomes (e.g., changes in underlying patient preferences for ASCs vs. HOPDs). In such a case, trends in our outcome variables of interest will likely change prior to vertical integration events. To test for these trends, we also estimate event study regressions that test for the “parallel trends” assumption common in difference-in-differences designs. We estimate both two-way fixed effects event studies and ”dynamic” event studies that account for the differential timing of hospital-physician vertical integration events, as well as potential heterogeneous differences that are correlated with treatment timing. We use the Sun and Abraham (2021) estimator for alternative event studies, presented in the Appendix.

3.1. Reduced form results

Table 1 reports the reduced-form difference-in-difference regression results that show changes in our primary outcome of interest, use of within-system HOPDs following acquisition of a Medicare beneficiary’s primary care provider organization by a hospital or health system, as well as our secondary outcomes, changes in Medicare facility payments, patient travel distance, and complications. The results in column 1 show a 5.0 percentage point increase in use of HOPDs vs. ASCs for arthroscopy and a 6.8 percentage point increase for colonoscopy procedures. These changes translates to relative increases of 8.2% and 9.3%, respectively. The increased use of HOPDs corresponding leads to approximately 2% increases in Medicare payments for each procedure. Notably, scaling the change in Medicare payments by the increased use of ASCs (e.g., column 2 divided by column 1) equate to IV estimates of 43% and 35% higher Medicare payments at HOPDs than ASCs for arthroscopy and colonoscopy, respectively. For our two measures of patient welfare, column 3, shows 4.8 and 3.2-unit increases in log patient travel distance to receive care for the two procedures. In column 4, we do not find any change in procedural complication rates. The estimated coefficients are small in magnitude and imprecise.

Table 1:

Reduced Form Difference-in-Differences Results

	(1) HOPD	(2) log payment	(3) log distance	(4) complication
Arthroscopy Surgery
Post integration	0.0496*** (0.00874)	0.0211 (0.0134)	0.0481*** (0.0174)	−0.00109 (0.00178)
Observations	818,980	818,980	818,980	818,980
R-squared	0.387	0.322	0.224	0.094
Colonoscopy
Post integration	0.0679*** (0.00968)	0.0235** (0.0101)	0.0318** (0.0131)	0.00108 (0.00139)
Observations	10,941,446	10,941,446	10,941,446	10,941,446
R-squared	0.450	0.239	0.148	0.083

Notes: This table presents difference-in-difference regression results for our main outcome, use of within-system HOPD following hospital-physician vertical integration, as well as secondary outcomes, log Medicare facility payment, log patient travel distance, and complication likelihood. All regressions include patient covariates (age, gender, HCC risk score), procedure code fixed effects, year and month fixed effects, and provider group fixed effects. The standard errors are clustered by provider group.

^***p < 0.01,

^**p < 0.05,

^*p < 0.1

In addition to primary care provider organization time fixed effects, our main results include patient covariates (age, gender, non-white race indicator, and Hierarchical Condition Category (HCC) risk scores). In Appendix Table A4, we test the sensitivity of these covariates. For each outcome, we test the inclusion of just the provider and time fixed effects (the first column for each outcome), the addition of the patient covariates (the second column), and the addition of state-by-year fixed effect interactions (the third column). Across all outcomes, we find little impact of covariate structure. In particular, the inclusion of patient covariates has no meaningful impact on our reduced form evaluation of the impacts of hospital-physician vertical integration. Perhaps not surprisingly, in supplemental analysis, we find no change in the underlying patient characteristics (using the same four observables) in a primary care provider’s patient panel. We also do not find changes in the composition of patients who receive care at an HOPD following integration, suggesting that the marginal patients shifted from an ASC setting to a HOPD setting are not advantageously “cream skimmed” or selected based on their underlying characteristics.

To further test the robustness of our reduced form results, Figure 2 presents our main event study results, which measure the change in use of within-system HOPDs for our two procedures of interest in the years before and after hospital-physician vertical integration. For both arthroscopy (Panel A) and colonoscopy (Panel B) procedures, we do not find pre-integration trends in HOPD use. However, following hospital acquisition, the treatment primary care providers in our sample have a sharp and immediate increased use of HOPDs of approximately 5 percentage points. These event study results are consitent with our above-discussed difference-in-differences results. Appendix Figures A1 and A2 find similar results when examining monthly trends and when adjusting for potential treatment effect heterogeneity using the Sun and Abraham (2021) approach, respectively.

Figure 2:

Event Study Results

Notes: The plots show the estimates and 95% confidence intervals from the event study regressions that measure annual changes in use of within-system hospital outpatient departments (HOPDs) following hospital acquisition of primary care provider organizations. Event studies are estimated using two-way fixed effect regressions. The top and bottom panel correspond to arthroscopy and colonoscopy procedures, respectively. In all the regressions, we control for patient covariates (age, gender, HCC risk score), procedure code fixed effects, year and month fixed effects, and provider group fixed effects. The standard errors are clustered by provider group.

4. Effect of Vertical Integration on Downstream Provider Choice

Our reduced-form approach measures the effects of vertical integration on dimensions that include resource allocation (ASC vs. HOPD), Medicare spending, and measures of patient welfare (travel distance and procedural quality). A limitation of this approach is that it is unable to aggregate these dimensions into a single metric of patient welfare. To mitigate this concern, we employ a discrete choice model for the choice of downstream providers. In this model, we include all three parties–patients, PCP groups, and downstream surgeons. Thus, this model enables us to incorporate patients’ characteristics into the referral decision.

4.1. Patient and provider utility function

To model patient choice of provider, we start with the patients’ perspective. In the most straightforward approach, patients select surgeons based on how much they will pay for the procedure, the quality of the surgeon, and how long they need to travel. In addition to provider characteristics, patients’ own characteristics also affect their utility, For example, patients with different age and gender may obtain different level of utility from a given provider. Therefore, we assume that patient $i$’s utility from choosing surgeon $k$ at time $t$ is

$$v_{ikt}=X_{it}^{{pat}^{\prime }}{\beta }^X+d_{ik}{\beta }^d+X_k^{{surg}^{\prime }}{\beta }^S+W_t^{\prime }{\beta }^W+{\delta }_t+{\xi }_{ikt},$$ (2)

where $X_{it}^{pat}$ is a vector of patient’s characteristics—age and gender, $d_{ik}$ is the travel distance between patient $i$ and provider $k,X_k^{surg}$ is a vector of surgeon’s characteristics such as average patient payments, complications rate, and the number of patients for whom the surgeon performed procedures, $W_t$ is a vector that contains region information, and ${\delta }_t$ is a time fixed effect.

However, the selection of downstream providers is not driven just by patient preferences, but instead incorporates provider decision-making and referral patterns. Several papers expand upon the model of patient choice in equation (2). Healthcare providers are altruistic—they take account their patients’ utility in addition to their own preferences (Arrow (1963); McGuire (2000)). Therefore, the selection of surgeon for a downstream service is a joint decision that depends on both patients’ utility and PCPs’ preference. When PCP groups are vertically integrated with hospitals or health systems, they will be strongly incentivized, if not required, to refer patients to their owning hospitals or systems, while independent PCP groups do not have such incentive.

We model this joint decision by expanding equation (2) to include both patient and provider predictors of choice. We estimate the choice utility of referring patient $i$ to provider $k$ at time $t$ as

$$u_{ij\left(i\right)kt}=X_{it}^{{pat}^{\prime }}{\beta }^X+X_{j\left(i\right)t}^{PC{P}^{\prime }}{\beta }^P+V{I}_{j\left(i\right)t}{\beta }^V+d_{ik}{\beta }^d+X_k^{{surg}^{\prime }}{\beta }^S+W_t^{\prime }{\beta }^W+{\delta }_t+{\epsilon }_{ij\left(i\right)kt}.$$ (3)

In this model, $X_{j\left(i\right)}^{pcp}$ is a vector of characteristics of patient’s assigned PCP such as group size, $V{I}_{j\left(i\right)t}$ is an indicator of PCP $j$ being integrated at time $t.{\epsilon }_{ij\left(i\right)kt}$ is the idiosyncratic error that follows the Gumbel distribution. In our model, patients’ PCP are assigned annually. Thus, for a given patient $i$ in market $t$, the patient’s PCP $j$ is given. We will eliminate the subscript $j$ thereafter for simplification.

4.2. Model structure

As discussed above, our two downstream services, colonoscopy and arthroscopy procedures, can be performed in two separate delivery organization types—HOPD and ASC. The preference of patients and PCPs over these two facility types is not completely captured by the observed characteristics. Because of the obvious difference between the two types, the unobserved provider characteristics could be systematically different. Therefore, we assume that patients and PCPs have different sensitivities for the two types. In the model, we group surgeons based on their facility types and use a nested-logit model to describe the joint choice of surgeon (McFadden (1980); Goldberg (1995)). Our model has two virtual levels—top-level for facility type and bottom-level for surgeons. The top-level includes the two aforementioned facility types. For each facility type, at the bottom-level, the choice set includes all the popular facilities and an “others” option for that facility type. The outside option is not having any procedure. Although our model has two levels, the choice of facility type and that of the surgeon are from one single decision that the patient makes jointly with his/her PCP.

Let $C_1\in \{H,I\}$ denotes the top level choice and $C_2\in \left\{K_H,K_I\right\}$ denotes the bottom level choice. At the bottom-level, we model the choice by a multinomial logit model. The conditional probability of patient $i$ choosing surgeon $k\in K_f$ given facility type $f$ is

$$Pr\left(C_2=k\mid C_1=f\right)=\frac{exp((d_{ik}{\beta }^d+X_k^{surg}{\beta }^S+{\delta }_t)/{\tau }_f)}{\sum _{l\in K_t} exp((d_{il}{\beta }^d+X_l^{surg}{\beta }^S+{\delta }_t)/{\tau }_f}$$ (4)

where ${\tau }_f$ is the dissimilarity parameter, indicating how similar the two types are. The more closer to one the ${\tau }_f$ is, the more similar the two types.

At the top-level, the probability of patient $i$ choosing facility type $f$ is

$$Pr\left(C_1=f\right)=\frac{exp\left(X_i^p{\beta }_f+{\tau }_f{I}_f\right)}{\sum _{s\in \{H,I\}} exp\left(X_i^p{\beta }_s+{\tau }_s{I}_s\right)}$$ (5)

where $X_f^p$ represents is a vector of patient-PCP group characteristics. Linking to the utility representation, $X_f^p$ includes $X_{it}^{pat}$, $X_{\left(j\right(i\left)t\right)}^{pcp}$, and $V{I}_{j\left(i\right)t}.I_f$ is the inclusive value associated with choice $f$,

$$I_f=ln\left(\sum _{k\in K_f} exp\left(\frac{d_{ik}{\beta }^d+X_k^{surg}{\beta }^S+{\delta }_t}{{\tau }_f}\right)\right)$$

Since the integration decision needs a relatively long preparation time and it is made at the PCP group level instead of PCP level, we assume away the endogeneity of the VI indicator in our model.

5. Choice Model Results

Columns (1) to (4) and (5) to (8) of Table 2 present the estimates from the choice model with different specifications for arthroscopy and colonoscopy. At the bottom of each column, we report the average marginal effect of vertical integration on the top-level choice probability. For joint arthroscopy procedures, column (1) shows that when a patient’s PCP group changes from independent to vertically integrated, this patient will be 3.37 percentage points more likely to go to a hospital-based facility for the arthroscopy procedure, which is a 6.1% relative increase. In column (2), we add three market characteristics—per capita income, number of non-federal PCPs, and number of non-federal MDs—to control for the geographical income variation and local market competition. After controlling for the market characteristics, the effect of vertical integration reduces slightly, from 3.37 percentage points to 3.03 percentage points. In Column (3), we allow for the interaction between OOP payment and complication rate. In this way, the trade-off between OOP payment and quality is flexible. This modification does not meaningfully change the estimates. The model corresponding to Column (4) allows for the interaction between surgeon characteristics and the PCP group integration status. This specification relaxes the assumption that patients and PCPs respond to the procedure price and quality in the same way, regardless of whether the PCP group is integrated or independent. When this interaction is included, the effect of vertical integration changes to 3.11 percentage points, a 5.7% relative increase. Table A5 provides the estimates of the full model. The bottom panel presents the dissimilarity parameter, which is a measure of similarity among alternatives intra- and inter-nest. As shown in the table, the dissimilarity parameters are 0.648 and 0.986 for ASCs and HOPDs, both smaller than one. Those estimates suggest that the options within one facility type are more substitutable than across-facility types.

Table 2:

Estimates from the discrete choice model

A. Arthroscopy (N = 1,882,870)
	(1)	(2)	(3)	(4)
Integrated	0.161*** (0.0203)	0.145*** (0.021)	0.144*** (0.0209)	0.149*** (0.0157)
Market characteristics		X	X	X
OOP X complication			X
Surgeon characteristics X integrated				X
AME (VI)	0.0337 (0.0068)	0.0303 (0.0061)	0.02999 (0.006)	0.0311 (0.0065)
B. Colonoscopy (N =18,259,705)
	(5)	(6)	(7)	(8)
Integrated	0.344*** (0.0214)	0.332*** (0.0195)	0.332*** (0.0193)	0.373*** (0.024)
Market characteristics		X	X	X
OOP X complication			X
Surgeon characteristics X integrated				X
AME (VI)	0.0748 (0.0013)	0.0719 (0.0013)	0.0719 (0.00134)	0.0807 (0.0014)

Notes: The eight columns report the estimates and the average marginal effect (AME) from eight different nested-logit regressions. The top panel focuses on arthroscopy while the bottom panel focuses on colonoscopy. All the nested-logit regressions control for patient characteristics (age and gender), PCP group size, surgeon characteristics (average OOP, complication rate, and the number of Medicare beneficiaries with the same procedures), traveling distance, and year fixed effects. Columns (2)-(4), and (6)-(8) also include market characteristics (per capita income, and number of non-federal PCPs). Columns (3) and (7) allow for the interaction between OOP payment and complication rate. Columns (4) and (8) allow for the interaction between surgeon characteristics and PCP groups’ integration status. Standard errors of estimates are clustered by year. The standard errors of AME are calculated by the Delta method.

^***p < 0.01,

^**p < 0.05,

^*p < 0.1

Results for colonoscopy procedures have similar patterns but larger magnitude. Column (7) shows that becoming vertically integrated leads to a 7.48 percentage points increase (15.9% relative increase) in the probability of referring to a hospital-based facility. The average marginal effect remains at similar level—7.19 percentage points, when control for market characteristics and allow for interaction between payment and quality. In our final specification that allows for the interaction between surgeon characteristics and PCP group integration status, we estimate a 8.07 percentage point increase in HOPD use, which translates to a 17.2% relative increase. The dissimilarity parameters, shown in Table A5, are 0.657 and 0.825 for ASCs and HOPDs, respectively.

5.1. Counterfactual analysis

Our estimates show that vertically integrated groups have significantly higher probability of referring patients to the hospital-based facilities rather than to free-standing facilities. In this section, we compare the observed scenario to two counterfactual market structures—a fully vertically integrated market and a fully independent market. By comparing to the two counterfactual scenarios, we estimate the welfare effects of vertical integration and the effects on non-monetary dimensions. To do so, we change the integration indicator to either one or zero for all PCP groups to reflect the counterfactual settings. Using parameters estimated from the choice model, we predict patients’ choice probability and assess the effect of vertical integration on volume, payment, travelling distances, and quality (Table 3).

Table 3:

Effects of vertical integration on patient welfare

	Arthroscopy		Colonoscopy
	All integrated (1)	All independent (2)	All integrated (3)	All independent (4)
N decisions (million)	4.42		10.89
Δ HOPD volume (thousand)	21.82	−12.55	542.07	−337.9
Δ OOP (million)	6.23	−1.17	56.85	−24.67
Δ Medicare total payment (million)	31.16	−5.85	284.26	−123.34
Δ Distance per patient	0.01139	−0.00195	0.1691	−0.05106
Δ Complication rate (z-score)	0.00041	0.00012	0.00993	−0.00449

Notes: The table shows the welfare analysis based on the choice model estimation. The first two columns are for arthroscopy surgeries and the last two columns are for colonoscopies. Columns (1) and (3) present changes when the market settings change from status-quo to all PCP groups integrated with hospitals or health systems. Column (2) and (4) present changes if all the integrated groups become independent.

When all the PCP groups are vertically integrated, 21,820 more arthroscopy surgeries and 542,066 more colonoscopy procedures will be performed at HOPDs—an 11.0% and 20.0% increase in volume, respectively. However, it is important to note that those estimates do not mean that all HOPDs will receive more patients. If changing to the other scenario, where all the PCP groups are independent, then the total volume of hospital-based visits will decrease by 12,553 for arthroscopy (a 6.3% relative change) and 337,899 for colonoscopy (a 12.5% relative change).

Due to site-of-care payment differentials, these changes in allocation lead to changes of millions of dollars in Medicare payments. To assess the effect on payments, we assume that the average level of payment for each surgeon remains the same in the counterfactual settings. This assumption is likely valid because PCP group integration has little influence, if any, on the Medicare fee schedule for downstream providers. Our estimates imply that if all PCP groups are vertically integrated, then patient payments will increase by $6.2 million for joint arthroscopy procedures and by $56.9 million for colonoscopy procedures. Converting to the total payment by Medicare and beneficiaries, the estimated increase in payments is $31.2 million for arthroscopy and $284.3 for colonoscopy procedures. In the other counterfactual setting, we estimate savings of $129.2 million for the two procedures together.

In addition to the two extreme scenarios, we calculated the effects of changing the integration rate from 30 percent to 40 percent, which is very close to the change from 2013 to 2019. The ten percentage points change leads to 95,065 more procedures in HOPDs and $20.4 million more total payments.

Similar to the assumptions for Medicare and patient payments, we assume that locations and the average complication rates for a surgeon are unchanged. This assumption could be strong for long-run analysis, but in short-run, it is unlikely that changes in provider locations or provider quality are induced by the vertical integration. Based on this assumption, the change in distance is not significant for patient seeking joint arthroscopy. The effect on distance is larger for colonoscopy patients. However, the magnitude is still not significant economically. We do not find meaningful changes in complication rates.

6. Discussion and Conclusion

The market environment for U.S. physicians is rapidly changing, as a growing share of physician practices have vertically integrated with hospitals and health systems. Changes in physician group ownership have led to concerns about how changes in physician financial incentives impact patient care. This concern is particularly noteworthy in the market for outpatient surgeries for Medicare patients. Due to the large difference in Medicare payment rates based on the site-of-care in which a procedure is performed, vertical integration can create incentives to perform procedures in hospital outpatient departments, rather than in ambulatory surgical centers. This paper show that the vertical integration increases the probability of referring patients to hospital-based facilities. Vertical integration of PCP groups leads to allocation changes in delivery settings for outpatient care. These allocation changes in turn lead to increased spending and increased travel time for patients. Importantly, we did not find any quality improvement brought by the vertical integration, for the two procedures. Future studies are needed to draw a broader conclusion for the total utilization and for the overall health outcomes. However, the lack of a measurable change in quality suggests that efficiency gains from vertical integration may not materialize to patients.

This paper is not without limitations. For one, we restrict our analysis to the Medicare fee-for-service population. Medicare sets prices administratively, which is in stark contrast to the negotiated-prices system among private insurers. Other work has shown that vertical integration leads to increase in prices for private insurers (Baker et al. 2014a). Thus, it is likely that the changes in procedure prices and spending are under-estimated in the Medicare system, relative to the private insurance system. In addition, we only examined two common outpatient procedures. However, the same site-of-care payment differentials exist for other outpatient surgeries, as well as for procedures like diagnostic imaging and laboratory tests (Whaley et al. 2021). Finally, we do not examine how increased Medicare payments are allocated between physician groups and hospitals.

Despite these limitations, the results of this paper indicate how payment incentives can impact the allocation of patients to providers. These results have several important policy implications. First, the extent to which Medicare and other insurer site-of-care differentials contribute to the incentives for health care systems to vertically consolidate suggests that reducing the “arbitrage” opportunities in payment models could limit the impacts of vertical integration. Second, the extent to which changes in patient allocation across providers represent allocation inefficiencies, such as requiring patients to travel further or increases in patient cost-sharing payments, suggests that more regulatory oversight of vertical consolidation may be warranted. While the Department of Justice and the Federal Trade Commission have provided guidelines for monitoring vertical consolidation, the impacts of regulatory activity on vertical consolidation of health care providers is uncertain (DoJ and FTC 2020).

• Hospital-physician vertical integration is a dominant trend in US health care markets.

• Due to site-of-care differences in Medicare payment policy, vertical integration creates an “arbitrage” opportunity to move patients to higher-priced providers.

• Vertical integration leads to increased use of hospital outpatient departments for surgical services, leading to increases in Medicare spending.