Does Multispecialty Practice Enhance Physician Market Power?

Abstract

In health care, vertical integration – common ownership of producers of complementary services – may have both pro- and anti-competitive effects. We use data on 40 million commercially-insured individuals from the Health Care Cost Institute to construct price indices for office visits to general-practice and specialist physicians for the years 2008-2012. Controlling for generalist market concentration, we find that generalists charge higher prices when they are integrated with specialists, and that the effect of integration is larger in more concentrated specialist markets. Conversely, controlling for specialist market concentration, specialists charge higher prices when integrated with generalists, with larger effects in more concentrated generalist markets. Our results suggest that multispecialty practice enhances physician market power.

1. Introduction

Economists have long understood that vertical integration – common ownership of producers of complementary services – may have both pro- and anti-competitive effects (Joskow 2005). In markets for health services, vertical integration is particularly important. There is almost universal agreement that closer relationships between providers of complementary services have the potential to improve communication and coordination, thereby enhancing the quality or reducing the cost of care. At the same time, there is also evidence that vertical integration may enhance providers’ market power and increase prices (Post, Buchmueller, and Ryan 2018).

This literature, however, has focused almost exclusively on integration between hospitals and physicians; the phenomenon of integration of physicians of different specialties has received significantly less attention. The lack of research on the consequences of multispecialty practice for prices and competition is surprising, since the majority of physicians work in a multispecialty practice and the share is increasing (Welch et al. 2013). In theory, the same mechanisms through which hospital/physician integration operates may also determine the consequences of integration among physicians.

This paper seeks to fill this gap. We study empirically the relationships between prices paid by commercial health plans to generalist and specialist physicians when they are in practices that also include their complementary colleagues. Holding fixed generalist market concentration, we find that generalists charge higher prices when they are integrated with specialists, and that the effect of integration is larger in concentrated specialist markets. We find the same thing in the reciprocal setting – specialist prices are higher when they are integrated with generalists, and the effect is larger in concentrated generalist markets. Our results suggest that multispecialty practice enhances physician market power.

Our paper proceeds in five parts. Part 2 discusses previous research on the effects of vertical integration on competition in markets for health care, focusing on the theoretical reasons why integration between generalists and specialists may enhance market power and increase prices, and describes the analytic approach we take. Part 3 presents our empirical models; our indices of prices of generalists and specialists using data from the Health Care Cost Institute (HCCI); our indices of concentration and integration between generalists and specialists using data from Medicare; and the characteristics of our sample. Part 4 presents our results, and Part 5 concludes.

2. Previous research and analytic approach

Economic theory offers several hypotheses about the effects of vertical integration in general (Bresnahan and Levin 2012), and multispecialty practice in particular, on markets for health services. One class of models suggests that multispecialty practice could improve productive efficiency. These models emphasize how multispecialty practice improves physicians’ ability to coordinate care and referrals through “one-stop shopping” (Burns, Goldsmith, and Sen 2013). A second class of models shows how vertical integration can reduce double marginalization. In the classic version of this model, a non-integrated “upstream” supplier chooses its markup without considering the effect on the profits of the “downstream” producer. If both parties have some market power, the supplier’s markup will exceed the jointly-profit-maximizing one. An integrated entity, however, will effectively price the upstream input at its marginal cost, thereby increasing efficiency (Spengler 1950). More recent work, discussed in Bresnahan and Levin (2012), explains how models of double marginalization apply in general to producers of complementary products who have some market power – such as physicians in a multispecialty practice – even if neither party actually buys from or sells to the other.

A third class of models hypothesizes that physicians may join multispecialty practices in order to make or receive hidden payments for referrals (Pauly 1979). Explicit “kickback” payments to referring physicians are banned by law in the Medicare and Medicaid programs, and banned by contract in essentially all commercial health plans, but enforcement of a ban on transfers across physicians within a practice is difficult or impossible. Models of hidden kickbacks therefore predict that multispecialty practices might couple a policy of referring patients within the practice with higher generalist prices, in order to compensate referring generalists for the extra revenue that they generate for referred-to specialists.

A fourth class suggests several channels through which multispecialty practice could enhance market power. Some of these models explain how vertical integration can be used in the same way as tying in order to facilitate price discrimination (Perry 1989). According to these models, integration could enable generalist and specialist physicians to extract greater rents from consumers in aggregate if consumers who value specialists more highly also demand more frequent generalist visits. Other models explain how vertical integration can foreclose competition (e.g., Whinston 2006). According to these models, generalist/specialist integration could enhance specialists’ market power and increase prices if it enables specialists to deprive their rivals of the referrals from the generalists with whom they affiliate.

In markets with purchasing intermediaries, two models show how vertical integration can enhance market power when the integrating parties share customers. These models are directly relevant to multispecialty practice. Health insurers typically act as purchasing intermediaries in markets for physician services, and physicians of different specialties in a multispecialty practice typically share patients. Peters (2014) shows how integration can enhance the integrating parties’ ability to deny the intermediary’s customers the benefits of their complementarity – benefits which the parties can jointly offer to one of the intermediary’s competitors. Dafny, Ho, and Lee (2019) show that the if the intermediary’s objective function is submodular (i.e., the value to the intermediary of either integrating party is higher if the other party refuses to sell to the intermediary), then the intermediary will suffer a larger profit reduction if both parties leave its network than the combined sum of profit reductions that would arise from removing each party separately. In both of these models, joint bargaining among physicians of different specialties in a multispecialty practice raises prices in much the same way as joint bargaining among competing physicians: it allows physicians to internalize and profit from contracting externalities which would have otherwise flowed to consumers’ benefit.

To investigate the validity of these theories, we begin by testing whether the prevalence of multispecialty practice in a small geographic area affects the price for a standard generalist physician office visit, holding constant the characteristics of the market for generalist physician services; patient characteristics; the procedure-code mix of office visits; the cost of medical care; the mix of patient age, gender, and plan types; and specialty*time- and specialty*area-fixed effects (thereby identifying the effects of interest using only within-specialty variation). We then perform an analogous test of whether the prevalence of multispecialty practice affects the price of a standard specialist physician office visit, holding constant the characteristics of the market for specialist physician services and the other factors listed above.

If the quality of physician services in multispecialty and other practices were identical, then higher prices for both generalists and specialists in multispecialty practices would be evidence that multispecialty practice transfers surplus from patients to physicians and enhances physician market power.

However, because we do not observe quality, we cannot definitively rule out the possibility that multispecialty practice benefits consumers, even if prices for both generalist and specialist services in multispecialty practices were higher. To investigate this possibility, we test for a second effect on prices: the interaction between physicians’ integration with complementary colleagues and the concentration of the complementary colleagues’ market, holding constant the concentration of the physicians’ own market and all of the factors described above. Such an effect is consistent with all of the models that predict that multispecialty practice enhances market power. Conversely, such an effect is inconsistent with all of the models that predict that multispecialty practice benefits (or at least does not harm) consumers, unless the quality advantage of a physician in a multispecialty practice was larger when the physician’s complementary colleague’s market was more concentrated. Under the assumption that no such differential quality advantage to multispecialty practice exists, the presence of a positive interaction effect rejects the hypothesis that multispecialty practice does not enhance physician market power.

Our analysis begins with models of the form:

$$p_{j,t}^s={\beta }_{1}HH{I}_{j,t-k}^s+{\beta }_2{d}_{j,t-k}^{-s}+{\beta }_3(d_{j,t-k}^{-s}\times HH{I}_{j,t-k}^{-s})+{\delta }_s{\gamma }_z+{\delta }_s{\theta }_t+{\epsilon }_{j,t}^s$$ (1)

where p_j,t^s is the price of physician services -- standardized for patient characteristics, procedure mix, and input costs; HHI_j,t-k^s is the Herfindahl-Hirschman index; s = (generalist, specialist); j = (1,…,J) indexes practices; and t = (1,… T;) indexes time. The HHI is a standard measure of the competitiveness of markets, here defined as the sum of the squared market shares of practices serving a market. The HHI will approach 0 for markets served by a large number of practices, each with a small market share. The HHI increases as the number of practices serving the market declines, and when the market share of any one practice increases relative to the others. The HHIs in this paper are scaled so they reach a maximum of 1 in a monopoly market. As we explain in detail below, we computed the HHIs using an approach that we have used in previous work, modified to incorporate the FTC and DOJ recommendations for measuring the market competitiveness facing Accountable Care Organizations (Baker et al. 2014; Federal Trade Commission 2011).

For s = generalist, we specify p_j,t^s and HHI_j,t-k^s as (1 x 2) vectors containing prices and HHIs for physicians reporting that their specialty is family practice/general practice or internal medicine; for s = specialist, we specify p_j,t^s and HHI_j,t-k^s as (1 x 10) vectors containing prices and HHIs for physicians reporting that their specialty is neurology, gastroenterology, hematology/oncology, urology, general surgery, otolaryngology, cardiology, dermatology, endocrinology, and orthopedics.¹

We specify HHI_j,t-k^−s and d_j,t-k^−s as scalars, defined over specialties complementary to s. For s = generalist, we define the ten specialist specialties above as complementary; for s = specialist, we define family practice/general practice and internal medicine as complementary. We allow HHI and d to affect prices with a lag of k years to account for the fact that prices at time t will generally be a function of negotiations with health plans conducted in earlier years; in our basic model we use a three-year moving average lag structure, but also present results from models that use a simple one-year lag. In particular, for s = generalist, d_J,t-k^−s = 1 for practices in which any specialist is present (= 0 otherwise), and HHI_j,t-k^−s = the patient-visit-weighted average of the HHIs of the 10 specialist specialties present in the practice; for s = specialist, d_j,t-k^−s = 1 for practices in which any generalist is present (= 0 otherwise), and HHI_j,t-k^−s = the patient-visit-weighted average of the HHIs of the family practice/general practice or internal medicine physicians present in the practice, γ captures fixed differences across zip codes of patient residence; δ captures fixed differences across specialties; θ captures differences across years; and ε is a mean-zero error term.

In this model, β₁ measures the effect of own-specialty HHI on prices, β₂ the effect of integration with complementary specialties, and β₃ the incremental effect of integration for physicians in practices in which complementary specialties’ markets are highly concentrated. This model assumes that the concentration of the market for physicians of complementary specialties does not affect p_j,t^s, unless they are part of practice j. Indeed, single-specialty practices do not have a defined value for HHI^−s. We examine the sensitivity of our results to this assumption below.

We expand our basic model in two ways. First, we allow the effect of integration to vary with the concentration of a physician’s own market as well as the concentration of the market for complementary physicians:

$$\begin{gathered} p_{j,t}^s={\beta }_{1}HH{I}_{j,t-k}^s+{\beta }_2{d}_{j,t-k}^{-s}+{\beta }_3(d_{j,t-k}^{-s}\times HH{I}_{j,t-k}^s)+{\beta }_4(d_{j,t-k}^{-s}\times HH{I}_{j,t-k}^{-s}) \\ +{\delta }_s{\gamma }_z+{\delta }_s{\theta }_t+{\epsilon }_{j,t}^s \end{gathered}$$ (2)

Second, we estimate models derived from (1) and (2) based exclusively on specialist physicians. In these models, s = (one of the 10 specialist specialties above, the other nine specialist specialties above); p_j,t^s and HHI_j,t-k^s are therefore scalars, as are HHI_j,t-k^−s and d_j,t-k^−s, with HHI_j,t-k^−s and d_j,t-k^−s defined for each specialist specialty over the other nine specialist specialties.

Three recent papers have examined a closely-related question: whether physician/hospital integration enhances provider market power. Baker, Bundorf, and Kessler (2014) construct county-level indices of the price of hospital services using claims from the nonelderly privately insured from Truven Analytics for 2001-07, matched with data on the types of relationships hospitals have with physicians from the American Hospital Association. They find that increases in the market share of hospitals that own physician practices are associated with higher hospital prices and spending. Neprash et al. (2015) construct MSA-level indices of inpatient and outpatient spending using Truven claims for 2008-12, matched with data on the extent of physician/hospital integration based on Medicare outpatient billing patterns. They find that increases in the market share of hospitals that are integrated with physicians are associated with statistically significantly higher outpatient spending. Capps, Dranove, and Ody (2018) show that integration of a physician’s practice with a hospital was associated with price increases for the integrated physician group of 14 percent on average, also holding constant other market characteristics, with larger effects when the integration was undertaken by a larger hospital. Whether vertical integration among complementary physicians has similar effects is the question to which we now turn.

3. Empirical approach

Ideally, we would estimate equations (1) and (2) with data at the physician practice level on prices, integration, and concentration. However, we were unable to obtain complete data on physician prices that identified practice group, and so were unable to match to prices information at the practice level on the extent of integration across specialties. Lacking practice-level data on prices and integration, we estimate our models instead at the ZIP code level. We construct ZIP code-year level generalist and specialist physician price indices for 2008-12 using data from the Health Care Cost Institute (HCCI). HCCI is an independent, nonprofit research institute that operates with the goal of advancing knowledge on health care use and spending in the US. The HCCI data include information from Aetna, Humana, and United HealthCare on approximately 40 million individuals from all 50 states, accounting for 27% of the nonelderly population covered by private health insurance, and has been used by other researchers to study the effects of market structure on prices (Cooper et al. 2019).

Estimates of the effect of concentration and integration from models at the ZIP code rather than the practice level could be subject to aggregation bias (Stoker 2016). Aggregation can result in bias if the aggregated unit’s responsiveness (i.e., the magnitude of the coefficient of interest) is correlated with the unit’s exposure to the independent variable. In order to rule out the possibility of aggregation bias, we assume that concentration and integration affect prices in the same way across practices, which Stoker (2016) describes as micro-linearity or “exact” aggregation.

We match to these indices ZIP code-year level measures of the characteristics of physician practices and physician markets based on Medicare Part B claims filed by physicians for the care of a 20% random sample of traditional Medicare beneficiaries. Medicare claims reflect care delivered by a very large share of active physicians and we expect the set of physicians who billed traditional Medicare to substantially overlap with the set of physicians who provided services to privately insured patients in the HCCI data.² Each claim reports the tax ID number (TIN) of the physician’s practice, the physician’s specialty, and the physician’s and patient’s zip codes, among other things. We define a physician’s practice as the set of physicians who file claims under a common TIN. This approach has been used in other work by some of us and others (Federal Trade Commission 2011; Welch et al. 2013; Baker et al. 2014; Clemens and Gottlieb 2017). Physicians who use the same tax ID are part of the same financially-integrated organization. Many financially-integrated organizations use the same tax ID for all physicians in their organization, though a single organization may use multiple tax IDs.

Based on this definition, we calculate at the ZIP code-year level the prevalence of generalist/specialist integration and HHI’s of the concentration of generalist and specialist physician markets. The models we estimate are of the form:

$$p_{z,t}^s={\beta }_{1}HH{I}_{z,t-k}^s+{\beta }_2{d}_{z,t-k}^{-s}+{\beta }_3(d^{-s}\times HH{I}^{-s}{)}_{z,t-k}+{\delta }_s{\gamma }_z+{\delta }_s{\theta }_t+{\epsilon }_{z,t}^s$$ (1’)

and

$$\begin{gathered} p_{z,t}^s={\beta }_{1}HH{I}_{z,t-k}^s+{\beta }_2{d}_{z,t-k}^{-s}+{\beta }_3(d^{-s}\times HH{I}^{-s}{)}_{z,t-k}+{\beta }_4(d^{-s}\times HH{I}^s{)}_{z,t-k} \\ +{\delta }_s{\gamma }_z+{\delta }_s{\theta }_t+{\epsilon }_{z,t}^s \end{gathered}$$ (2’)

where z = (1,…Z), and p, HHI^s, d^−s, and (d^−s × HHI^s), (d^−s × HHI^s) are ZIP-code level averages constructed as described below.

Intuitively, this specification measures at the area level the effect of the concentration of a physician’s market, the prevalence of integration into multispecialty practice, and the incremental effect of integration in areas where the integrated physicians practice with complementary specialties that have concentrated markets. If all of the practices serving patients in one ZIP are multispecialty, so d^−s = 1, variations in HHI^−s are relevant for all prices in the ZIP. If in another ZIP, half of the practices are multispecialty, so d^−s = 0.5, then the effects of HHI^−s only affect half of the prices in the price index.

3.1. Calculating the price index

We focus on prices paid for office visits with new patients (Current Procedural Terminology [CPT] codes 99201-99205) and established patients (CPT codes 99211-99215). These 10 codes are the most commonly-billed in the United States, and unlike other procedures, represent a service that is common to all specialties. Codes 99201-99205 include all office visits for new patients not previously seen by the physician. These 5 services are intended to span the range of intensity across office visits from 99201, a “basic” office visit with minimal patient complexity and short duration, to 99205, an “advanced” office visit with a high degree of complexity and a long duration. The range 99211-99215 similarly captures the range of office visits for established patients who have been previously seen by the physician.

To construct our ZIP-code level price index, we began by extracting from the HCCI database claims for office visits provided by physicians who reported on the claim that their specialty is family practice/general practice or internal medicine (for generalist physicians), and neurology, gastroenterology, hematology/oncology, urology, general surgery, otolaryngology, cardiology, dermatology, endocrinology, and orthopedics (for specialist physicians). We included only fee-for-service claims from health plans identified as a commercial PPO, EPO, POS, or HMO plan, where the claim was for a single service, the place of service was specified as a physician office, the claim was for professional (as opposed to facility) fees, the CPT modifier code was either blank or had the value 25,³ and the claim reported the practice ZIP code of the physician providing the service.

For each claim, we obtain the allowed amount, which is the amount the physician agreed to accept for the service after the application of contractual discount provisions and other plan rules. We refer to this as the “price” for the service. The physician may have received this partly from the plan and partly from the patient in the form of applicable copayments, coinsurance, or deductibles. We retained for analysis claims that had a positive allowed amount. We excluded claims when there appeared to be a billing error.⁴ Separately for each CPT code each year, we dropped as outliers claims where the allowed amount fell in the top 1% or bottom 1% of claims for the code.

We matched the claims to HCCI enrollment data to obtain information about the age, gender, and residence ZIP code of the patient. We retained claims where this information was nonmissing,⁵ the patient was age 0-64, and where the enrollment data indicated that the patient was enrolled in commercial insurance and not also enrolled in Medicare. We also obtained the Medicare Geographic Practice Cost Index (GPCI) for work, practice expenses, and malpractice, by ZIP code by year, matched this to the claims based on physician practice ZIP code, and dropped the small number of cases where GPCIs could not be matched to the listed physician ZIP. This resulted in a base of approximately 181 million claims for generalist physicians and 94 million claims for specialists, with characteristics shown in Table 1.

Table 1:

Characteristics of HCCI Physician Claims Used in Analysis

	Generalists	Specialists
Patient age
0-17	0.0822	0.1288
18-24	0.0703	0.0600
25-34	0.1456	0.1062
35-44	0.2049	0.1637
45-54	0.2575	0.2475
55-64	0.2395	0.2938
Gender
Female	0.5782	0.5595
Plan Type
HMO	0.1215	0.1285
POS	0.6494	0.6517
EPO	0.0790	0.0754
PPO	0.1502	0.1374
Procedure code
99201	0.0014	0.0059
99202	0.0163	0.0461
99203	0.0401	0.1072
99204	0.0220	0.0413
99205	0.0061	0.0097
99211	0.0159	0.0155
99212	0.0402	0.0983
99213	0.5139	0.4002
99214	0.3163	0.2383
99215	0.0278	0.0374
Modifier codes
None	0.9238	0.8776
25	0.0762	0.1224
GPCI
Work	1.0136	1.0169
Practice Expense	1.0181	1.0411
Malpractice	1.0341	1.1122
Year
2008	0.2078	0.1902
2009	0.2147	0.2008
2010	0.1987	0.2027
2011	0.1923	0.2032
2012	0.1865	0.2031
Number of claims	180,652,033	93,790,666

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. Table entries are shares except for GPCI indices.

We then estimate the following regression separately for each specialty and year:

$$\text{in}(p_{k,i,z,u})={\beta }_0+{\beta }_{1}AG{E}_i+{\beta }_{2}SE{X}_i+{\beta }_{3}PLA{N}_i+{\beta }_4(PRO{C}_i\times MO{D}_i)+{\beta }_{5}GPC{I}_u+{\gamma }_z+{\epsilon }_{k,i,z,u}$$ (3)

where p_k,i,z,u is the price paid for claim k for patient i resident in ZIP code z by a provider in ZIP code u. AGE is a vector of dummy variables for patient age in 6 categories (0-17, 18-24, 25-34, 35-44, 45-54, 55-64); SEX is a dummy variable for patient sex; PLAN is a vector of controls for plan type (HMO, POS, EPO, PPO); PROC×MOD is a set of interactions between procedure and modifier codes; GPCI is a vector of the three GPCIs for work, practice expenses, and malpractice; and γ is a vector of fixed effects for ZIP codes.⁶ The estimates of the ZIP fixed effects γ for each ZIP code-specialty-year define our adjusted ZIP code level price index p_z,t^s.

3.2. Calculating market characteristics

Our analytic approach follows that in Baker et al. (2014), which adapts the approach of Kessler and McClellan (2000) to the case of physician practices. Using Medicare data as described above, we calculate ZIP-code level market characteristics in three steps. In the first step, we construct an HHI for each ZIP code-specialty-year equal to the sum of squared market shares of billed charges C for each practice serving the ZIP code:

$$ZIPHH{I}_{z,t}^s=\sum _{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingZIPz\hfill \end{array}}{\left[\frac{C_{j,z,t}^s}{{\sum }_{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingzipz\hfill \end{array}}{C}_{j,z,t}^s}\right]}^2$$

These ZIP HHIs do not impose any a priori market size; they are instead based on the set of physicians who actually provide services to patients in each ZIP code. We exclude from this calculation claims where the physician is more than 100 miles from the patient ZIP, to reduce the potential for mismeasurement from patients who, perhaps while traveling, visit a distant physician who does not play a substantial role in competition for patients residing in the ZIP code.

Using this measure of concentration directly would assume that physician practices differentiate among patients based on their ZIP code of residence. More realistically, physician pricing decisions depend on the total demand for services from the practice’s core market area. Thus we take a second step to construct practice-level market characteristics with this feature. In order to do this, we identified for each practice j a market area equal to the nearest set of patient ZIP codes served by the practice that accounted for 75% of the practice’s billed charges.⁷ We averaged the patient ZIP code HHIs for the ZIP codes in the market area, weighting by the charges incurred by patients from that ZIP at the practice-specialty-year:

$$HH{I}_{j,t}^s=\sum _{\begin{array}{c}\hfill zipszin\hfill \\ \hfill practicejarea\hfill \end{array}}\left[\frac{C_{j,z,t}^s}{{\sum }_{\begin{array}{c}\hfill zipszin\hfill \\ \hfill practicejarea\hfill \end{array}}{C}_{j,z,t}^s}\right]\times ZIPHH{I}_{z,t}^s$$

The third and final step in our analysis is to construct patient-ZIP-code-area market characteristics equal to the average practice-level HHI across the practices serving the ZIP code, weighted by the share of office visits V at the practice experienced by patients resident in the ZIP code:

$$HH{I}_{z,t}^s=\sum _{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingZIPz\hfill \end{array}}\left[\frac{V_{j,z,t}^s}{{\sum }_{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingzipz\hfill \end{array}}{V}_{j,z,t}^s}\right]\times HH{I}_{j,t}^s$$

We weighted practice-level HHIs in this calculation by visits rather than billed charges so that the aggregation of practice HHIs was similar to the aggregation implicit in the computation of the patient ZIP-code level price indices from the HCCI data.

We calculate d^−s – our key measure of integration -- as the share of practices serving patients in a ZIP code in which a complementary specialty is present, weighted by the number of office visits by ZIP-code residents to physicians in each practice:

$$d_{z,t}^{-s}=\sum _{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingZIPz\hfill \end{array}}\left[\frac{V_{j,z,t}^s}{{\sum }_{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingzipz\hfill \end{array}}{V}_{j,z,t}^s}\right]\times d_{j,t}^{-s}$$

We similarly computed the ZIP-code level weighted average of (d^−s × HHI^−s) over practices serving patients resident in the ZIP code with non-missing values for HHI^−s:⁸

$$(d^{-s}\times HH{I}^{-s}{)}_{z,t}=\sum _{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingZIPz\hfill \end{array}}\left[\frac{V_{j,z,t}^s}{{\sum }_{\begin{array}{c}\hfill practicesj\hfill \\ \hfill servingzipz\hfill \end{array}}{V}_{j,z,t}^s}\right]\times d_{j,t}^{-s}\times HH{I}_{j,t}^{-s}$$

3.3. Sample Construction

For each specialty, we excluded ZIP codes meeting any of the following conditions: a missing price index or area market characteristics in any year; a price index in the top 1% or bottom 1% of the distribution; fewer than 10 HCCI office visit claims underlying the price index; and fewer than 20 Medicare beneficiaries underlying the construction of the practice characteristics.

Tables 2 and 3 report ZIP code level descriptive statistics for the key variables for generalists and specialists, respectively, for 2012. Table 2 is based on 32,194 observations (16,705 for family practice/general practice and 15,489 for internal medicine) covering 17,284 ZIP codes with at least one type of generalist; table 3 is based on 92,416 observations covering 13,505 ZIP codes with at least one type of specialist (with the distribution across the 10 specialties reported in the table). The statistics in the table were calculated by weighting each ZIP code by the number of HCCI claims underlying the price index. The price indices, by construction, are very close to zero (they are not identically zero because of the sample trimming discussed in the paragraph above). The market for generalist physicians is relatively unconcentrated, with an average HHI equal to 0.1396 (Table 2). The overall lack of concentration in the market for generalist physician services, however, masks differences across areas; in the least concentrated quartile, the average HHI for generalists is 0.0564, but in the most concentrated quartile, the HHI is almost five times as large (0.2639). The market for specialists is, not surprisingly, more concentrated (average HHI = 0.2753), ranging from 0.1329 in the least concentrated quartile to 0.4850 in the most concentrated quartile (Table 3).

Table 2:

Descriptive Statistics For Key Variables, 2012 Levels Markets for Generalist Physician Services

	Mean	Standard Deviation	Mean in bottom quartile	Mean in top quartile
Price index	0.0034	0.1725	-0.1932	0.2358
Own-specialty HHI	0.1396	0.0912	0.0564	0.2639
Integration with Specialist	0.4034	0.2290	0.1371	0.7252
Integration with specialist*mean(spec HHI)	0.1166	0.0966	0.0280	0.2531
Integration with specialist* own HHI	0.0723	0.0836	0.0121	0.1842

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. N = 32,194 = 16,705 ZIP codes for family practice/general practice + 15,489 ZIP codes for internal medicine. Table statistics weight each ZIP code/physician type by the number of claims underlying the price index.

Table 3:

Descriptive Statistics for Key Variables, 2012 Levels Markets for Specialist Physician Services

	Mean	Standard deviation	Mean in bottom quartile	Mean in top quartile
Price index	-0.0028	0.1562	-0.1828	0.2078
Own-specialty HHI	0.2753	0.1488	0.1329	0.4850
Integration with generalist	0.3218	0.2735	0.0449	0.7280
Integration with generalist * mean(generalist HHI)	0.0440	0.0581	0.0043	0.1195
Integration with other specialist	0.2214	0.2189	0.0302	0.5396
Integration with other specialist*mean(HHI oth spec)	0.0598	0.0758	0.0066	0.1607
Integration with generalist * own HHI	0.1037	0.1367	0.0091	0.2902
Integration with other specialist* own HHI	0.0695	0.1074	0.0063	0.2018
	share of claims		# of zips 8542
neurology	0.0544		8542
gastroenterology	0.0699		9150
hematology/oncology	0.0602		7699
urology	0.0619		9622
general surgery	0.0475		8956
otolaryngology (ENT)	0.0865		8698
cardiology	0.0941		11695
dermatology	0.2645		11183
endocrinology	0.0434		4853
orthopedics	0.2177		12018

Note: Generalists include family practice/general practice and internal medicine. N = 92,416 ZIP code*specialties. Table statistics weight each ZIP code/physician type by the number of claims underlying the price index.

Consistent with previous descriptive work, integration is relatively common in our sample: 40.34 percent of generalists are integrated with a specialist (Table 2), and 32.18 percent of specialists are integrated with a generalist (Table 3). Integration of specialists with specialists of another specialty is slightly less common (22.14 percent, Table 3), indicating that multispecialty practices that include a generalist in many cases do not include all of the 10 types of specialists we examine. Appendix Tables 1 and 2 report the matrix of correlations among the key independent variables. The tables show that, especially for generalist physicians, integration is positively correlated with generalist market concentration: the correlation between the integration rate and own-specialty HHI is 0.5307. This highlights the importance of controlling for both concentration and integration in order to assess the independent effect of each on competition and prices. Appendix Tables 3 and 4 report descriptive statistics for generalists and specialists, analogous to those in Tables 2 and 3, but for 2008-2012 changes instead of 2012 levels. Appendix Tables 3 and 4 show that there is economically important variation over time in both the dependent and independent variables in our analysis. For example, in the bottom quartile of ZIP codes (ranked by 2008-2012 changes), the price index for generalist visits fell by 0.1019, but in the top quartile, the price index rose by 0.1026 (Appendix Table 3). Integration of generalists with specialists rose overall by 0.0671 from 2008-2012 (Appendix Table 3), but the aggregate increase was composed of ZIP codes with modest decreases in integration (bottom quartile average = −0.0585) and ZIP code with large increases (top quartile average = 0.2208).

4. Results

Table 4 presents estimates from equations (1’) and (2’) for generalist physicians. The Table reports the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile. All models are based on balanced panels for the years 2008-2012. The Table reports standard errors allowing for arbitrary correlation within a ZIP code and spatial correlation across ZIP codes within a radius of 25km (Conley 1999; Hsiang 2010). Columns (1)-(4) report results from different variants of equation (1’). Column (1) imposes the constraint β₂ = β₃ = 0; we report an estimate from this specification for purposes of comparison to the previous literature. Column (1) shows that an increase in the HHI from 0.0564 (mean in the bottom quartile, table 2) to 0.2639 (mean in the top quartile) is associated with an increase in the generalist price index of approximately 4.2 percent. This is comparable to, although slightly smaller than, the effect of generalist physician market concentration on prices found in Baker et al. (2014). Column (2) relaxes the constraint that β₂ = 0, i.e., allows own- specialty concentration and integration to have separate effects. Integration with a specialist has an economically important and statistically significant positive effect on generalist prices. Moving a generalist from a relatively unintegrated area (claim-weighted integration rate of 0.1371, table 2) to a relatively integrated one (0.7252) would allow her to increase her prices for a standard office visit by approximately 2.7 percent. This implies a practice-specific effect of integration on generalist prices of 4.6 percent (0.027 / (0.7252 – 0.1371)). Controlling for integration reduces the estimated effect of own-specialty concentration to approximately 3.5 percent for a bottom-to-top-quartile change in HHI, suggesting that a substantial fraction of previous estimates of the effect of own-specialty concentration may have been due to the effects of integration rather than concentration per se.

Table 4:

Effect of Concentration and Integration with Specialist Physicians On Prices of Generalist Physician Services (standard errors in parentheses)

	(1)	(2)	(3)	(4)	(5)
Own-specialty HHI	0.0417 (0.0068)	0.0345 (0.0068)	0.0308 (0.0068)	0.0307 (0.0067)	0.0393 (0.0086)
Integration with specialist		0.0267 (0.0052)	0.0088 (0.0101)	0.0093 (0.0101)	0.0142 (0.0114)
Integration with specialist*mean(spec HHI)			0.0235 (0.0097)	0.0227 (0.0096)	0.0282 (0.0096)
Integration with specialist*own HHI					−0.0141 (0.0084)
Mean(specialist HHI)				0.0025 (0.0057)

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. Sample period is 2008-12. N = 160,970 = 5 years * (16,705 ZIP codes for family practice/general practice + 15,489 ZIP codes for internal medicine). All models include specialty*zip code and specialty*year fixed effects. Standard errors clustered at the zip code level (N ZIP codes = 17,284, some ZIP codes are not served by all types of generalists) allowing for spatial correlation across zip codes within a radius of 25km. Table entries are the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile.

Column (3) estimates the fully-specified version of equation (1’). It shows that the effect on generalist prices of integration with a specialist increases with specialist concentration. Moving a generalist from a relatively unintegrated area to a relatively integrated one, when the markets in both areas for the specialists with whom she integrated were unconcentrated (HHI = 0), is associated with a small and statistically insignificant effect on her prices; but allowing the interaction between integration and specialist concentration also to increase from the bottom to the top quartile is associated with an additional 2.4 percent increase. Unless there is unobserved time-varying heterogeneity across ZIP code level markets in the determinants of the price of a standard generalist office visit that is correlated with generalist/specialist integration in concentrated specialist markets, and not due to market power, this estimate represents a rejection of the null hypothesis that multispecialty practice does not enhance market power.

Using the information in table 2 this coefficient can be rescaled in a number of ways. For example, moving a generalist from a relatively unintegrated area to a relatively integrated area (change in integration of 0.5881 = 0.7252 – 0.1371) in a monopolized specialist market (mean specialist HHI = 1) would increase prices by an additional 6.1 percent = (0.5881 * (0.0235 / (0.2531 – 0.0281)), relative to the increase in an unconcentrated specialist market (mean specialist HHI = 0).

Column (4) presents estimates from an extension to (1’) that also includes a control for specialist HHI not interacted with integration. As we explain above, the models underlying (1’) assume that concentration in the market for specialist physicians does not affect the price of services of generalist physicians, except insofar as those physicians work together in an integrated practice. Indeed, a generalist physician working in a practice without specialists does not have a defined value for HHI^−s. Column (4) confirms that the results reported in column (3) are not sensitive to this assumption, and are consistent with integration enhancing physician market power. The magnitude of the effect of the interaction between specialist HHI and integration does not change materially when uninteracted specialist HHI is included – and uninteracted specialist HHI has a small and statistically insignificant effect on generalist prices in the absence of integration.

Column (5) presents estimates of equation (2’), which allow the effect of integration to vary both with the generalist’s own HHI and the average HHI of specialists in the generalist’s multispecialty practice. The estimates in column (5) show that the positive effect of integration on generalist prices falls off as the concentration of the generalist’s market increases. A bottom-to-top quartile change in the interaction between integration and the generalist’s own HHI is associated with an effect of half the magnitude and opposite sign of the effect of the interaction between integration and average specialist HHI (−1.4 percent versus 2.8 percent). This finding can be accommodated by each of the four classes of models of market-power-enhancing effects of integration discussed above, none of which require the effect of integration to vary in any particular way with the concentration of the own-physician’s market.

Table 5 presents estimates from equations (1’) and (2’) for specialist physicians, with columns defined analogously to those in Table 4. Table 5 shows that the standardized effect of own-specialty HHI for specialists, approximately 4.8 percent (column (2), controlling for the extent of integration with generalist physicians) is larger than the effect of own-specialty HHI for generalists. The larger magnitude of the effect is due to a wider interquartile range of specialist market concentration (0.3521 = 0.4850 – 0.1329 for specialists, table 3, versus 0.2075 = 0.2639 – 0.0564 for generalists, table 2) rather than a larger unit effect of concentration on specialist prices (compare 0.1375 = 0.0484 / 0.3521 to 0.1663 = 0.0345 / 0.2075). By contrast, the average effect of integration is very similar between specialists and generalists (compare column (2), table 5 to column (2), table 4). This is consistent with insurers’ benchmarking multispecialty practice prices off of those for single-specialty practices in the same market (e.g., Clemens and Gottlieb 2017).

Table 5:

Effect of Concentration and Integration with Generalist Physicians On Prices of Specialist Physician Services (standard errors in parentheses)

	(1)	(2)	(3)	(4)	(5)
Own-specialty HHI	0.0504 (0.0057)	0.0484 (0.0055)	0.0467 (0.0056)	0.0468 (0.0056)	0.0596 (0.0060)
Integration with generalist		0.0200 (0.0032)	0.0063 (0.0053)	0.0103 (0.0050)	0.0307 (0.0070)
Integration with generalist * mean(generalist HHI)			0.0168 (0.0048)	0.0109 (0.0043)	0.0217 (0.0051)
Integration with generalist * own HHI					−0.0293 (0.0068)
Mean(generalist HHI)				0.0178 (0.0059)

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. Sample period is 2008-12. N = 462,080 = 5 years * 92,416 ZIP code*specialties. All models include specialty*zip code and specialty*year fixed effects. Standard errors clustered at the zip code level (N ZIP codes = 13,505, some ZIP codes are not served by all types of specialists) allowing for spatial correlation across zip codes within a radius of 25km. Table entries are the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile.

Analogous to Table 4, column (4) presents estimates from an extension to (1’) that also includes a control for generalist HHI directly. Column (4) confirms that, although the magnitude of the interaction between integration and generalist HHI declines, the results in column (3) are not due to the concentration of the market for generalists per se or one of its correlates. Column (5) presents estimates of equation (2’) for specialists, showing that as with generalists, the effect of integration declines with the concentration of the specialist’s own-specialty’s market.

Table 6 presents estimates of equations (1’) and (2’) based exclusively on specialist physicians. In these models, instead of s = (generalist, specialist), we define s = (one of the 10 specialist specialties above, the other nine specialist specialties above). These models therefore identify the effect on specialist prices of integration among specialists of different specialties rather than the effect of integration with a generalist. Column (1) shows that integration with another specialist is associated with higher specialist prices, of similar magnitude to integration with a generalist (compare to table 5, column (2)). Column (2) shows that the effect of specialist/specialist integration, like the effects of generalist/specialist (table 4) and specialist/generalist (table 5) integration, is driven by concentration in the complementary markets. Column (3) of table 6 presents estimates that include a control for complementary specialties’ HHI uninteracted with integration as well as the interacted effect. As in the models generalist/specialist integration (table 4), the complementary specialties’ market concentration has a small and statistically insignificant effect on the physician’s own price in the absence of integration, and the effect of the interaction between complementary specialties’ HHI and integration does not change materially when uninteracted complementary specialties’ HHI is included.

Table 6:

Effect of Concentration and Integration with Other Specialist Physicians On Prices of Specialist Physician Services (standard errors in parentheses)

	(1)	(2)	(3)
Own-specialty HHI	0.0481 (0.0057)	0.0459 (0.0056)	0.0461 (0.0056)
Integration with other Specialist	0.0321 (0.0034)	0.0007 (0.0066)	0.0033 (0.0063)
Integration with other specialist*mean(oth spec HHI)		0.0365 (0.0063)	0.0335 (0.0057)
Mean(oth spec HHI)			0.0091 (0.0061)

Note: Specialist specialties are listed in Table 3. Sample period is 2008-12. N = 462,080 = 5 years * 92,416 ZIP code*specialties. All models include specialty*zip code and specialty*year fixed effects. Standard errors clustered at the zip code level (N ZIP codes = 13,505, some ZIP codes are not served by all types of specialists) allowing for spatial correlation across zip codes within a radius of 25km. Table entries are the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile.

To investigate further the effects of competition and integration on specialist prices, table 7 examines the two specialties with the broadest geographic coverage (Table 3): cardiology and orthopedics. In both cardiology and orthopedics, the price of a standard office visit is responsive to both the specialist’s own-specialty HHI and to integration with a generalist (columns (1) - (2)), although the effect of both own-specialty HHI and integration is much larger for cardiology than for orthopedics. The estimated effects of integration and the interaction between integration and generalist HHI are invariant to the inclusion of a control for generalist HHI uninteracted in cardiology, although in orthopedics the interaction effect declines and becomes statistically insignificant when uninteracted generalist HHI is included as a control (column (3)).

Table 7:

Effect of Concentration and Integration on Prices of Cardiology and Orthopedic Physician Services (standard errors in parentheses)

			cardiology						orthopedics
	(1)	(2)	(3)	(4)	(5)	(6)	(1)	(2)	(3)	(4)	(5)	(6)
Own-specialty HHI	0.0996 (0.0176)	0.0814 (0.0157)	0.0802 (0.0157)	0.0861 (0.0160)	0.0794 (0.0154)	0.0809 (0.0153)	0.0473 (0.0087)	0.0452 (0.0087)	0.0451 (0.0088)	0.0482 (0.0087)	0.0435 (0.0085)	0.0435 (0.0086)
Integration with generalist	0.0403 (0.0082)	0.0081 (0.0126)	0.0130 (0.0125)				0.0129 (0.0050)	−0.0001 (0.0085)	0.0051 (0.0087)
Integration with generalist mean * (generalist HHI)		0.0559 (0.0146)	0.0425 (0.0135)					0.0168 (0.0079)	0.0086 (0.0079)
Integration with other specialist				0.0809 (0.0092)	0.0336 (0.0193)	0.0422 (0.0188)				0.0116 (0.0046)	−0.0290 (0.0103)	−0.0287 (0.0097)
Integration with other specialist*mean(oth spec HHI)					0.0558 (0.0197)	0.0420 (0.0188)					0.0459 (0.0099)	0.0455 (0.0092)
Mean(generalist HHI)			0.0312 (0.0108)						0.0126 (0.0056)
Mean(oth spec HHI)						0.0332 (0.0106)						0.0008 (0.0061)

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. Sample period is 2008-12. Cardiology N = 58,475 = 5 years * 11,695 ZIP codes; orthopedics N = 60,090 = 5 years * 12,018 ZIP codes. All models include zip code and year fixed effects. Standard errors clustered at the zip code level allowing for spatial correlation across zip codes within a radius of 25km. Table entries are the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile.

Columns (4) – (6) of Table 7 replicate the analysis of all specialists together (table 6) for cardiology and orthopedics individually. Column (4) shows that integration with another specialist is associated with higher specialist prices for both cardiology and orthopedics, but especially for cardiology; column (5) shows that the price-increasing effects of integration are driven by concentration in the complementary specialties’ markets. In addition, in these specialties, the estimated effects of integration and the interaction between integration and other-specialist HHI are invariant to the inclusion of a control for other-specialist HHI uninteracted.

Table 8 presents results from variants of equation (1’) in order to explore the robustness of our results. Each pair of columns replicates column (2) of tables (4) and (5), which estimate the effect of own-specialty HHI and the average effect of integration – under various alternative assumptions. Columns (1) and (2) of Table 8 replicate column (2) of tables 4 and 5, respectively, but use as a dependent variable a price index constructed without controls for plan type, procedure, or modifier codes. Estimates of the effect of concentration and integration based on this alternative price index are similar to those using our standard price index (although the effect of own-specialty HHI is higher for generalist physicians). Columns (3) and (4) use a simple one-year lag structure for the control variables instead of a 3-year moving average; the magnitudes of our estimated effects decline but remain statistically significant, suggesting that the 3-year moving average captures the market conditions affecting price with slightly greater accuracy. Columns (5) and (6) substitute the one-year lead of own- specialty HHI and integration for the lag. The effects of the lead of own-specialty HHI are small but remain statistically significant; the effects of the lead of integration are very small and statistically insignificant. Columns (7) and (8) include ZIP-code level controls for inpatient hospital market conditions and the extent of hospital/physician integration used in Baker, Bundorf, and Kessler (2014); the magnitude and significance of our estimates remain virtually unchanged, indicating that our key results are not driven by unrelated characteristics of markets for health services.

Table 8:

Effect of Concentration and Integration on Prices of Physician Services, Alternative Models (standard errors in parentheses)

	Price index adjusted for patient age/sex only		One-year lag structure for regressors		One-year lead structure for regressors		Control for hospital HHI & hospital/physician integration
	generalists (1)	specialists (2)	generalists (3)	specialists (4)	generalists (5)	specialists (6)	generalists (7)	specialists (8)
Own-specialty HHI	0.0269 (0.0081)	0.0546 (0.0093)	0.0162 (0.0047)	0.0183 (0.0036)	0.0149 (0.0040)	0.0164 (0.0038)	0.0303 (0.0067)	0.0484 (0.0056)
Integration with generalist		0.0256 (0.0039)		0.0085 (0.0018)		0.0025 (0.0018)		0.0199 (0.0032)
Integration with specialist	0.0328 (0.0061)		0.0141 (0.0033)		0.0025 (0.0027)		0.0255 (0.0052)
N	160970	504180	160970	504180	160970	504180	152990	453200
number of zip codes	17284	13552	17284	13552	17284	13552	16311	13050

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. Sample period is 2008-12. All models include specialty*zip code and specialty*year fixed effects. Standard errors clustered at the zip code level allowing for spatial correlation across zip codes within a radius of 25km. Table entries are the effect of moving from the mean of the bottom quartile of the listed variable to the mean of the top quartile.

5. Conclusion

The consequences of vertical integration for the competitiveness of markets for health services have been the topic of considerable debate. On one hand, there is almost universal agreement that fragmentation among providers of complementary health services has negative consequences for productive efficiency. On the other hand, a wide range of economic models indicate that vertical integration can be anticompetitive, and recent empirical evidence supports the hypothesis that physician/hospital integration enhances provider market power, leading to increased prices. Yet despite the high and increasing prevalence of multispecialty physician practices – which raise exactly the same possibilities for pro- and anticompetitive effects in theory – no empirical work has examined how vertical integration among physicians affect prices.

In this paper, we estimate the extent to which changes in multispecialty practice from 2008-2012 in approximately 17,000 US ZIP codes affect the price of a standard office visit with a generalist and specialist physician. We obtain data on prices from the Health Care Cost Institute (HCCI) on approximately 40 million individuals from all 50 states, accounting for 27% of the nonelderly insured population, making it one of the largest sources of health insurance claims data ever assembled. We match ZIP-code level price indices from the HCCI data to measures of the characteristics of physician practices and physician markets based on Medicare Part B claims, grouping together physicians based on their receipt of payments under a common Taxpayer Identification Number (TIN). This approach has been used extensively in previous work, because physicians who receive payment under a common TIN are generally part of the same financially-integrated organization. We estimate the effect on generalist physician prices of integration with a specialist and the effect on specialist physician prices of integration with a generalist or another specialist of a different type. Our empirical models control for the concentration of the markets for generalist and specialist physician services; patient characteristics; the procedure-code mix of office visits; the cost of medical care; the mix of patient age, gender, and plan types; and specialty*area- and specialty*time- fixed effects.

We report three key findings. First, generalist physicians charge higher prices when integrated with a specialist, holding constant the factors above. The effect of integration with a specialist on generalist prices is of the same order of magnitude as the effect of the generalist’s own market characteristics. Moving a generalist from a relatively unintegrated area (the average ZIP code in the bottom quartile of integration) to a relatively integrated one (the average in the top quartile) is associated with an increase in prices of approximately 2.7 percent. By comparison, moving a generalist from a relatively unconcentrated area (the average ZIP code in the bottom quartile of concentration, as measured by the generalist’s Hirschman-Herfindahl index) to a relatively concentrated area (the average in the top quartile) is associated with an increase in prices of approximately 3.5 percent. We find the same thing in the reciprocal setting – specialist prices are higher when they are integrated with generalists, although the effect on specialist prices of integration with a generalist is not as large, relative to the magnitude of the effect of the specialist’s own market characteristics. Because integration and own-market concentration are highly positively correlated, our result shows that a substantial fraction of previous estimates of the effect of own-specialty concentration may have been due to the effects of integration rather than concentration per se.

Second, the effect on generalist physician prices of integration with a specialist is larger in concentrated specialist markets (and the effect on specialist physician prices of integration with a generalist is larger in concentrated generalist markets). The presence of a positive interaction effect between integration and the concentration of the market of the complementary physician is important for two reasons. First, it is a consequence of all of the economic models that predict that vertical integration enhances market power. Second, it is likely to be inconsistent with the hypothesis that the positive correlation between integration and price is due to a simple difference in quality between single- and multispecialty group services. In order for differences in quality to explain the positive interaction effects, the unobserved quality of a generalist office visit in a multispecialty practice would have to increase in the extent of specialist market concentration – as would the unobserved quality of a specialist office visit in a multispecialty practice have to increase in the extent of generalist market concentration. Of course, as with any observational analysis, we cannot definitively rule out the possibility of unobservables, but this test – along with other validity checks we discuss above – make such explanations for our results implausible.

These findings also are not plausibly explained by models in which physicians join multispecialty practices in order to make or receive hidden payments for referrals. Models of hidden kickbacks can explain why referring physicians’ unit prices are higher in multispecialty practice. Such models can even explain the positive effect on generalist prices of the interaction between integration and specialist concentration, if specialists with market power valued referrals more than specialists without it. But they cannot explain why multispecialty practice is associated with higher specialist prices, along with a positive effect of the interaction between integration and generalist concentration. Because specialists do not generally refer patients to generalists, payments for referrals (and therefore hidden kickbacks) generally flow from specialists to generalists.

Third, we also find that integration of one specialist with another of a different (complementary) specialty is associated with higher prices. As in our models of specialist/generalist integration, there is a positive average effect on specialist prices of the type of specialist/specialist integration described above, and the effect of this specialist/specialist integration is larger when the complementary specialist’s market is more concentrated. These results hold for the two specialties that we examine individually – cardiology and orthopedics – suggesting that this type of specialist/specialist integration is as important, from the perspective of antitrust, as specialist/generalist integration.

Our analysis has several important limitations. First, our measure of competitiveness (the HHI) is not directly derived from a theoretical model of physician/insurer bargaining, despite the fact that this is process through which the prices we study are determined. Although it is used extensively, the limitations of the HHI as a measure of competitiveness or market power are well-known (Bresnahan 1989). Second, although we used numerous strategies in order to minimize the possibility that our results were due to unobserved differences across geographic areas – including controls for area-fixed- effects that vary by specialty, patient characteristics, and other related market factors; the choice of a standardized service (a simple office visit) that is relatively comparable across specialties, areas, and over time; and a hypothesis test that rules out at least simple differences in unobserved quality between single- and multispecialty practices – our analysis is fundamentally observational in nature and so cannot definitively exclude that our findings are due to the endogeneity of the integration decision, other aspects of physician markets, or unobserved quality more broadly. Third, we do not assess the consequences of multispecialty practice for social or consumer welfare in aggregate. Our analysis is limited to standard office visits only (for the reasons discussed above), but physicians provide many more specialized services that we do not consider. Exploration of the relevance of these limitations is a topic for future research.

Appendix

Appendix Table 1:

Correlations Among Key Independent Variables, 2012 Levels Markets for Generalist Physician Services

	own-specialty HHI	integration with specialist	integration with specialist*mean(spec HHI)	integration with specialist* own HHI
Own-specialty HHI	1.0000
Integration with specialist	0.5307	1.0000
Integration with specialist*mean(spec HHI)	0.7250	0.8796	1.0000
Integration with specialist* own HHI	0.8500	0.7964	0.9010	1.0000

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. N = 32,194 = 16,705 ZIP codes for family practice/general practice + 15,489 ZIP codes for internal medicine. Table statistics weight each ZIP code/physician type by the number of 2012 claims underlying the price index.

Appendix Table 2:

Correlations Among Key Independent Variables, 2012 Levels Markets for Specialist Physician Services

	own- specialty HHI	integration with generalist	integration with generalist* mean(generalist HHI)	integration with other specialist	integration with other specialist *mean(oth spec HHI)	integration with generalist* own HHI	integration with other specialist* own HHI
Own-specialty HHI	1
Integration with generalist	0.3111	1
Integration with generalist * mean(generalist HHI)	0.4095	0.7570	1
Integration with other specialist	0.2370	0.6172	0.6239	1
Integration with other specialist*mean(oth spec HHI)	0.3311	0.5777	0.7533	0.9007	1
Integration with generalist * own HHI	0.6270	0.8444	0.7644	0.5328	0.5710	1
Integration with other specialist* own HHI	0.4936	0.5515	0.6442	0.8712	0.8715	0.6800	1

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. N = 92,416 ZIP code*specialties. Table statistics weight each ZIP code/physician type by the number of 2012 claims underlying the price index.

Appendix Table 3:

Descriptive Statistics For Key Variables, 2008-2012 Changes Markets for Generalist Physician Services

	Mean	Standard deviation	Mean in bottom quartile	Mean in top quartile
Price index	−0.0044	0.0826	−0.1019	0.1026
Own-specialty HHI	0.0005	0.0332	−0.0350	0.0361
Integration with specialist	0.0671	0.1192	−0.0585	0.2208
Integration with specialist*mean(spec HHI)	0.0190	0.0434	−0.0236	0.0718
Integration with specialist* own HHI	0.0110	0.0369	−0.0206	0.0510

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. N = 32,194 = 16,705 ZIP codes for family practice/general practice + 15,489 ZIP codes for internal medicine. Table statistics weight each ZIP code/physician type by the number of 2012 claims underlying the price index.

Appendix Table 4:

Descriptive Statistics for Key Variables, 2008-2012 Changes Markets for Specialist Physician Services

	Mean	Standard deviation	Mean in bottom quartile	Mean in top quartile
Price index	−0.0015	0.0884	−0.1030	0.1114
Own-specialty HHI	0.0047	0.0520	−0.0485	0.0657
Integration with generalist	0.0507	0.1613	−0.1146	0.2494
Integration with generalist * mean(generalist HHI)	0.0078	0.0303	−0.0187	0.0416
Integration with other Specialist	0.0578	0.1457	−0.0746	0.2381
Integration with other specialist*mean(oth spec HHI)	0.0155	0.0432	−0.0205	0.0664
Integration with generalist * own HHI	0.0150	0.0769	−0.0498	0.0944
Integration with other specialist* own HHI	0.0186	0.0711	−0.0300	0.0896

Note: Generalists include family practice/general practice and internal medicine; specialist specialties are listed in Table 3. N = 92,416 ZIP code*specialties. Table statistics weight each ZIP code/physician type by the number of 2012 claims underlying the price index.