The Impacts of Restricting Mobility of Skilled Service Workers

A. Physician Practices

Although the questions we discuss apply generally to firms that provide high-skilled services, our empirical analyses focus on primary care physicians, including family medicine, general internal medicine, and pediatrics. Primary care physicians, which represent about 39 percent of practicing physicians in the United States, typically serve as the first point of contact for patients and as coordinators of their care. Consequently, they tend to have lasting relationships with their patients.

Physician practices are generally organized as either solo practices with one physician or as group practices with multiple physicians. In group practices, physicians may be either owners or employees. Group practices typically use compensation contracts that combine a fixed salary, an individual productivity component, and a firm profit component. This creates incentives for physicians who make referrals to keep patients within a practice whenever reasonable.⁵ Operating a group practice generally requires substantial investments in recruiting new physicians and in developing relationships with a stock of patients to maintain demand for the new physicians’ services. Accordingly, practices have strong incentives to protect these investments from competitive threats. This is also evident from the fact then when a practice is sold, the main factor determining the transaction price is often the stock of patients.

B. Noncompete Agreements

Our survey data provide the first known systematic documentation of how prevalent NCAs are among physicians—we find that about 45 percent of primary care physicians in group practices are bound by NCAs.⁶ Section V. A presents several important aspects of heterogeneity in the use of NCAs.

Whereas NCAs often restrict workers from joining another firm in the same industry, in service industries like healthcare that have geographic markets, NCAs are geographic in scope. Physician NCAs, for example, prevent a departing physician from practicing any medicine anywhere in an ex ante defined geographic area for a specified period of time. NCA restrictions without geographic limits are generally not enforceable for physicians, which allows physicians to continue to work in medicine without a career detour, but only if they move out of the local area and bear the associated costs. Although states differ in how large the defined geographic markets can be, examples of common market definitions used in NCAs involving physicians include the county containing the practice or a 20-mile radius around the practice. NCAs must also not be excessive in duration, and two to three years is often deemed reasonable. Some NCA contracts allow physicians to pay damages in lieu of leaving the market. For example, under an NCA recently upheld in Kansas, a family physician leaving a medical group was prohibited from practicing for three years in the same county as the group unless she paid the group 25 percent of her earnings during the period.⁷

As elements of labor contracts, NCAs are subject to review at the state level, and the ability to enforce NCAs is based on state case law and applicable statutes. Currently, NCAs are enforceable to at least some extent in every state except North Dakota. Thirty-nine states follow common law, while the remaining 11 states have passed specific legislation that guides the enforcement of NCAs.⁸ In common-law states, the precedent that determines enforceability of NCAs was generally shaped long before modern healthcare and insurance markets existed.⁹ The inertial nature of common law makes NCA laws slow to adapt to changing market conditions, although they do evolve somewhat over time (Bishara 2011).

While variation in the ability to enforce NCAs across states offers an opportunity for comparative analysis, a difficulty for empiricists has been finding a way to characterize this variation. Popular summaries of the enforcement of NCAs, such as Wilson (2006), broadly divide states into three groups: those where noncompetes are judged “unenforceable” (seven states), those where they are judged “enforceable” (36 states), and those where case law is judged uncertain (nine states). While this categorization has the appeal of being easy to apply, in practice issues of enforcement are much more nuanced then these summaries suggest.¹⁰ Recently, a much more careful and precise quantification system was developed by Bishara (2011). Based on legislation and case law in each state, Bishara (2011) scores the overall ability to enforce NCAs on a state-by-state basis along each of eight different dimensions. Our empirical analyses make use of these quantified restrictiveness scores, which we refer to as “Bishara Scores,” and Online Appendix Table A2 reports the questions and rules used in developing these scores.

C. Legal Constraints in Physician Compensation Contracts

One important stylized feature of physician labor markets is that there are many constraints on the factors that can be used to determine compensation. One constraint of particular importance is that accepting payments for patient referrals is not only viewed as unethical, it is per se illegal in the United States. Yet, in many other settings, commissions or bonuses for client referrals are both accepted and commonplace.

Specifically, the Ethics in Patient Referral Act of 1989 (and amendments made under the Comprehensive Physician Ownership and Referral Act of 1993) requires that “any amount paid by an employer to a physician. who has a bona fide employment relationship with the employer” must be “consistent with the fair market value of the services, and.not determined in a manner that takes into account (directly or indirectly) the volume or value of any referrals by the referring physician.…”(42 CFR § 411.357 (c)).¹¹ The Code of Federal Regulations (42 CFR § 411.352 (g)) adds restrictions on compensation arrangements in practices that treat Medicare patients, requiring that “no physician who is a member of the group practice directly or indirectly receives compensation based on the volume or value of his or her referrals.”¹² However, it is explicitly permitted to base compensation on a physician’s output or on the amount of revenue they generate, or to use per capita group profit sharing (Olson and Stanley 2004).

A. Basic Model Setup

We consider a two-period model of a firm owned by a physician proprietor, indexed by a, who is endowed with P patients, which the physician can treat to generate revenue where f is assumed to satisfy: f(0) = 0, f(P) > 0, and f’(P) < 0. The strictly concave production function f can be interpreted as the monetary equivalent of the utility the owner would receive by treating the patients, net of any utility lost to providing the effort and time required to treat the patients.

Alternatively, the owner could hire a worker physician, indexed by w. In this case, the owner can choose to allocate (“refer”) P_w ≡ P – P_a patients to the worker, and the firm’s per-period profit is given by: where S is the cost of paying the worker’s salary. Since f is strictly concave, it is potentially advantageous for the owner to share the patients with the worker and pay the worker’s salary. However, any allocated patients P_w become loyal to the worker physician, who may then poach the patients.

The worker may exit the firm in the second period for two reasons. First, with probability (1 − ρ), the worker and firm exogenously separate, where 0 < ρ < 1. Second, if the worker can earn a higher salary in the outside competitive market, the worker will voluntarily exit. In either case the worker takes allocated patients with her. The outside option salary for a physician without any patients is denoted , and the outside option increases to for a worker with P_w > 0 loyal patients.

In order to prevent the worker from poaching patients in the second period, the owner may require the worker to agree to an NCA. If a worker signs an NCA and the job is then terminated for any reason, allocated patients are returned to the owner, and the worker must exit the geographic market. At the beginning of period 1, workers have heterogeneous geographic location preferences R_w, expressed in monetary units, which are distributed uniformly , and are private knowledge of the worker. Larger R_w values indicate high willingness to pay for staying in the geographic market, which increases the expected cost of signing an NCA. At the end of period 1, workers receive geographic preference shocks from a discrete uniform (Bernoulli) distribution . Therefore, the sum is a continuous uniform distribution. If R_w + ϵ is sufficiently negative, relative to earnings potential, workers may increase their utility by moving to a new geographic market.

The timing of events occurs as follows. At time zero, firms post take-it-or-leave-it offers that have three elements: (i) noncompete agreements {N, C}, where N corresponds to a contract with an NCA, and C to a contract without; (ii) first-period compensation, S₁; and (iii) second-period compensation, S₂. Workers observe all posted offers and choose jobs that maximize earnings S₁ + S₂, net of any expected relocation costs .¹³ Firm owners then make patient referral choices. Production occurs, workers and firms earn payoffs, and then exogenous separation draws ρ are realized. Workers then announce whether they wish to voluntarily exit the job.

Contractual commitments to allocate P_w are forbidden, and as discussed in Section II. C, compensation in each period must be based on fair market value and may not include an implicit purchase or sale of patient referrals. We impose this legal constraint by assuming a minimum salary in each period, which is consistent with fair market value and prevents workers from forgoing salary to implicitly purchase referrals.¹⁴

We begin the model by considering fixed salary compensation only, and allowing one-sided forward commitments by the firm to guarantee S₂. We then consider an extension of the model in which future salary commitments have limited credibility—firms can guarantee not to cut earnings, but they may not credibly commit to guaranteed salary increases. These assumptions about contract structures play an important role, because once a worker has signed an NCA, their reservation salary decreases in the second period due to the cost of relocating.

Firms maximize the sum of expected profits over the two periods π₁ + π₂. Workers choose jobs that maximize two-period earnings net of expected relocation costs, .

Hedonic wage theory (Rosen 1974) says that the competitive market salary will be determined by the preferences of the marginal worker, who has a value of R* that makes them indifferent to accepting an NCA. Since we are interested in studying a mixed equilibrium, in which some jobs include NCAs and others do not, we assume that R is sufficiently large that some workers would never accept an NCA at any price that firms are willing to pay. The hedonic equilibrium is therefore characterized by a single worker with preferences R* that determines assignment to jobs: workers with R_w < R* sort into jobs with NCAs, and workers with R_w > R* sort into jobs without NCAs. For simplicity, we also assume that R* > e, which implies that workers who sort into jobs without NCAs will never choose to relocate (as long as their earnings do not decrease in period two.)

B. Contracting Frictions, Bargaining, and Earnings

A stylized fact of labor markets, however, is that forward commitments to guaranteed salary increases are rarely observed. If firms cannot credibly commit to a contract specifying a second-period salary, then NCAs create a bargaining problem. Once a worker has signed an NCA their bargaining position decreases in the second period, since the firm knows that the worker’s reservation wage has declined due to the cost of relocating. Without credible forward commitments, workers may demand front-loaded compensation in order to accept a job with an NCA. All else equal, this incentive may force the earnings path to be flatter than the profit-maximizing path derived above. Flattening the earnings path increases the probability of worker separations in the second period, and reduces welfare relative to the equilibrium with credible forward commitments.

Our goal in this section is to demonstrate that there exists an incentive compatible revenue-sharing contract in which the loss of ex post bargaining position due to NCAs does not cause distortions that flatten earnings paths, avoiding potential deadweight loss from excess turnover. The existence of such a contract suggests that when turnover is costly to firms, as is the case in the model presented above, then share-based contracts may be Pareto-improving relative to front-loaded or flat compensation paths.

To see this, suppose compensation structures may depend linearly on output: where α is the share of output that the worker keeps as compensation. A contract is now defined as (i) first-period compensation, M₁, (ii) noncompete agreements {N, C}, and (iii) forward “sticky wage” commitments by the firm to not reduce S or a in the second period. The sticky wage commitment reflects the limited credibility of guaranteed future salary increases, but allows firms to credibly commit to not decreasing either compensation parameter.¹⁵

To pin down the intuition behind the model equilibrium, suppose there is a small amount of stochasticity in output. We also introduce an upward-sloping output function, by assuming that output grows in the second period at the rate δ > 1. Firms without NCAs have no compelling reason to use revenue-sharing contracts. Since the firm is risk-neutral, they will insure the worker against output shocks by offering the contract in period 1. The worker can then renegotiate the contract in the second period by threatening to separate, .

Workers with NCAs, however, cannot increase their compensation in the second period by threatening to exit, since the worker’s expected outside option yields a payoff of only . Anticipating that their bargaining position will decline in the second period, workers must negotiate an ex ante incentive-compatible contract with fixed compensation components { S_N, α_N}.

To gain intuition, suppose for simplicity that output shocks are very small, so the profit-maximizing equilibrium earnings path can be approximated by re-solving the model with log utility: subject to the legal constraint on minimum salaries and the worker’s participation constraint:

When , and the participation constraint binds, the profit maximizing earnings path is

Now, introducing revenue-sharing contracts, the equilibrium compensation contract {S_N, α_N} that matches this profit-maximizing earnings profile must satisfy: 4 4

Equation 3 implies . Subtracting Equation 3 from 4 gives:

Notice that the RHS is strictly positive because of the earnings constraint .¹⁶ The LHS is also strictly positive since δ > 1. This implies α_N > 0.

The economic intuition behind this result is straightforward. Although limited credibility constrains the set of contracts, this constraint can be overcome if the firm uses fixed revenue-sharing rates to match the profit-maximizing earnings path that would occur under perfect forward credibility. This equilibrium requires the existence of an upward-sloping function to which a can be tied; growing output, δ > 1, is one natural example of such a function. When this occurs, firms can bundle NCAs with revenue-sharing contracts, which allows compensation to increase along with output, without the need to renegotiate contract terms in the second period.

In this simple model we abstract from explaining which firms choose to use NCAs, and the hedonic equilibrium is driven entirely by sorting on worker preferences. Of course, in a more realistic setting the decision by a firm to impose NCAs is unlikely to be random. For example, firms in geographic markets with fewer patients per physician (lower endowments of P per firm) may derive more benefits from protecting the marginal patient from being poached, increasing R*, and hence the fraction of employees with NCAs. Similarly, if production is augmented by a persistent productivity shifter τf(P), more productive firms may derive greater benefits from NCAs. Finally, if firms differ in hiring costs, higher cost firms may benefit more from NCAs. Although our theoretical discussion abstracts from many of these issues, appropriate interpretation of our empirical estimates depends on the extent to which potentially unobserved factors directly affect both the decision to use NCAs as well as the outcomes of interest in our hypotheses. In Section V we return to discuss these selection issues, and the conditions under which our parameter interpretations may be affected by selection.

C. Summary of Testable Hypotheses

The goal of our empirical analyses is to test for evidence that physician practices use NCAs to prevent patients from being poached, protecting firms’ investments in client relationships, which we model as intrafirm referral choices in the stylized model above. Our primary analyses test Hypothesis 2, that NCAs increase the rate of return to job tenure. We test this hypothesis by estimating the relationship between the use of NCAs and within-job earnings growth and decomposing the earnings growth differential into components due to experience and job tenure.

We also make use of several other predictions from the model to provide corroborating suggestive evidence. Hypothesis 1 is that firms that use NCAs allocate more patients to employed physicians. In the survey data, we are able to observe the distribution of patients to physicians. We test for evidence of disparities in the allocation of patients between employed physicians and those that have equity ownership in the firm. If NCAs reduce referral holdups, firms that use NCAs should have more balanced distributions of patient loads across physicians. In the medical context, however, all patients are not alike. Physicians who treat privately insured patients tend to receive higher reimbursements than those who treat Medicaid patients, for example. In addition to testing for overall disparities in the number of clients, we also examine heterogeneity in the allocation of clients by their source of insurance coverage.

Hypothesis 3 is that NCAs may be bundled with share-based compensation incentives to overcome the effects of changes in bargaining position. We use data on the fraction of earnings that come from incentive payments tied to individual production to provide stylized summary statistics on this hypothesis. We also empirically evaluate the alternative hypothesis that physician practices use NCAs solely to reduce job turnover.

A. Physician Survey

We use the Physician Perspectives on Patient Care Survey, which we conducted in 2007,¹⁷ using a sampling frame drawn from the American Medical Association (AMA) Masterfile. The population from which potential respondents were drawn included only primary care physicians (family practice, general practice, general internal medicine, and general pediatrics) in five states: California, Texas, Illinois, Georgia, and Pennsylvania. Using a state-based sample rather than a national survey permitted collection of larger samples from local market areas, and these states were chosen to be representative of a variety of practice environments, while being geographically diverse. The target population excluded residents, fellows, physicians not in clinical practice, and those over 70 years old. Pediatricians and minority physicians were oversampled.

The AMA database provides information on physician location and contact information, specialty and training, age, and race. Telephone calls verified contact information and whether sample physicians were providing patient care. A multimode (mail and web) self-administered survey was conducted. A packet was sent by Federal Express to a total of 2,831 physicians containing a mail survey accompanied by an advance letter, a prepaid business return envelope, and an honorarium check of $100. Physicians were given the option of responding by web. Followup was conducted for those physicians who did not respond, with separate followup with those who did not respond but cashed their checks. Altogether, a total of 1,967 usable responses were received, 216 (11 percent) of which were by web. The overall response rate was 69.8 percent.¹⁸

The survey questionnaire included detailed questions on the following topics: physician characteristics, practice characteristics, physician demographics, practice financial performance, physician earnings over several years, patient mix, practice administrative controls, average prices negotiated with insurance companies, and patient vignettes to elicit knowledge of clinical guidelines, diagnoses, and treatment recommendations.

One section of the survey that is key to the analyses includes income and revenue generated by each of the respondents. Although the survey was conducted at a point in time, respondents were asked about longitudinal variation in their earnings. This was done by asking for earnings levels in 2005, and then asking by what percentage their earnings differed from 2005 levels in 2006, and three years prior, in 2002. We use responses to these questions to estimate the rate of growth of earnings within jobs. We also observe the year in which each respondent completed medical school and the year in which they joined their current practice, from which we calculate potential experience (in medical care) and job tenure.

The survey also asks detailed questions about the structure of compensation for each worker. Specifically, it asks, “What percent of your 2005 earnings was paid as flat salary?”, “What percent of your 2005 earnings was based directly on fees-for-services you provided, or on your productivity?”, and “What percent of your 2005 earnings was in the form of pay-outs from withholds, practice bonuses, or other incentive payments, including pay-for-performance bonuses?” We use responses to these questions to estimate how NCAs are related to the use of share-based compensation structures.

Physicians were asked how many medical practices they worked in and their ownership status in their main practice. If they responded they were a sole-owner, the survey proceeded to questions about general practice characteristics. However, if the physician indicated that they were not a sole-owner, they were asked about their employment status and the following question regarding NCAs: “Were you to leave your (main) practice, would you be subject to a noncompete clause?”

A copy of the survey questionnaire containing the wording of all questions is included in the Online Appendix.

B. State NCA Laws

We use data from Bishara (2011) to measure the relative strength of enforceability of NCA laws across different states. The measure was created by analyzing case law in each state and comparing laws based on eight different dimensions. Each dimension was assigned a weight based on legal knowledge about the relative importance of the dimensions. The specific questions that define the eight dimensions, along with benchmarks for how states were scored and relative weights of each question, are included in Online Appendix Table A2. For example, one important dimension in which state laws differ is whether NCAs are still enforceable in the event that the employer makes the decision to terminate the relationship. In some states NCAs would still be enforceable, while in others NCAs apply only to voluntary separations made by workers.

We normalize the NCA enforceability measures by dividing by the maximum score (Florida, 470), to create a continuous measure between zero and one. Online Appendix Table A3 shows that there is substantial variation in the Bishara (2011) enforceability scores for the five states in our sample (CA, GA, IL, PA, TX), which are respectively ranked 50th, 43rd, 4th, 23nd, and 32rd in stringency of enforcement out of 50 states plus Washington, DC.

A. Summary Statistics on NCA Use

Table 1 reports the share of physicians with employment contracts that contain NCAs in each state in our sample. The use of NCAs varies substantially across states, ranging from 31.3 percent in California to 60.6 percent in Pennsylvania, with an average of 45.1 percent of all physicians in group practices in the sample subject to an NCA. This variation in usage is consistent with differences in enforceability—physicians are more likely to have contracts with NCAs if they work in states in which NCAs are more enforceable.

NCAs are also used more frequently with physicians who are employees (49.2 percent) rather than part-owners (43.1 percent) of a practice. Part-owners have some deterrent to competing against their current practice because doing so could devalue their share of the practice’s equity, which may be a relatively illiquid investment.¹⁹

Given the lack of prior evidence documenting how prevalent NCAs are in physician groups, one may expect that NCAs are a relatively recent phenomenon in physician organizations. Table 2 suggests that this is not the case—the probability that physician in the sample is bound by an NCA is fairly stable across all levels of physician experience and practice tenure.

Since understanding why practices differ in the use of NCAs requires first knowing what motivates practices to use NCAs, our main question of interest in this study, we begin by simply documenting summary statistics on the use of NCAs by practice type. Table 3 describes the heterogeneity in the percentage of physicians in group practices that are bound by NCAs.

The first notable pattern in the table is that NCAs are quite common in every specialty, practice type, demographic group, and geographic market. There is relatively little variation across specialties within our sample of primary care physicians, ranging from 41 percent of physicians with two or more specialties to 46 percent for pediatric specialists. There is, however, some variation across practice settings, with 50 percent of office-based physicians having NCAs, compared to only 37 percent of physicians employed at hospitals or free-standing care centers. This pattern is consistent with the hypothesis that NCAs are more likely to be used in settings in which ongoing patient–doctor relationships are more valuable. The least likely group to have signed NCAs are physicians who are independent contractors.

The last group of summary statistics shows that there is relatively little variation in NCA use by demographics or by characteristics of the geographic market in which the practice is located. However, there is substantial variation in NCA use across states. In Online Appendix Table A8 we estimate the marginal effect of the Bishara Score on the probability of having an NCA, and show that state enforceability is strongly predictive of NCA use. An increase in enforceability from the least restrictive state (ND) to the most restrictive state (FL) is associated with a 32 percentage point increase in the probability that a physician will have an NCA. This suggests that firms consider state laws to be important factors in calculating the expected benefits to imposing NCAs. It is also worth noting that although employment-based NCAs are virtually unenforceable in California, we still find that 31 percent of physicians in group practices have signed them. This may suggest that some practices use NCAs simply as deterring threats, even when they are not enforceable.

B. Do Physician Practices Use NCAs to Deter Poaching?

We begin by estimating the impact of NCAs on earnings growth and the rate of return to job tenure. These analyses are based on Hypothesis 3 from Section III, which suggested that if practices were motivated by deterring exiting physicians from poaching patients, NCAs should increase the rate of return to tenure, indicative of an increase in firmspecific human capital investments. We use both extensive-margin variation in the use of NCAs and intensive-margin variation state-level NCA enforceability laws to show that NCAs increase the rate of return to tenure and that this effect appears to be driven by state NCA policies rather than selection. We also provide a variety of suggestive stylized summary statistics that are consistent with the other predictions from Section III, although our main results come from the longitudinal models of earnings differentials.

2. Suggestive corroborating evidence

Across-Job Earnings Differentials

Although we have focused on earnings growth differences associated with NCAs, hedonic wage theory also suggests that earnings levels should also be higher in jobs with NCAs. To show the conditional mean difference in average earnings associated with NCAs, which we view only as corroborating suggestive evidence, we estimate the following fixed-effects model: 11 where Y_ijm is the log hourly earnings of physician i at firm j in market m; NCA_ij is either a binary indicator or interaction between NCA and enforceability; X_ij includes characteristics of the worker, such as gender, experience, medical school location, tenure, specialty, and equity status, and characteristics of the firm, such as the number of physicians, university affiliations, and whether the practice is multispecialty; and θ_m is a fixed PCSA market effect.

Table 6 presents estimates of Equation 11. The first column shows that the average conditional hourly earnings of physicians with NCAs are about 13 percent higher. The second column includes a continuous measure of NCAs that accounts for variation in state enforceability interacted with a binary NCA indicator, and interacted with experience to test whether earnings rise faster for physicians with NCAs in states where NCAs are easier to enforce. Since it is difficult to interpret each coefficient on its own, Figure 1 graphically depicts the predicted wage profiles from Column 3.

View this table:

• View inline
• View popup

Table 6

Fixed-Effects Wage Models

Figure 1 Cross-Sectional Wage Profiles, by Experience and Tenure

Notes: Vertical axis is expected hourly earnings conditional on PCSA effects and covariates in Model 3 in Table 6 (top), and with tenure instead of experience interacted with NCA (bottom). Sole owners are excluded. Lines are best-fitting quadratic functions, with 95 percent confidence intervals.

The upper figure graphs the predicted wage profile as a function of potential experience for physicians with and without NCAs, and the lower panel graphs similar tenure profiles. There are two notable features of these graphs. First, hourly earnings of physicians with and without NCAs are similar at the beginning of a career or a new job. Second, earnings rise faster over time in jobs with NCAs. After about 15 years of either potential experience or job tenure, hourly earnings of workers with NCAs are more than 20 log points higher than observably similar workers without NCAs. Although these estimates do not account for unobserved worker heterogeneity and job-matching, as the panel-based estimates do, they provide corroborating evidence that the increase in the rate of earnings growth documented in Section V.B.1 leads to differences in hourly earnings levels.

Figure 2 shows kernel density distributions of conditional hourly earnings for physicians with and without NCAs, in low enforceability states compared to high enforceability states. In Figure 2a there is a large positive shift in the distribution of hourly earnings for physicians with NCAs; however, as Figure 2b shows, there is no such shift when NCAs are very difficult to enforce.

Figure 2 Kernel Density of Log Hourly Earnings Distribution

Notes: Kernel density of the predicted hourly earnings distribution conditional on county effects, physician specialty, practice type, practice size, experience, experience squared, ownership status, gender, and race. High enforceability defined as Bishara Score above 0.8, which includes Illinois, and low enforceability defined as Bishara Score below 0.2, which includes California.

The compensating differentials framework can be useful for understanding why earnings effects of NCAs may be larger in states where NCA enforceability is higher. If workers have heterogeneous preferences for occupational mobility, theory suggests they will sort into jobs with NCAs according to these preferences, and the equilibrium wage differential will be determined by the preferences of the marginal worker. If more workers are bound by NCAs in a market, the marginal worker ought to be more averse to accepting an NCA, increasing the observed wage premium. Summary statistics suggest that in states where NCAs are easier to enforce, firms are more likely to use them, consistent with earnings differentials being larger in high enforceability states.

Evidence from Compensation Structures

Prediction 2 from our model is that physicians with NCAs have compensation contracts that are more strongly tied to individual output. Table 7 shows summary statistics on compensation components in the survey data. We find that physicians with NCAs receive about 62 percent of their total annual income as guaranteed fixed salary, compared to 80 percent of income for physicians without NCAs. Consistent with our prediction, the share of total earnings that is tied to individual productivity is more that twice as high for physicians with NCAs, 31.2 percent compared to 15.1 percent. The disparity in the level of individual incentive payments in even larger, since physicians with NCAs earn more on average. Other incentive payments tend in the same direction, accounting for about 5.5 percent of total earnings for physicians with NCAs and 4.0 percent for physicians without.

View this table:

• View inline
• View popup

Table 7

NCAs and Compensation Structure

Although our theoretical model focuses on the role of incentive pay in generating upward-sloping earnings paths despite the reduction in bargaining power caused by NCAs, the literature more broadly has discussed several alternative explanations for the use of incentive pay in physician compensation. Ellis and McGuire (1986) develop a model of physician behavior under alternative compensation structures. They show that fee-for-service reimbursement, which is more strongly incentive-based, can lead to overutilization of services, while purely prospective reimbursement, which is less incentive based, can lead to underutilization if physicians care sufficiently strongly about profits relative to patient welfare. They conclude that a mixed reimbursement system in which physicians have incentives that are only partially linked to services provided can be superior and can prevent underutilization of care. In this model, incentive pay primarily affects the quantity of services provided because physicians are assumed to have perfect control over quantity as agents of their patients. The model suggests that physicians who more strongly value patient welfare relative to profits have less need for incentive pay.

To help interpret the observed difference in incentive pay associated with NCAs, we test whether these alternative theories of incentive pay from the literature can explain the patterns we observe. The models in Ellis and McGuire (1986) and Clemens and Gottlieb (2014) suggest that incentive pay should be lower among more altruistic physicians who are less motivated by profits. Our survey data contain several questions related to providing charity care. In Online Appendix Table A7 we show that the gap in incentive pay associated with NCAs cannot be explained by controlling for these measures of altruism, by observed physician and practice characteristics, or by geography. Relative to the unconditional mean difference in the share of earnings tied to individual production of 18.5 percent, the difference drops to 16.1 percent after conditioning on the full set of controls.

One additional potential concern is that incentive pay may be used in areas with relatively low supply of physicians to encourage doctors to work more hours. However, we find that NCAs are less common in areas with low supply, suggesting that this explanation is unlikely to drive the observed patterns in incentive pay. Specifically, we find that in counties in which less than 20 percent of the population lives in an urban area, which tend to be areas with low physician supply, NCAs are 6 percent less likely to be used.

Another possible alternative explanation for this pattern is discussed by Gaynor and Gertler (1995), who argue that in partnerships the influence of productivity on earnings could be motivated in part by risk protection. This suggests that if firms using NCAs tend to be smaller, earnings may be more directly influenced by productivity because there are fewer physicians over which to smooth fluctuations in output. However, Table 8 shows that very small practices with two to three physicians are less likely to use NCAs (31 percent) than practices with four to 499 physicians (45−50 percent).

View this table:

• View inline
• View popup

Table 8

NCA Use by Firm Size

Evidence from Patient Allocations and Generated Revenue

For evidence on Prediction 1, that firms using NCAs have more intrafirm patient referrals, we look directly at client allocations and worker-level revenue within firms. Table 9 shows the mean number of weekly patient visits for employees and practice owners with and without NCAs. Employed physicians with NCAs see over 11 percent more patients per week than those without NCAs, while the number of patients seen by practice owners does not vary much with NCA use. However, even more important differences underlie these totals—the composition of patients by source of insurance coverage is also substantially affected by NCA use. Physicians with NCAs have substantially more privately insured patients and Medicare patients, which have the highest reimbursement rates, but treat fewer patients with Medicaid coverage or no insurance.²⁴ Medicaid payment rates averaged roughly half of the private insurance rates in our data.

View this table:

• View inline
• View popup

Table 9

Patient Stocks, Revenue, and Hours

We also estimate the weekly revenue generated by each physician using data on the number of patients served, the shares of patients covered by private insurance, Medicare, and Medicaid, and the associated reimbursement rates from each insurer. For privately insured patients, data on negotiated reimbursement prices are based on responses to the survey question: “On average, what is your net fee after discount for an initial office visit with a private, commercially insured patient?” Similarly, we apply reimbursement rates in the corresponding geographic area for primary care services to new patients making initial visits who are insured by Medicare or Medicaid.²⁵ Although this estimated revenue index cannot account for unobserved variation in the mixture of services provided by different physicians, it does provide an estimate of the effect of variation in the absolute number of patients as well as the composition of patients across physicians.

Both higher patient levels and more favorable patient mixes contribute to employed physicians with NCAs generating 35 percent more revenue per week ($8,964 vs. $6,637). With only about 11 percent more hours worked per week, this translates to 17 percent more revenue per hour of patient care.²⁶

More revealing are the differences between firms that use NCAs and those that do not in the disparities in the allocation of clients between owners and workers. In firms that use NCAs, employed physicians generate about $8,964 in revenue per week, while owners generate about $9,246. In firms that do not use NCAs, however, the difference is substantially larger and statistically significant—employed physicians generate about $6,637 per week, compared to $8,253 for owners. This difference in disparities is consistent with greater intrafirm referrals, as predicted by our model.

Still, the exact mechanism behind these patterns is unclear. Clearly, patient mixes can affect the incentive to use NCAs, and firms with the largest potential benefit from NCAs are most likely to impose them. However, it is also possible that firms using NCAs have more valuable patients because they have successfully prevented workers from poaching valuable clients in the past.

In Online Appendix Table A6 we look for evidence that the difference in revenue generated by physicians with NCAs is due to selection based on the types of workers or firms that choose to use NCAs. We find that the difference in revenue associated with NCAs cannot be explained by physician and practice characteristics; in fact, conditioning on observable characteristics increases the estimated revenue disparities associated with NCAs. The table also shows that these patterns hold within geographic markets.

This evidence on patient allocations is of course not causal, but it is consistent with the first prediction from our model. Along with the additional results on returns to tenure, earnings effects, and compensation structures, the combination of evidence is broadly consistent with all three predictions from the model, in which NCAs are used by physicians to prevent poaching and protect the value of patient stocks.

C. Do Physician Practices Use NCAs to Reduce Turnover and Hiring Costs?

It is also possible that physician practices use NCAs to reduce turnover and hiring costs. If so, we are interested in estimating the relative importance of these different sources of benefits in affecting firms’ decisions to use NCAs.

There are several mechanisms through which NCAs could increase the length of job spells. They could deter exit directly by making it more costly, they could induce self-selection by workers with private knowledge about their expectation for remaining at the firm, or they could reduce the probability that an outside offer will exceed earnings by increasing the returns to tenure. While it is not possible to identify each of these effects separately with available data, we are able to estimate the overall relationship between NCA use and the duration of job spells.

Previous studies in the literature have shown that NCAs have causal effects on worker job mobility rates. Marx, Strumsky, and Fleming (2009) use an exogenous inadvertent change in enforcement of NCAs in Michigan in 1985 and find that the average mobility of inventors producing patents in Michigan fell relative to the mobility of inventors in other states as a result of the increase in the enforceability of NCAs. Marx (2011) finds 40 percent of electrical engineers surveyed had signed NCAs, and that workers who left firms were more likely to switch industries if they were subject to an NCA.

Table 10 shows the unconditional distribution of job tenures for physicians with and without NCAs. As expected, physicians with NCAs are significantly less likely to have begun their job within the prior seven years. Figure 3 shows similar patterns conditional on observed characteristics. The figure plots the estimated difference in the conditional probability of observing a given year of tenure for a physician with an NCA relative to one without. Physicians with NCAs were significantly less likely to have tenures between one and seven years and significantly more likely to have tenure of nine or more years. These patterns imply that the CDF of job tenures of physicians with NCAs first-order stochastically dominates the distribution for those without NCAs.

Figure 3 Marginal Effects of NCA on Years of Tenure

Notes: Graph shows the effects of a one unit increase in “Bishara Score*NCA” on the conditional probability of observing a given year of tenure based on estimates from an ordered probit model with covariates identical to Model 1 in Table 11. The 95 percent confidence intervals are shown in bars.

Table 11 presents our main estimates on job spell durations from fixed-effects negative binomial models of job tenure, conditional on observed worker and firm characteristics and primary care market effects. Column 1 shows physicians with NCAs have 12.3 percent longer job spells. Column 2 shows that the difference in job spell lengths increases to 29.6 percent when using variation in state enforceability. Although these are not causal effects, there is some suggestive evidence on the mechanism behind these patterns that is consistent with NCAs causally lengthening job spells. Since firms are more likely to use NCAs where they are more enforceable, this suggests that if the entire difference in job tenures were due to sorting on unobserved preferences for mobility, then as enforceability increases, the marginal workers who accept NCAs should be more likely to switch jobs. Column 2 suggests the opposite—where enforceability is higher, job spells are longer.

Column 3 adds the Bishara Score without interacting with NCA use. The coefficient reveals information about sorting on unobservables. For workers without NCAs, who should not be directly affected by state enforceability laws, higher enforceability is actually associated with shorter job spells. This suggests that physicians sort into jobs on the basis of preferences for mobility. In combination, the three models provide suggestive evidence that each of the potential mechanisms we discuss appears to play some role in increasing the length of job spells. However, with only a single cross-sectional sample, a direct relationship between tenure and job-spell length requires a stationarity assumption in the difference between the rates of job flows for jobs with and without NCAs (Kiefer, Lundberg, and Neumann 1985).

Although these increases in job-spell length may be valuable to firms, are the reductions in hiring costs large enough to explain the estimated difference in career earnings? To answer this question, we calculate the net present value of the earnings differences associated with using NCAs as a function of the length of completed job spells by integrating over the discounted difference between the two functions in the bottom of Figure 1 from zero to year T. We then back out the hiring costs that firms would have to face in order to be indifferent about paying the present value of the wage differential to achieve these reductions in turnover. The estimates of course require assumptions about workers’ time preferences for money. At the mean completed job-spell length of 14 years, the present value of the future wage differential associated with NCAs is between $149,000−$274,000 when the rate of time preference varies between 10 percent and 2 percent. The largest estimate of the effect of NCAs on job-spell lengths is a 29.6 percent increase, which implies that the cost of hiring one primary care physician would have to be at least $653,000−$1,200,000 to justify paying workers this earnings differential solely to reduce turnover. Online Appendix Figure A1 plots the full range of estimates as a function of tenure and shows sensitivity of the estimates to the assumed discount rate. For the full range of discount rates the average implied threshold hiring costs are implausibly high for primary care physicians, suggesting that although turnover reductions appear to be substantial, they are very unlikely to be the primary motivation behind the use of NCAs among physician practices.

This conclusion may of course be specific to physicians or other skilled service workers, in part because the nature of job mobility is somewhat different. In other markets where markets are defined by products rather than by geography, NCAs can have very different types of mobility effects since workers may have to switch industries to avoid violating the contract. Fallick, Fleischman, and Rebitzer (2006) and Gilson (1999) discuss the strong importance of job mobility on the formation of agglomeration economies in technology-producing markets like Silicon Valley and suggest that the lack of enforceability of NCAs contributed to the microfoundations of agglomeration economies in the technology sector.

D. Robustness Analyses