Primary Care Physicians In Medicare Advantage Were Less Costly, Provided Similar Quality Versus Regional Average

Abstract

The use of many services is lower in Medicare Advantage (MA) compared with traditional Medicare generating cost savings for insurers, whereas the quality of ambulatory services is higher. This study examines the role of selective contracting with providers in achieving these outcomes, focusing on primary care physicians. Assessing primary care physician costliness based on the gap between observed and predicted costs for their traditional Medicare patients, we found that the average primary care physician in MA networks was $433 less costly per patient (2.9 percent of baseline) compared with the regional mean, with less-costly primary care physicians included in more networks than more-costly primary care physicians. Favorable selection of patients by MA primary care physicians contributed partially to this result. Quality measures of MA primary care physicians were similar to the regional mean. In contrast, primary care physicians excluded from all MA networks were $1,617 (13.8 percent) costlier than the regional mean, with lower quality. Primary care physicians in narrow networks were $212 (1.4 percent) less costly than primary care physicians in wide networks, but their quality was slightly lower. These findings highlight the potential role of selective contracting in reducing costs in the MA program.

Enrollment in Medicare Advantage (MA) has risen steadily in the past two decades, with the share of eligible Medicare beneficiaries enrolled in private MA plans reaching just over 50 percent in January 2023.¹ Compared with traditional Medicare beneficiaries, MA enrollees have lower use of specialists, emergency departments, and elective procedures,^2–4 and they are more likely to receive higher-quality ambulatory services.⁵ These utilization differences may be affected by unmeasured favorable selection of enrollees into MA plans or by plans’ care management and utilization management activities. This article considers whether a third potential mechanism contributes to these differences: MA plans’ selective contracting with providers for participation in MA networks.

MA networks can be broad or narrow, but they almost always exclude providers in a market.^6–10 Some authors have hypothesized that MA plans selectively include cost-efficient and high-quality providers in their networks as a strategy to achieve lower spending and higher quality.¹¹ Selection of primary care physicians may help achieve these outcomes, as primary care physicians manage the continuum of care provided to their patients, including through referrals to particular specialists and hospitals.^12–14 Plans may also use selective contracting as a tool to attract more profitable enrollees.

In this national observational study, we examined whether primary care physicians in MA networks differ in costs and quality compared with the average primary care physician in their area. We also examined how findings differ for narrow versus broad networks.

Study Methods And Data

Our analysis quantified the costliness and quality of primary care physicians by studying care provided to their traditional Medicare patients. We compared the mean costliness and quality of primary care physicians included in MA networks with their regional average, using data for these primary care physicians’ traditional Medicare patients. We repeated the comparison for primary care physicians excluded from all MA networks.

Traditional Medicare Patient Sample

To quantify the costliness and quality of primary care physicians, we used a 20 percent sample of traditional Medicare beneficiaries, pooling three years of data (2016–18) to include sufficient sample sizes per primary care physician. Our analytic sample included beneficiaries who had both Part A and Part B Medicare coverage for the whole year (or until they died) and who were not enrolled in a Medicare Advantage plan in any two consecutive months throughout the year. We excluded patient-years for beneficiaries in their first year in Medicare or in a year when they moved. We also excluded patients with end-stage renal disease.

Patient And Primary Care Physician Characteristics

Using the Medicare Master Beneficiary Summary File for 2016–18, we assessed beneficiary age, sex, race and ethnicity, and Medicaid enrollment and whether disability was the reason for their original Medicare eligibility. Annual cost data from the Master Beneficiary Summary File included the amount spent by Medicare, as well as out-of-pocket payments and payments from other insurers. This cost included spending on prescription drugs in Medicare Part D. We used diagnoses that were recorded in claims from the preceding year to calculate prospective risk scores, using the Centers for Medicare and Medicaid Services (CMS)–Hierarchical Condition Categories risk-adjustment model.

We defined primary care physicians as those identified with a general practice, family practice, internal medicine, or geriatric medicine specialty in the Medicare claims data. Additional primary care physician characteristics, including gender and professional affiliations, were drawn from the December 2018 CMS Care Compare Doctors and Clinicians national downloadable file.¹⁵

Assigning Beneficiaries To Primary Care Physicians

To estimate primary care physicians’ costliness and quality, we identified the primary care physician most responsible for managing and coordinating the care of each beneficiary. We followed the Medicare Shared Saving Program’s attribution rule to annually assign traditional Medicare beneficiaries to primary care physicians based on outpatient office-based evaluation and management claims from the CMS carrier files from 2016 to 2018 (excluding claims in nursing facilities and inpatient settings), assigning beneficiaries to the primary care physician with the highest total payments.¹⁶ We excluded the almost 40 percent of beneficiaries who did not see a primary care physician in any given year—a share consistent with prior work.¹⁷

Identifying Medicare Advantage Primary Care Physician Networks

To identify primary care physicians’ participation in MA networks, we used 2019 data from Ideon (formerly, Vericred), a firm that collects data on network participation from insurers and online plan directories. Such directories are notoriously prone to errors, but the Ideon data set and our analysis minimize this issue through several validation methods (see more details in online appendix section 1).¹⁸ Among other steps, we followed prior research⁶ and validated primary care physicians’ practice locations in specific markets and their primary care specialty, using additional 2019 data from IQVIA to capture information on office-based physicians, including their clinics’ locations, and from CMS’ Care Compare Doctors and Clinicians data. We also used specialty data from traditional Medicare claims and data from the Master Beneficiary Summary File on the location of attributed traditional Medicare patients. We elected to focus on individual primary care physicians, rather than practices (defined by Taxpayer Identification Numbers), because for many practices in our data, some primary care physicians participated in MA networks, whereas others did not. We defined networks at the geographic level of the hospital referral region (HRR) and excluded networks in Puerto Rico and the other US territories. We focused on HRRs rather than counties because MA plans generally are offered in multiple counties within a geographic area, and the affiliated networks are not constructed at the county level. Our conclusions remained the same when we conducted a county-level analysis (see exhibit S6 in appendix section 3).¹⁸ Nonstandard MA plans, including employer group waiver plans, were excluded from the Ideon data.

We used the 2019 CMS plan characteristics file to link networks to their parent insurer. Using the Master Beneficiary Summary File and Ideon’s data on the MA contracts linked to each MA network, we calculated the number of MA enrollees in each HRR who used each network. We excluded from our sample very small network HRRs with either fewer than fifty primary care physicians or fewer than fifty enrollees. HRRs with no MA network were excluded from the analysis. Appendix section 3 details the effects of these selection criteria on our sample.¹⁸

Statistical Analyses To Measure Primary Care Physician Costliness

To assess primary care physicians’ costliness, we estimated a linear regression model of patients’ annual costs that included a fixed effect for each primary care physician, controlled for patients’ county of residence and their risk score, and also for the year. After estimation, the primary care physician fixed effects were “shrunk” toward their county average, using the Empirical Bayes method.¹⁹ The resulting primary care physician effects were then standardized to the average in their HRR. See appendix section 5 for further details on these calculations.¹⁸

The primary care physician–level fixed effects measured the extent to which average patient spending for each primary care physician was higher or lower than the costs predicted by patient risk scores and county of residence. As risk scores and county determine payments to MA plans, the estimated fixed effects can be interpreted as the relative costliness of each primary care physician to the MA plan.

To examine the extent to which our costliness measures reflected selection of patients to primary care physicians’ panels (versus direct effect of primary care physicians’ practice style), we examined patients’ characteristics by the costliness quintile of their primary care physicians. In addition, in a sensitivity analysis, we estimated an alternative model of patients’ costs that also controlled for their race, Medicaid eligibility, disability status, and whether they died during the year.

Measuring Primary Care Physician Quality

For each physician, we computed eight Healthcare Effectiveness Data and Information Set²⁰ (HEDIS) quality measures that could be calculated using claims for their assigned traditional Medicare patients. We included two measures for comprehensive diabetes care, one for breast cancer screening in women at ages 65–69, one for osteoporosis management in women who had a fracture, two for pharmacotherapy management of chronic obstructive pulmonary disorder exacerbation, and two for annual monitoring for patients on persistent medications. The measures were mostly uncorrelated (see exhibit S22 in appendix section 8).¹⁸

Calculating Network HRR–Level And National-Level Quality And Costliness Measures

We calculated HEDIS quality measures at the network HRR level as the mean quality of in-network primary care physicians. We then calculated the “network HRR quality,” a composite quality measure that is the simple average of all HEDIS measures in the network HRR. Similar to the underlying HEDIS measures, the composite measure ranged from 1 to 100 percent, indicating the average share of patients who received care according to clinical guidelines.

“Network HRR costliness” is the mean costliness of primary care physicians included in each network HRR. For primary care physicians excluded from all MA networks, we calculated quality and costliness at the HRR level, as if they formed a separate network HRR. We also calculated the overall “HRR costliness” and “HRR quality,” using the means of all primary care physicians in the HRR. These HRR means serve as benchmarks for comparison of network HRRs’ costliness and quality.

We calculated the costliness and quality gaps between each network HRR and the HRR mean. These gaps measured how different the average primary care physician in the network was from the surrounding market and could indicate the extent of selective contracting. For national inference we calculated an average of the network HRR measures, weighted by the number of MA enrollees who used each network HRR (clustering the standard errors at the network HRR level). For primary care physicians who were excluded from all MA networks, we calculated an average of the HRR-level means, weighted by the number of traditional Medicare beneficiaries in each HRR (clustering errors at the HRR level).

Heterogeneity

We examined heterogeneity in our results by network breadth. Network HRRs were defined as narrow if they included at most 15 percent of the primary care physicians in the HRR, and were defined as wide if they included at least 35 percent of primary care physicians. We also examined heterogeneity among the five largest MA insurers (see appendix section 6).¹⁸

Limitations

Our approach relied on a key assumption, which is that primary care physicians’ costliness and quality in traditional Medicare and MA were strongly correlated, so information on the treatment of traditional Medicare patients was relevant to MA selective contracting. The required assumption is weaker than the “norms hypothesis,” which postulates that physicians adapt a mostly uniform practice style, best suited to the insurance mix across their patients.²¹ Our assumption allowed us to interpret within–traditional Medicare results as indicative of primary care physician–driven costs and quality differences between MA and traditional Medicare patients. Our approach has the advantage that MA plans’ incentives to physicians or patients do not directly affect traditional Medicare patients’ care, which therefore better captures the practice style and coding practices of primary care physicians. A limitation of our approach is that primary care physicians that exclusively treat MA patients are not represented in our sample.

There are additional limitations in our use of contemporary traditional Medicare data. First, MA policies may have spillovers to traditional Medicare patients and affect their care.^22–26 If spillovers from MA were substantial for primary care physicians, any cross-sectional difference we observed may partly have been the result of inclusion in MA networks, and not the reason for this inclusion. Second, some primary care physicians in our sample had a small panel of associated traditional Medicare patients, causing their costliness measures to be heavily shrunk to the county average. However, our results remained very similar when we excluded network HRRs where primary care physicians had an average panel of fewer than fifty traditional Medicare patients.

Inaccurate provider directories may have introduced errors into the Ideon data set. However, we verified primary care physicians’ specialties and locations, using other data sources.

Last, our costliness measures were estimated using patient-year observations, ignoring possible serial correlation between the predicted versus actual costs of the same patient over the years.

Study Results

We linked data on 4,456,037 traditional Medicare patients (9,975,761 patient-years) to 151,679 primary care physicians. During 2016–18, the average primary care physician had sixty-six assigned patient-years from the 20 percent Carrier file. Most primary care physicians (81 percent) were included in at least one MA network, ranging from 69 percent in the fifth percentile of HRRs to 90 percent in the ninety-fifth percentile. Women made up 58 percent of the assigned beneficiaries, and the average age was seventy-four. Most patients were White (85 percent), with 7.8 percent Black, 1.5 percent Hispanic, and 1.9 percent Asian. Patients’ mean annual costs were $15,431. The national HRR-level mean of patients’ costs was $14,999 (data not shown.)

Our sample included 3,719 network HRR combinations: 495 networks in 299 HRRs, with 363 of these networks spanning multiple HRRs (ten on average). These network HRRs were used by 13,756,550 MA enrollees, out of 23.3 million MA enrollees in December 2019. The average network HRR operated in an HRR with 1,905 primary care physicians (median: 1,166) and seventeen MA networks (data not shown).

Exhibit 1 presents unweighted summary statistics for the network HRRs in our sample and for primary care physicians excluded from MA. Network HRRs had a mean of 416 in-network primary care physicians (27 percent of all primary care physicians in the HRR).

Exhibit 1:

Characteristics of Medicare Advantage (MA) primary care physicians (PCPs), by participation in MA networks and by network breadth, 2019

Characteristics	TM-only PCPsa	MA PCPsb	Narrow networksc	Wide networksc
PCPs
Female PCPs, %	30	37	37	38
Years since graduation	23	24	24	23
Group affiliation, %	86	87	83	92
Hospital affiliation, %	85	88	80	94
In ACO, %	43	55	51	59
Patients
No. of Attributed patient-years per PCP	33	82	73	82
Mean age, years	73	73	73	73
White, %	84	86	82	91
Black, %	8	7	8	5
Dual eligibility, %	19	17	20	14
Annual mortality rate, per 1,000	41	27	27	27
MA Network-HRRs
Network-HRR breadth, %d	—e	27	9	43
No. of In-network PCPs per MA network	—e	416	235	519
No. of PCPs with attributed TM patients	97	306	172	379
Sample Counts
No. of TM beneficiaries	39,997,547	—e	—e	—e
No. of Enrolled MA beneficiaries	—e	13,756,550	3,139,735	3,685,475
No. of Sample PCPs	28,504	123,175	19,230	44,097
No. of Network-HRRs	—e	3,719	848	1,002

SOURCES Authors’ analysis. For PCP participation in MA networks, we used 2019 data from Ideon. PCP characteristics were drawn from the Centers for Medicare and Medicaid Services’s December 2018 Care Compare Doctors and Clinicians file, from 2019 IQVIA data, and from 2019 Ideon data. Associated patients’ data came from 2016–18 claims files for a 20 percent sample of traditional Medicare beneficiaries.

NOTE

^aMeans were calculated at the HRR level for PCPs who treated only traditional Medicare (TM) patients and then averaged nationally.

^bFor PCPs who participated in MA networks, means were calculated at the network HRR level and then averaged nationally.

^cMeans were calculated at the network HRR level and then averaged nationally. Statistics were calculated separately for PCPs who participated in narrow MA networks (defined as including at most 15 percent of PCPs in the HRR) and PCPs who participated in wide MA networks (defined as including at least 35 percent of the HRR’s PCPs).

^dThe breadth of each network HRR was measured as the share of the HRR’s PCPs who were in-network.

^eNot applicable.

Network HRRs had an average of 306 primary care physicians with associated traditional Medicare patients in our sample, with each primary care physician having an average of eighty-two attributed patient-years. In each HRR, there were, on average, ninety-seven primary care physicians who were excluded from all MA networks. These primary care physicians had an average of thirty-three traditional Medicare patient-years. The share of women was higher among primary care physicians who participated in MA networks, and the share of primary care physicians who participated in a Medicare accountable care organization was higher among those who participated in MA networks. The mortality rate of patients associated with MA primary care physicians each year was lower by 34 percent than the annual mortality rate for patients of primary care physicians excluded from MA (twenty-seven deaths per 1,000 population versus forty-one deaths).

Costliness Of Primary Care Physicians

The average primary care physician who participated in MA networks was $375 less costly than the average of all primary care physicians in the HRR (exhibit 2). Weighted by enrollment, the average primary care physician who participated in the average MA network was $433 less costly than the average of all primary care physicians in the HRR (2.9 percent of the national HRR-level mean of patients’ costs). Exhibits S19 and S20 in appendix section 7 demonstrate that primary care physicians who generated lower costs were included in more network HRRs, and network HRRs with less-costly primary care physicians had higher MA enrollment.¹⁸ Primary care physicians in the average narrow network were $212 (1.4 percent of baseline) less costly than primary care physicians in the average wide network (exhibit 3). Last, the small group of primary care physicians who served only traditional Medicare patients was $1,534 more costly than the average of all primary care physicians in the HRR (exhibit 2).

Exhibit 2:

Primary care physician (PCP) costliness and quality, by inclusion in Medicare Advantage (MA) networks, 2019

Costliness and quality, relative to HRR mean	TM-only PCPs	95% CIs	MA PCPs	95% CIs	MA network HRRs	95% CIs	MA versus TM-only difference	95% CIs
PCP costliness, $a,b	1,534	1,524, 1,543	−375	−377, −373	—e	—e	−1,909	−1,915, −1,903
Network HRRs’ costliness, $ (weighted by enrollment)a,c	1,617	1,499, 1,734	—e	—e	−433	−469, −397	−2,050	−2,172, −1,927
Network HRRs’ quality, percentage points (weighted by enrollment)d	−2.1	−2.4, −1.7	—e	—e	0.1	−0.3, −0.4	2.1	1.7, 2.6

SOURCES Authors’ analysis. For PCP network participation, we used 2019 data from Ideon. PCPs’ hospital referral region (HRR) location was determined using 2019 Ideon data, 2019 IQVIA data, and December 2018 data from the Centers for Medicare and Medicaid Services Care Compare Doctors and Clinicians file. Costliness and quality measures were estimated using 2016–18 claims files for a 20 percent sample of traditional Medicare (TM) beneficiaries.

NOTE

^aPCPs’ costliness was measured as the difference between the observed annual costs of their TM patients and their predicted costs, based on risk scores and county of residence. All measures in the table examine differences in costliness relative to the average of all PCPs in the same HRR. The national average of HRR-level means of costs per patient-year in our sample of patients associated with TM PCPs was $14,999. Annual cost data included the amount spent by Medicare, out-of-pocket payments, and payments from other insurers. This included spending on prescription drugs in Part D.

^bAverage costliness was examined at the PCP level.

^cAn average of the mean costliness of PCPs in each network HRR was calculated, weighted by the number of MA enrollees using each network HRR. For PCPs not participating in any MA network, we calculated the average of HRR-level means, weighted by the number of TM beneficiaries in the HRR.

^dAn average of the mean quality of PCPs in each network HRR was calculated, weighted by the number of MA enrollees using each network HRR. For PCPs not participating in any MA network, the measure is an average of HRR-level means, weighted by the number of TM beneficiaries in the HRR. The mean quality in our sample was 69.9 percent.

^eNot applicable.

Exhibit 3:

Primary care physician (PCP) costliness and quality, by network breadth, 2019

Costliness and quality, relative to HRR mean	Narrow networksa	95% CIs	Wide networksb	95% CIs	Wide versus narrow difference	95% CIs
PCP costliness, $c,d	−466	−473, −459	−317	−319, −315	149	143, 155
Network HRRs’ costliness, $ (weighted by enrollment)c,e	−557	−676, −438	−345	−366, −325	212	91, 333
Network HRRs’ quality, percentage points (weighted by enrollment)f	−0.8	−1.9, 0.2	0.3	0.3, 0.4	1.1	0.2, 2.2

SOURCES Authors’ analysis. For PCP network participation and networks’ breadth, we used 2019 data from Ideon. PCPs’ hospital referral region (HRR) location was determined using 2019 Ideon data, 2019 IQVIA data, and December 2018 data from the Centers for Medicare and Medicaid Services Care Compare Doctors and Clinicians file. Costliness and quality measures were estimated using 2016–18 claims files for a 20 percent sample of traditional Medicare beneficiaries.

^aNarrow MA networks include at most 15 percent of the PCPs in the HRR.

^bWide MA networks include at least 35 percent of the PCPs in the HRR.

^cPCPs’ costliness was measured as the difference between the observed annual costs of their TM patients and their predicted costs, based on risk scores and county of residence. All measures in the table examine differences in costliness relative to the average of all PCPs in the same HRR. The national average of HRR-level means of costs per patient-year in our sample of patients associated with TM PCPs was $14,999. Annual cost data included the amount spent by Medicare, out-of-pocket payments, and payments from other insurers. This included spending on prescription drugs in Part D.

^dAverage costliness was examined at the PCP level.

^eAn average of the mean costliness of PCPs in each network HRR was calculated, weighted by the number of MA enrollees using each network HRR. For PCPs not participating in any MA network, we calculated the average of HRR-level means, weighted by the number of TM beneficiaries in the HRR.

^fAn average of the mean quality of PCPs in each network HRR was calculated, weighted by the number of MA enrollees using each network HRR. For PCPs not participating in any MA network, the measure is an average of HRR-level means, weighted by the number of TM beneficiaries in the HRR. The mean quality in our sample was 69.9 percent.

Exhibit S13 in appendix section 5 presents the characteristics of patients by the costliness quintile of their associated primary care physicians.¹⁸ Compared with the lowest quintile, primary care physicians in the highest (most costly) quintile served more Black patients, fewer Asian patients, and more patients who are dually eligible for Medicare and Medicaid; the mortality rate among their patients was 2.4 times higher.

Primary Care Physician Quality

The quality of primary care physicians in the average MA network HRR was similar to the quality of all primary care physicians in the HRR, just 0.1 percentage point above the HRR mean (exhibit 2). The quality of primary care physicians excluded from all MA networks in the average HRR was lower than the quality of all primary care physicians in the HRR, by 2.1 percentage points (exhibit 2). We found that the quality of primary care physicians in the average narrow network HRR was slightly lower than that of primary care physicians in the average wide network HRR, by 1.1 percentage points (exhibit 3). Exhibit S21 in appendix section 7 presents the eight underlying HEDIS measures.¹⁸

Exhibit 4 presents primary care physician costliness and quality in 892 network HRRs with above-average MA enrollment (at least 3,700 enrollees) and the HRR-level means of costliness and quality for primary care physicians excluded from MA. These excluded primary care physicians (marked by black dashed circles) were markedly concentrated in the lower-right quadrant, indicating that they had both higher costliness and lower quality compared with the means for all primary care physicians in their HRRs. MA networks were mostly concentrated in the two left quadrants, indicating lower costliness than their HRR average. The figure demonstrates that the costliness and quality in narrow networks varied more than in wide networks, and that many narrow networks had markedly lower costliness or quality, or both, than the average wide network. Exhibit S18 in appendix section 7 presents 95% confidence ellipses of costliness and quality for the whole sample.¹⁸

Exhibit 4: PCP Costliness and Quality, by Inclusion In MA Networks and by Network-HRR Breadth, 2019 and by Network-HRRs Breadth.

Source/Notes: SOURCE Authors’ analysis. For primary care physician network participation and networks’ breadth, we used 2019 data from Ideon. Primary care physicians’ HRR location was determined using 2019 Ideon data, 2019 IQVIA data, and December 2018 data from the Centers for Medicare and Medicaid Services Care Compare Doctors and Clinicians file. Costliness and quality measures were estimated using 2016–18 claims files for a 20 percent sample of traditional Medicare beneficiaries.

NOTES The scatter plot shows the mean costliness and quality of primary care physicians in 892 network HRRs with above-average MA enrollment (at least 3,700 enrollees). Network-excluded primary care physicians are treated as if they formed separate network HRRs (“TM-only HRRs”), with their cost and quality calculated at the HRR level. Narrow MA networks include at most 15 percent of the primary care physicians in an HRR. Wide MA networks include at least 35 percent of the primary care physicians in an HRR. Network HRRs with mean costliness less than −$2,000 and greater than $2,000 were excluded. For network HRRs, circle size represents the relative number of MA enrollees using each network HRR. For network-excluded primary care physicians, circle size represents the relative number of traditional Medicare beneficiaries in each HRR.

Sensitivity Analysis

Our results were robust to estimating costliness using an alternative model that controlled for a richer set of patients’ characteristics (exhibit S11 in appendix section 5),¹⁸ shrinking our main estimates by only 11–13 percent.

Discussion

In this national study, we found that primary care physicians who participated in Medicare Advantage networks were less costly to MA plans than the average of all primary care physicians in the same HRR while providing similar quality of care. These findings provide insights into one mechanism that may lower MA plans’ spending and increase their profits: selective contracting with providers. Whether plans directly choose which providers to include in their networks or offer contract terms that deter certain providers from joining, the resulting selection of primary care physicians has the potential to improve plans’ margins. Additional evidence for selective contracting comes from examining primary care physicians who were excluded from all MA networks; these primary care physicians were markedly more costly and had lower quality than the average for all primary care physicians in their HRRs. We found that selection of less-costly primary care physicians was stronger in narrow MA networks compared with in wide networks and that there was slightly lower average quality in these narrow networks.

Relatively little research has studied the extent to which MA plans select providers on the basis of their performance. One study suggested that MA plans selected average-quality hospitals, excluding both high- and low-quality hospitals.²⁷ Others studied the star ratings of plans with narrow versus wide networks without directly examining physicians in these networks.^8–10

Because patients were not randomly assigned to primary care physicians, our costliness measure is not the causal effect of primary care physicians’ practice style on costs. Our findings that patients of the costliest quintile of primary care physicians had higher mortality rates and higher shares of Black patients suggest that costliness was indeed partly a result of selective sorting of patients to primary care physicians. However, from the point of view of the MA plan, a primary care physician is less costly either if she causally decreases the costs of her patients or if she manages to attract less-costly patients to begin with. Hence, selective contracting with primary care physicians may also function as a mechanism through which plans can select more-profitable enrollees. Primary care physicians may also be less costly to MA plans if they code their patients more intensively than the average primary care physician in their area, increasing plans’ revenues without a similar increase in costs. Our costliness measure essentially sums up these three possible effects, all of which influence the profits of MA insurers, and thus may affect selective contracting. Our results were robust to estimating costliness, using an alternative model that controlled for a richer set of patients’ characteristics. This could indicate that the lion’s share of our costliness measure could be attributed to primary care physicians’ causal effect on costs.

Our finding that narrow MA networks included less-costly primary care physicians is consistent with prior evidence that plans with narrow networks have lower premiums,²⁸ they reduce quantities of care and prices paid to providers, and they shift care from hospitals and specialists to primary care.²⁹ Several studies have demonstrated how excluding costly hospitals may allow insurers to avoid unprofitable enrollees.^30,31

Conclusion

Our findings suggest that managed care tools, particularly selective contracting with primary care physicians, contribute to lower costs in Medicare Advantage. This may create a trade-off for policy makers shaping network adequacy regulations. The more the rules require improved access, the more likely they are to limit the scope of selective contracting, potentially increasing costs for MA insurers, limiting their ability to offer additional benefits to enrollees. Although our results indicate that selective contracting by primary care physician costliness is mainly related to primary care physicians’ practice style, selective contracting may also serve as a mechanism for patient selection by MA plans. When evaluating MA networks, regulators could scrutinize efforts by MA plans to select healthier patients through selective contracting with physicians.