Tiered Physician Network Plans and Patient Choices of Specialist Physicians

Key Points Question Are tiered physician network health plans associated with patient choices of physicians over time, and do the associations vary with patient income? Findings In this cross-sectional study of a setting where tiered physician networks have been used for a decade, there was no association between physician tier ranking and market share among any group of patients. However, patients with low income lived closer to physicians with lower copays and paid slightly lower out-of-pocket costs for tiered physician office visits than patients with high income. Meaning This study suggests that maintaining the saliency of physician tiers for patients may be important for prolonging the benefits associated with tiered networks.


Research overview
We used the largest employer-based purchaser of health insurance in Massachusetts to study the impacts of tiered physician networks on patient choices of specialist physicians in a setting where tiered networks have been in place for several years.The GIC tiered physician networks include tiered copayments for office visits with a specialist physician.Members paid the same deductible and had the same out-of-pocket maximum regardless of their physician's �er.About half of the plans also had �ered hospital networks, where inpa�ent admission copays varied with the �er of the hospital, but this was a separate �ering program.The only variation in employee premium contribution is based on employee date of hire.Members hired before FY2003 paid 20% of the monthly premium out-of-pocket.Members hired after FY2003 paid 25% of the monthly premium out-of-pocket.
If pa�ent choices are unaffected by �ered networks, then �ered networks will not be a useful tool for encouraging pa�ent choice of low-cost or high-quality care.Answering this question required going beyond simple comparisons of patient volumes across low-copay-tier and high-copay-tier physicians, because physician tier ranking is correlated with other physician characteristics preferred by patients.We therefore used regression discontinuity design, an analy�cal approach that iden�fied the associa�on of a physician's �er rankings with study outcomes.Our second research question was how the tiered physician networks differentially affected low-income pa�ents.Even if �er rankings did not affect pa�ent choices, �ered physician networks could still exacerbate health care disparities if the physicians serving these communities were disproportionately likely to be in the least favorable, highest-copay tier.We therefore examined the distribu�on of �er rankings within specialty across pa�ent incomes.

Details of GIC Scoring and Tiering Methodology
Physicians in tiered specialties were assigned to one of three tiers in each health plan in which they were in-network based on performance on a set of quality and cost-efficiency measures.
Within each plan-specialty-year, the majority of physicians-approximately 65 percent of each specialty-were assigned to the medium-copay tier (Tier 2).Another 20 percent or so were assigned to the low-copay tier (Tier 1), leaving the remaining approximately 15 percent to be assigned to the high-copay �er (Tier 3).However, each carrier set the performance cutoffs for each �er, resul�ng in slight differences in the cutoffs across carriers.Physicians' efficiency and quality scores were calculated using all commercial claims from par�cipa�ng insurers, including claims from the insurers' non-GIC plans; that is, they were calculated using the insurers' full books of business.This method increased the sample size for a physician's score calculations, while avoiding incentivizing physicians to treat their GIC pa�ents differently from other commercially insured pa�ents.

Efficiency scores.
Physician efficiency was a summary measure of the per-episode quan�ty and complexity of care a physician's pa�ents received, separate from heterogeneity in physician prices.The measure was case-mix adjusted, meaning physicians' efficiency scores were adjusted to compare observably similar patients.More specifically, medical claims were binned into episodes using Optum's proprietary Symmetry Episode Treatment Groups (ETG) software.This software grouped together claim lines into mutually exclusive sets of relatively self-contained, interrelated claims.It aimed to atribute follow-on care and care due to complications to the same episode as the original health care service.For example, all health care services received for surgical removal of kidney stones for a pa�ent without other comorbidi�es, including follow-up care, were grouped into a single episode of care.
Each episode was attributed to a single physician, typically the physician whose claim lines generate the plurality of the spending within the episode.Importantly for episode attribution, the GIC uses a unique physician identifier that can be linked across GIC insurers and over �me, constructed as part of the data infrastructure suppor�ng the GIC's �ering efforts.This physician iden�fier is constructed from a range of informa�on including name, prac�ce location, and license information.The result is more reliable for uniquely identifying physicians than tax IDs or National Provider Identifiers (NPIs) because it is never shared with other physicians in the same prac�ce.
Each physician's efficiency for each episode type was then calculated as the average quantity of services provided under that episode type (adjusted for patient severity) multiplied by a standardized price schedule regardless of the identity of the physician.This price standardization abstracted away from the differences in negotiated prices across physicians and compared physicians by treatment styles.As a result, physicians were unable to change their efficiency score, or their tier-ranking, by changing negotiated prices.
Episode types were fairly granular; for example, surgical removal of kidney stones for a pa�ent without other comorbidi�es was a separate episode type from surgical removal of kidney stones for a patient with comorbidities.Each physician's final efficiency score used for tiering was equal to the average of her efficiency performance across episode types (weighted by episode type volume).

Quality scores.
Physician quality was defined as percent compliance with prespecified clinical guidelines.
Specifically, quality was measured as the fraction of a physician's patients with a given medical condition for whom the guidelines are followed.Higher compliance rates translated to better quality scores.For example, endocrinologists were assessed primarily on quality measures related to diabetes care.A compliance rate was calculated for each physician-quality measure pair.Each physician's final quality score used for tiering was the weighted average of the relative compliance rates across quality measures (weighted by number of pa�ents per measure).

How efficiency and quality scores were used in tiering.
In principle, tier assignments were based on a combination of quality of care and resource efficiency (also described as intensity).In practice, tiers depended primarily on efficiency, but a small number of physicians with sufficiently low quality scores were always assigned to the highcopay �er.More specifically, physicians were placed into one of three quality groupings as a func�on of their quality score and the statistical confidence in that score.A small number of physicians with the very lowest quality scores were always assigned to the high-copay �er.For the majority of other physicians, quality score did not influence tier assignments with two exceptions.
First, a small frac�on of physicians with top quality scores who were placed in the best quality grouping realized an improvement in their final tiers.Specifically, physicians in the top quality grouping who missed the 20th percentile efficiency cutoff for the low-copay tier but who were under the 30th percen�le of efficiency within their specialty were moved from the mediumcopay tier to the low-copay tier.Second, physicians below the 10th percentile of quality within their specialty were at risk of reassignment to the high-copay tier regardless of their efficiency scores.We excluded physicians whose quality score had the poten�al to impact their �er eligibility in these ways by dropping the top and bottom quality groupings.This reduced the sample of physician-carrier-zip code-year observations by 19.21%.Physicians whose scores were missing or based on fewer than 30 observations were placed by default in the medium-copay tier.In our sample period, all GIC insurers used the same set of quality and efficiency measures to assign physicians to tiers.As a result, physicians' tier assignments were nearly identical across insurers.This motivated our analytical approach of comparing physicians on either side of a �er cutoff within an insurer rather than comparing within a physician across insurers as in Sinaiko and Rosenthal (2014). 1 The GIC required carriers to tier physicians based on these individual physician efficiency and quality scores until the start of FY2018.Starting in FY2018, two carriers began to use group-level �ering, whereby all physicians prac�cing in the same prac�ce loca�on within the same specialty were assigned the same �er.We dropped data from those two carriers star�ng in FY2018 because our regression discontinuity design is not valid under group-level tiering.In FY2015 only, one of the GIC carriers (Fallon) only used tiering for the larger of its two GIC plans.We retained in the sample both of Fallon's plans for all years; the FY2015 observations from the non-�ered plan contributed only 0.27% of the new pa�ent office visits in our sample.

Data linkage and defining physicians' geographic markets
To account for different patient densities across geographic markets, we measured each physician's market share only among patients who resided within a reasonable distance of the physician's primary practice location.We linked the claims data to physician data using a unique physician iden�fier and specialty designa�on that is constructed by the GIC for the purpose of tiering physicians and designed to be consistent across carriers, across physician prac�ce loca�ons, and over �me.
To arrive at the defini�on of a reasonable distance among all pa�ents treated by each specialty, we constructed the distribu�on of distances between the prac�ce loca�on zip code and patient residence zip code.We then measured the 90 th percentile of the distances traveled for care within each specialty.Finally, we defined each physician's potential patient market as all patients who saw a physician in that specialty and lived within a distance no farther than the 90 th percentile of distances traveled.We then calculated each physician's market share as their number of new pa�ents in the specialty-plan-year within the physician's zip code, divided by the total number of new pa�ents in the specialty-plan-year within the physician's zip code.Some physicians practiced in multiple zip codes and had multiple market share measures.

Regression Discontinuity (RD) design
We used a regression discontinuity (RD) study design to assess the association of tier ranking with physicians' market shares among new patients.Regression discontinuity analyses take advantage of clinical or policy decision rules that result in assignment to an interven�on on the basis of an arbitrary cutoff for a continuous variable. 2,3The validity of our RD design relies on the assump�on that all other determinants of a pa�ent's choice of specialist physician, except for the physician's assigned �er, would have a smooth rela�onship with efficiency scores if not for the arbitrary cutoffs.
The RD study design addresses omitted variable bias that would arise in a simple regression model.There were likely unobserved factors that drove both patient choice of physician and physician's tier assignment, for example, a physician's practice style.For example, some new patients may have avoided physicians who frequently recommend surgery over medical management.These physicians would also have had worse efficiency scores and would have been more likely to be in a higher-copay tier.Using a simple regression of market share on �er would therefore overstate the loss of pa�ent volume atributable to being in the higher-copay tier, since it would not account for some new patients' underlying preferences to avoid the surgeryrecommending physicians.
To implement the RD analyses, we focused on a subset of physicians whose �er was quasirandomly assigned.GIC carriers determined �er rankings by first checking whether a physician passed a minimum quality threshold, and if so, then assigning a tier ranking depending on whether the physician's efficiency score fell above or below a specialty-plan-year specific cutoff.While there were meaningful differences across physicians that may have affected pa�ent choices, physicians whose efficiency score placed them just to either side of the cutoff between the lowand medium-copay �ers or the cutoff between the medium-and high-copay �ers should not have had discon�nuous jumps in any other atributes that would impact a patient's choice to see them for care.There was also no reason to expect any other underlying differences in care provided by physicians on either side of these cutoffs other than those stemming from the impact of the �er ranking.
We checked the underlying assumptions of the RD design using widely accepted methods for RD validation. 4 First, to facilitate comparisons only between physicians whose efficiency scores were similar, we used narrow bandwidths to subset to physicians whose scores are close to the cutoff.We used data-driven mean squared error-op�mal bandwidths for our primary specifica�on, and we tested the sensi�vity of our results to other choices of the bandwidths.We also used density tests to check for discon�nuous jumps in the distribu�on of physicians' continuous scores or their attributes on either side of the tier cutoffs.Finally, we tested the sensi�vity of our results to different choices of polynomial degrees.
Our main analyses were estimated using the following linear regression: The outcome variable,   ℎ  , measured physician 's share in year  of all GIC pa�ents seeing a physician in that specialty for the first �me.This new pa�ent share variable the difference between the physician's efficiency score and the nearest �er cutoff.For example, in a specialty where an efficiency score of 0.7 or lower was required for the low-copay �er and a score of 1.4 or higher was required for the high-copay �er, a physician with a score of 0.68 would be closer to the cutoff between the low-copay and medium-copay �ers and would be coded as having a difference of -0.02 points to the cutoff with a nega�ve value to indicate being on the low-copay side of the cutoff.

Relative copayment differences: Difference-in-Differences design
Our secondary analysis used a difference-in-differences design with continuous treatment to examine the association between copayment amounts (the monetary dimension of tiering) and physicians' new patient market share.This analysis leveraged variation in copayments withinphysician due to administrative changes in copayments in FY2016 and FY2018.These analyses were estimated using the following linear regression: The outcome variable,   ℎ  , was defined as in the RD analysis.The explanatory variable of interest,   , was defined as the difference between the highcopay (�er 3) copayment amount in year  and the physician 's copayment amount in year .(In just the subset analysis using only the subsample of physicians who were in �er 3, this variable is redefined rela�ve to the lowest-copay �er's copayment amount.)If   > 0, then new pa�ent market share is higher when a physician's copayment amount is much smaller than the high-copay �er's copayment amount; in other words, new pa�ent market share is higher when pa�ents can save more money by choosing that physician.The physician fixed effects,   , ensured that   was es�mated by making only within-physician, within-�er comparisons to isolate the effect of changing copayment amounts.These fixed effects were defined at the granular level of physician-�er-carrier cells, but we use the   nota�on as a shorthand.These fixed effects controlled for persistent differences across physicians in their prac�ce styles, bedside manners, and other persistent characteris�cs.These difference-in-differences analyses thus measure the rela�onship between new pa�ent market share and the financial differences pa�ents face across �ers.

eFigure 1. Distribu�on of efficiency scores rela�ve to �er cutoffs
Authors' analysis of Group Insurance Commission (GIC) physician-year efficiency scores used for �er assignment.Notes: The horizontal axis represents a physician's score rela�ve to the cutoff score between �ers.The cutoffs between �ers vary by specialty, year, and insurance carrier.The plots are therefore normalized by subtrac�ng the cutoff for the corresponding specialty, year, carrier, and �er pair from the physician's raw score.The top panel plots the score rela�ve to the cutoff between the low-copay and medium-copay �ers.The botom panel plots the score rela�ve to the cutoff between the medium-copay and high-copay �ers.Scores to the right of the cutoff (0 on the horizontal axis) place physicians into higher-copay �ers.eFigure 2. Associa�on of assignment to higher-copay �er with new pa�ent market share using regression discon�nuity, by salary of primary subscriber Notes: The horizontal axis shows the es�mated percentage point impact of assignment to a higher-copay �er on a physician's market share among new pa�ents.Standard errors are clustered by physician.Error bars represent 95% CIs.A nega�ve and sta�s�cally significant coefficient would be consistent with pa�ents being less likely to choose physicians in highercopay �ers.Es�mates with error bars that did not cross the ver�cal line at zero were significantly different than zero.Baseline includes all new pa�ents.eFigure 3. Associa�on of difference between physician's copayment and highest �ered copayment with new pa�ent market share, difference-in-difference analysis Notes: The horizontal axis shows the es�mated impact in percentage points of the difference between a physician's copayment and the highest �ered copayment on a physician's market share among pa�ents choosing a physician for the first �me.Standard errors are clustered by physician.Error bars represent 95% CIs.A posi�ve and sta�s�cally significant coefficient would be consistent with pa�ents being less likely to choose physicians in higher-copay �ers.Es�mates with error bars that did not cross the ver�cal line at zero were significantly different than zero.Notes: Each curve plots the share of pa�ents of a given income range who live no farther than the ploted distance from a low-or medium-copay physician in that specialty.Income is measured as median household income in the pa�ent's zip code.Distances are calculated between 5-digit zip code centroids; if a pa�ent lives in a zip code in which at least one low-or medium-copay is also located, the calculated distance is zero.

eTable 1 .
Evalua�on and management (E&M) codes used to select study sample These Current Procedural Terminology (CPT) codes for Evalua�on & Management (E&M) visits are used to select the sample of pa�ents newly seeing a specialist within a given specialty ("new pa�ent visits").This excludes E&M codes for established pa�ents.

eTable 2 .
Distribu�on of ac�ve state employees by salary band, FY2015-2019Authors' analysis of Group Insurance Commission (GIC) primary subscriber wage data.Notes: FY = Fiscal Year.Fiscal Years run from July 1 of the prior year through June 30 (e.g., FY2015 is July 1, 2014 -June 30, 2015).

eFigure 4 .
Travel distance to nearest low-or medium-copay physician, by household income and specialty