2. Background

Ten states passed laws protecting patients from OON balance bills in non-emergency, elective situations between 2013 and 2018: Arizona, California, Connecticut, Florida, Maine, Minnesota, New Hampshire, New Jersey, New York and Oregon. All of these states protected patients from balance billing by OON ancillary physicians, using a variety of methods to regulate OON reimbursement. The laws either used a specific regulated payment or employed an IDR process to determine reimbursement by the insurer to the OON provider. The former regulated payment states employed a variety of payment rules. The latter IDR states used a variety of arbitration mechanisms to select cases eligible for arbitration and guide arbiters in their decision-making. Table 1 summarises the ten surprise billing laws.

Table 1. State surprise billing laws

Five of the ten state laws specified a regulated amount for OON reimbursement of physicians. California's law was effective 1 July 2017 and required that health plans pay OON physicians working at in-network hospitals the greatest of the average contracted rate or 125 per cent of the amount Medicare reimburses on a fee-for-service basis for the same or similar services in the general geographic region in which the services were rendered.Footnote ⁴ Connecticut's law, which started on 1 July 2016, specified that OON physicians providing non-emergency services at in-network hospitals be reimbursed at the health plan's in-network rate if the health plan and OON provider do not agree on a reimbursement amount.Footnote ⁵ Maine's law, which became effective on 1 January 2018, uses the same approach as Connecticut for reimbursement of OON providers working in in-network hospitals.Footnote ⁶ Florida's law became effective on 1 July 2016 and requires that the health plan reimburse the OON provider the lesser of (1) the provider's charges, (2) the usual and customary provider charges for similar services in the community where the services were provided or (3) the charge mutually agreed to by the insurer and the provider within 60 days after submittal of the claim.Footnote ⁷ Oregon's law, which started on 1 March 2018, required OON reimbursement to equal the median in-network rate in 2015 (adjusted for subsequent inflation).Footnote ⁸

The remaining five states relied on an IDR process for reimbursement of OON physicians practicing at in-network hospitals. Arizona's law became effective in January 2019 and it used an IDR process, but the patient must initiate it and the bill only qualifies for arbitration if the balance bill and cost-sharing is greater than $1,000.Footnote ⁹ Minnesota's law became effective in January 2018 and implements an IDR process if the health plan and physician cannot agree to reimbursement. If arbitration is necessary, the arbiter is directed to consider the health plan's other OON payments for similar services, the complexity of the treatment and the usual and customary payment calculated from an independent claims database (e.g. FAIR Health).Footnote ¹⁰ New York's law – which was signed by the governor in October 2014, but became effective in April 2015 – employed final offer arbitration for OON reimbursement disputes in which the arbiter is shown the 80th percentile of charges as calculated by FAIR Health (Adler, Reference Adler2019). New Hampshire's law became effective on 1 July 2018 and also employs an IDR system, but it relies on the state insurance commissioner's determination of the reasonableness of the insurer's payment to qualify for arbitration.Footnote ¹¹ New Jersey's law protecting patients from balance bills for OON care occurring in in-network hospitals became effective on 30 August 2018 and uses an IDR system based on charge percentiles (Chartock et al., Reference Chartock, Adler, Ly, Duffy and Trish2021). However, the difference between the insurer's offer and the provider's offer must be greater than $1000 (at the episode level) to qualify for the IDR system. Prior to this law, New Jersey also banned balance billing by OON ancillary physicians, but required that insurers pay the OON physician's full charge. Some argued that this led to reduced levels of network participation for physicians (Mattke et al., Reference Mattke, White, Hanson and Kotzias2017). New Jersey's new law allows self-funded health plans to opt into the protections of the law and its IDR system.

3. Data and methods

We use the Health Care Cost Institute's (HCCI) 2.0 Commercial Claims Research Dataset for years 2012 through 2019. The data capture the medical claims for over 55 million covered lives in employer sponsored health plans and represent multiple insurers, including Aetna, Blue Cross and Blue Shield, and Humana.

We analyse prices, charges and network participation for three hospital-based physician specialties: anaesthesiologists, pathologists and radiologists. For anaesthesiologists,Footnote ¹² we isolate the physician claims with Current Procedural Terminology (CPT) codes 100 through 1999. For pathologists, we isolate the physician claims with CPT codes 88300 through 88399. For radiologists, we isolate the physician claims with CPT codes 70010 through 79999, excluding CPTs 77261 through 77799. For all three specialties, we restrict the claims to those with place of service codes indicating a hospital (either outpatient or inpatient) and also exclude claims that may not represent commercially insured patients with managed care plans or that have other coding problems.Footnote ¹³

The HCCI claims data do not include the provider's charge. For anaesthesiologists and pathologists, the HCCI claims data are merged with charge information from Medicare's Provider Utilization and Payment (MPUP) data for 2013 through 2019Footnote ¹⁴. The MPUP data include the provider's average charge for Medicare patients for each CPT and year when the place of service is an inpatient hospital. The MPUP data are merged with the anaesthesiology and pathology physician claims by the provider's masked NPI, CPT and year. (2012 MPUP data are not available. Analysing charges for radiologists is not feasible because MPUP data are available for too few radiologists.)

Payments for anaesthesiologists and nurse anaesthetists are based on a different system than payments for other physicians.Footnote ¹⁵ Anaesthesiology payments are calculated using base units, time units, anaesthesiology-specific conversion factors that differ by region, and other adjustments depending on whether services were provided by an anaesthesiologist or by a nurse anaesthetist under the supervision of an anaesthesiologist. To case-mix-adjust prices and charges, we use base unit and anaesthesia conversion factor files for 2012 through 2019 from the Centers for Medicare and Medicaid Services (CMS).Footnote ¹⁶ The base unit files are merged with the anaesthesia claims by CPT and the anaesthesia conversion factors are merged with the anaesthesia claims by zip code.

For pathologists and radiologists, we use Medicare physician fee schedule files for 2012 through 2019 from CMS to case-mix-adjust prices and charges.Footnote ¹⁷ Using the Medicare physician fee schedule files, we calculate the total geographically adjusted relative value units (RVUs) for each five-digit zip code and CPT when the service is performed in a facility setting. We calculate the work RVUs multiplied by the work geographic practice cost index (GPCI) plus the facility practice expense RVUs multiplied by the practice expense GPCI plus the malpractice expense RVUs multiplied by the malpractice expense GPCI. This total geographically adjusted RVU amount is merged onto the claims data by zip code and CPT.

For pathology and radiology claims, we convert prices (and charges for pathology) to amounts that are comparable to Medicare's standard physician conversion factor (which is used to determine Medicare payment) by dividing the allowed amount (and, separately, the average charge for pathology) by the total geographically adjusted RVUs for each claim. This converts prices and charges to amounts per RVU. Because the geographic adjustment for anaesthesiology payments is included in the anaesthesiology conversion factors, we cannot use this same approach for anaesthesiology prices and charges. Instead, we calculate how much Medicare would have paid for each anaesthesiology claim by applying the base units, time units and region-specific conversion factors using Medicare's payment formula. We then divide the claim's allowed amount and (separately) the average charge by the hypothetical Medicare payment to convert anaesthesiology prices and charges into multiples of Medicare's payment. Whether prices and charges are amounts per geographically adjusted RVU for pathology and radiology or multiples of the payment that would apply if the patient had been a Medicare patient for anaesthesiology, our prices and charges are adjusted to reflect geographic and treatment intensity differences using Medicare's payment adjustments for these factors.

To measure network participation, we calculate the percentage of claims that are in-network. We separately measure outcomes for fully insured and self-funded claims. Summary statistics for price, charge and network participation outcomes are presented in Tables 2–4 for anaesthesiology, pathology and radiology, respectively.

Table 2. Anaesthesiology summary statistics for price (allowed/Medicare), charge/Medicare and network participation

Table 3. Pathology summary statistics for price (per RVU), charge (per RVU) and network participation

Table 4. Radiology summary statistics for price (per RVU) and network participation

Potential surprise billing episodes are defined as inpatient treatment episodes in which the facility is in-network, but at least one non-facility claim is OON. We identify patient treatment episodes as groups of inpatient facility and physician claims with the same patient ID and overlapping treatment dates. To focus on potential surprise billing episodes in elective situations, we exclude episodes that include claims for emergency services. We analyse potential elective surprise billing episodes regardless of physician specialty (i.e. the facility is in-network, but at least one physician, regardless of specialty, is OON), but we also analyse elective surprise billing episodes that are attributable to an OON anaesthesiologist, an OON pathologist and an OON radiologist, respectively. As with prices, charges and network participation, we separately estimate changes in potential surprise billing episodes for fully insured and self-funded health plans. Summary statistics for the prevalence of potential surprise billing episodes are presented in Table 5. For all analyses, we excluded 2020 due to the COVID pandemic.

Table 5. Elective potential surprise medical bill prevalence by type

To estimate the overall average effect of the state laws in each quarter after implementation, we use the event study method of Sun and Abraham (Reference Sun and Abraham2021). Estimating dynamic (i.e. by-period) average treatment effects using the standard two-way fixed-effects (TWFE) regression with treatment indicators for each period from treatment can lead to biased estimates if the treatments (e.g. state laws) are implemented at different times and the effects of treatment are heterogeneous. The method of Sun and Abraham (Reference Sun and Abraham2021) estimates dynamic average treatment effects accounting for treatment effect heterogeneity and staggered adoption. Specifically, we first estimate a TWFE model with dynamic time-from-treatment indicators interacted with treatment-time cohort indicators:

(1)$$Y_{ijt} = \alpha _i + \lambda _t + \mathop \sum \limits_{b\in B} \mathop \sum \limits_{l\in L} \mu _{lb}D_{ib}D_{it}^l + \epsilon _{ijt}, \;$$

where Y_ijt is the outcome of interest for provider j in state i in quarter t, α _i is a state-specific effect, λ _t is a quarter-specific effect, D _ib is an indicator that equals one for each state in treatment-time cohort b (taken from the set B of treatment-time cohorts), $D_{it}^l$ is an indicator that equals one when state i is l quarters from treatment, and $\epsilon _{ijt}$ is the idiosyncratic error. (For the analysis of surprise-billing episodes, j indexes episodes instead of providers and the estimation is based on a random sample of one million episodes.) From the estimation of equation (1), we recover μ _lb, which are average treatment effects for cohorts b∈B and quarters l ∈L from treatment. Next, we calculate weights as the share of cohort b in quarter l from treatment. Finally, we calculate a weighted average of μ _lb across cohorts as the average treatment effect in quarter l from treatment, using the cohort weights calculated in the previous step.Footnote ¹⁸ Following convention, we exclude indicators for l = −1 and set the quarter before treatment as the reference quarter. We restrict L such that l ≥ −12. This is the same approach for estimating dynamic treatment effects as in Garmon et al. (Reference Garmon, Li, Retchin and Xu2024).

The control group we use for the estimation of equation (1) is the set of states that passed laws in 2020 or 2021 that protect patients from surprise OON bills in elective situations. Chhabra et al. (Reference Chhabra, Sheetz, Nuliyalu, Dekhne, Ryan and Dimick2020) and Nikpay et al. (Reference Nikpay, Nshuti, Richards, Buntin, Polsky and Graves2022) highlight the significant geographic heterogeneity in surprise billing from hospital-based physicians, but offer no explanation for it. The differences in surprise billing prevalence across states may be due to factors that are difficult to observe and quantify. States with high rates of surprise billing (e.g. Florida, New York, New Jersey) are more likely to have passed a law to protect patients in response to the elevated risk. Thus, states that passed a surprise billing law may differ from states that never passed a surprise billing law and the differences may not be measurable. For this reason, we restrict our control group to states that passed an elective surprise billing law after our observation period (i.e. after 2019). These states are Georgia, Michigan, New Mexico, Ohio, Virginia and Washington.

The dynamic quarterly estimates of the overall average treatment effect may obscure differences across states in the impact of the laws. Heterogeneous effects across states that passed elective surprise billing laws are likely because of the various methods used to regulate OON reimbursements. To explore this potential heterogeneity, we estimate equation (1) separately by the type of state law: a regulated OON reimbursement vs an IDR system. We also estimate the average effect of each state law relative to controls. To estimate the effect of each state law, we use the method of synthetic controls (Abadie, Reference Abadie2021). Specifically, we estimate:

(2)$$Y_{it} = \alpha _i + \lambda _t + \mu D_{it}^{l \ge 0} + \epsilon _{it}, \;$$

where Y_it is the state-quarter mean, α _i is a state-specific effect, λ _t is a quarter-specific effect, $D_{it}^{l \ge 0}$ is an indicator that equals one for all observations in the treated state and all quarters after treatment, and $\epsilon _{it}$ is the idiosyncratic error. The treated state is stacked with its synthetic control, a weighted average of the outcomes across the pool of control states, where the weights are selected to match the treated state with its synthetic control in the pre-treatment period, based on the outcome and predictors of the outcome. We only use synthetic controls for the price and charge outcomes and use the Herfindahl–Hirschman Index (HHI) of the state's insurance market and the state's health expenditures per capita as predictors of price.Footnote ¹⁹ To increase the likelihood of a close pre-treatment match between the treated state and its synthetic control, we include all states that did not have an elective surprise billing law during 2012–2019 in the pool of potential synthetic controls, including the states that passed a law in 2020–21.Footnote ²⁰

For inference, we follow the recommendation of Abadie (Reference Abadie2021) and create the permutation distribution of placebo effects, i.e. the distribution of effects if each state in the control pool is considered a treated state. P values are calculated using the distribution of placebo ratios of post-to-pre root mean squared errors (RMSE), where the RMSE captures the difference between the ‘treated’ state and its synthetic control. The p value is the proportion of placebo ratios of post-to-pre RMSE that exceed the post-to-pre RMSE ratio of the actual treated state. In this way, the p value captures both the size of the treatment effect relative to the distribution of placebo effects and the precision of the pre-treatment match of the treated state and its synthetic control. Following the recommendation of Abadie (Reference Abadie2021), we also report (in Appendix Tables A7–A10) the range of estimates when leaving each control state out of the pool of potential controls. As a robustness check on the state-level results, we also estimate the standard TWFE model (using the control group for equation (1)) and these results are reported in Appendix Tables A11–A14.Footnote ²¹ The price and charge-dependent variables of equation (2) are logged and we report the exponentiated coefficient minus one.

By prohibiting physicians from balance billing patients and regulating OON physician reimbursement, surprise billing laws may change physicians' incentives to join managed care networks or remain OON. If a surprise billing law changes the network participation of hospital-based physicians, mean in-network or OON reimbursements may change even if the underlying reimbursements have not changed. For example, if a law made it less lucrative for a physician to be OON (e.g. because the physician could no longer balance bill patients), physicians with smaller OON reimbursements may sign contracts and join networks after the law. In this case, the mean OON reimbursement would increase even if all existing OON reimbursement levels stayed the same because of the elimination of low-reimbursement OON physicians from the OON reimbursement distribution. We address this composition issue by analysing network participation (i.e. the proportion of claims that are in-network) and combined prices and charges, regardless of whether the claim is in-network or OON.

For network participation, the outcome is an indicator for whether the claim was in-network. We only estimate average effects for each state, and do so using the standard TWFE model:

(3)$$Y_{\,jkt} = \alpha _j + \lambda _t + \mu D_{\,jt}^{l \ge 0} + \epsilon _{\,jkt}, \;$$

where Y_jkt is the in-network indicator for claim k in quarter t, α _j is a state-specific effect, λ _t is a quarter-specific effect, $D_{jt}^{l \ge 0}$ is an indicator that equals one for all providers in the treated state and all quarters after treatment, and $\epsilon _{jkt}$ is the idiosyncratic error.

For potential surprise billing episodes, in addition to the Sun and Abraham (Reference Sun and Abraham2021) method described above, we also estimate the effect of each state law with the following model:

(4)$$Y_{ijkt} = \alpha _j + \lambda _t + \rho _k + \mu D_{\,jt}^{l \ge 0} + \epsilon _{ijkt}, \;$$

where Y_ijkt is the potential surprise billing indicator for episode i, α _j is a state-specific effect, λ _t is a quarter-specific effect, ρ _k is a plan-specific effect (e.g. fixed effect for PPO, HMO, EPO, etc.), $D_{jt}^{l \ge 0}$ is an indicator that equals one for all episodes in the treated state and all quarters after treatment, and $\epsilon _{ijkt}$ is the idiosyncratic error. Like equation (1), the control group used in the estimation of equations (3) and (4) is the group of states that passed an elective surprise billing law in 2020 or 2021. For both network participation and potential surprise billing episodes, we calculate p values using the randomisation inference procedure described in MacKinnon and Webb (Reference MacKinnon and Webb2020), in which placebo effects are estimated for each state in the control group (as if it were the treated state) and the p value is calculated as the proportion of placebo effects that exceed the treated state's effect in absolute value.

Although the data provide detailed information about each patient, provider and treatment episode, some information is not included. For instance, we cannot observe and control for the employment of hospital-based physicians by physician staffing companies. In addition, each claim includes only the insurer's allowed amount (i.e. the insurer's payment plus the patient's plan-based cost-sharing). We cannot observe the claim's specific charge, whether the OON physician sent a balance bill to the patient or, if so, how much was paid by the patient to the OON provider. Although we cannot observe the claim's specific charge, we can observe the provider's average charge for Medicare patients for each CPT and year and we use this as a proxy for the provider's average charge to privately insured patients. The same approach has been used in prior papers that analysed charges for privately insured patients (Adler et al., Reference Adler, Lee, Hannick and Duffy2019b).