Data on Healthcare Expenditures
We analyzed all preferred provider organization insurance claims processed for care through BCBSTX for 2014 through 2016 in the four largest Texas metropolitan statistical areas (MSAs; Austin, Dallas, Houston, and San Antonio). The population in these MSAs totaled 18.9 million in 2017,14 which is greater than the entire population of 46 other US states.15 The sample includes all claims for healthcare services, except for prescription of drugs.
For each patient, we summed all claims in each calendar year to calculate the annual allowed amount for each patient, which represents payments to providers by BCBSTX, as well as out-of-pocket expenses (deductibles and copayments) that the patient is responsible for. We refer to the allowed amount on each claim as the price for that particular service. Physicians influence patient spending within their practice, as well as what hospitals or other facilities that patients may be referred to outside of their practice.16 Thus, annual spending for each patient represents all services used by patients, not just the services directly provided by the attributed physician organization. Patients with > $100,000 in costs in a calendar year were excluded from the sample, in order to exclude the effect of small numbers of very sick patients on average expenditures per patient.11, 17 A flow diagram illustrating the change in sample size resulting from each exclusion is in Appendix Figure 1.18
In addition to comparing overall spending by physician ownership type, we decomposed spending by clinical category, as defined by the Berenson-Eggers Type of Service (BETOS) classification system, and by site and type of care. The BETOS coding system was developed by the Center for Medicare and Medicaid Services to analyze growth in Medicare expenditures, grouped by readily understood clinical categories. BETOS codes are assigned to each Health Care Procedure Coding System (HCPCS) code.17, 19
Attributing Patients to Physician Organizations
We limited the analysis to adults ages 19 to 64 who were continuously enrolled in at least one of the calendar years 2014 to 2016. Ideally, we would have attributed each patient to a primary care physician (PCP) in each calendar year based upon the PCP who accounted for the majority of their visits in that same year. But if we followed this approach, patients with no visits to a PCP in a calendar year could not be attributed to a physician. Therefore, we searched for claims for visits to a PCP in a 24-month window that included the calendar year and additional months closest to that year. Additional details of our attribution method are provided in Appendix eMaterial 1.18 Primary care physicians were those doctors who listed their specialty as family practice, general practice, geriatrics, internal medicine, or pediatrics.
Quality Measures
Among hospitalized patients, we identified those readmitted within 30 days of discharge. We also constructed four process measures adapted from the Healthcare Effectiveness Data and Information Set (HEDIS; National Committee for Quality Assurance); for patients with diabetes, hemoglobin A1c and LDL cholesterol screening, and diabetic retinal examination; screening mammography for women ages 50–64 years.
Defining Physician Organizations
We determined the ownership status of physicians based on their recorded network for reimbursement in the BCBSTX internal database. Each claim contains the tax identification number (TIN) of the treating physician or hospital. The Network Management group at BCBSTX maintains records of the contracts negotiated with physician and hospital organizations, which also contain information on the identities of physicians included in each contract. These groupings were used to assign each TIN to a particular physician or hospital organization.
We compared the annual healthcare spending for patients attributed to a physician treated by local hospital-owned physician groups versus multi-hospital system–owned physician groups, versus physician-owned organizations. Following past research, local-owned hospital organizations are single hospitals or hospital chains which do not extend across geographic regions. Multi-hospital system–owned physician groups are affiliated with a hospital system that extends across geographic regions.11
Control Variables
The annual expenditures for each patient were adjusted for the relative disease burden of each patient, differences in the cost of delivering care across regions of Texas, the annual number of patients attributed to each physician organization, whether the patient was in a consumer-directed health plan (CDHP), and unobserved but systemic differences in spending across the four MSAs by year.
A CDHP combines a high deductible health plan (HDHP) as defined by the Internal Revenue Service (IRS) with a health savings account or a health reimbursement account. The minimum annual deductible for an HDHP was $1300 for self-only coverage and $2600 for family coverage.20 Patients with CDHPs may be less likely to use healthcare services than patients without these plans.
Relative disease burden was accounted for using patient age categories, gender, and a concurrent risk score. BCBSTX calculates the concurrent risk score based on claims in the calendar year of treatment using the Verscend Technologies DxCG concurrent risk score. The Verscend DxCG is a risk scoring model that performs as well as, or better than, other available risk models in predicting the current year’s healthcare expenditures based on diagnoses reported in the claims.21
Differences in the cost of care across regions of Texas were accounted for by including as a control variable the wage index that is used to construct the geographic adjustment factor (GAF).22 The GAF is used by the Center for Medicare and Medicaid Services to adjust fee-for-service reimbursement rates for regional differences in input prices and is updated yearly. The wage index reflects the average hourly hospital wage in each local market divided by the national average hourly hospital wage.23
Statistical Analyses
Annual payments by physician organization type were first compared using descriptive statistics. We then estimated regression-adjusted analyses, distinguishing both local hospital-owned and multi-hospital-owned physician groups from physician-owned practices. Given that the estimated spending differences between the two hospital-owned practice types were not statistically significantly different from each other, we report regression-adjusted spending differences for physician-owned practices versus all hospital-owned organizations combined.
We estimated spending regressions using all expenditures as the dependent variable, as well as spending by BETOS category and site and type of care. The regressions were estimated using SAS Enterprise Guide, version 7.1. In these regressions, the explanatory variable of interest is an indicator for whether the patient was treated by a hospital-owned or physician-owned practice. We then decomposed the spending change into a price effect and a utilization effect by replacing the price on each claim with the median price in its respective BETOS category. Spending regressions estimated using the annual sums of standardized claims reflect only differences in utilization.17, 24 To further understand the source of spending differences by organization type, we compared the percent of patients with positive (versus no) annual spending in each BETOS category and site and type of care for patients treated by physician-owned versus hospital-owned practices.
The number of patients attributed to each physician group was specified in four categories: practices with up to 500 patients; 501 to 10,000 patients; 10,001 to 15,000 patients; and 15,001+ patients per year. These categories were chosen to obtain a broad separation of practice sizes within our sample, but also reflect information from previous studies on physician practice panel size and the distribution of physicians by practice size.3, 25
The regressions include year fixed effects to measure trends in spending over time, as well as the control variables outlined above. Patient age was included using categorical indicator variables for ages 19 to 29, 30 to 39, 40 to 54, and 55 to 64. The DxCG concurrent risk score was specified using categorical indicators for deciles of patient values in our sample, to allow for flexibility in how patient risk influences spending as patient illness severity increased. The wage index and its squared value were also included.
Interactions of indicator variables for the four MSAs multiplied by year indicator variables control for unobservable systematic factors (e.g., regional practice patterns) that may be correlated with both vertical integration and spending. With the inclusion of these variables, the regression results yield estimates of spending differences associated with practice ownership-type within MSAs and year, unconfounded by unmeasured differences across MSAs and over time.
Annual spending in dollars for each patient was converted to logarithmic units in order to obtain regression estimates expressed in percentage differences associated with each covariate. The regressions were estimated using a quasi-likelihood generalized linear model with a log link. For cases where a patient has $0 spending in a BETOS category, this approach includes these observations, whereas they would be excluded from an ordinary least squares regression with the natural log of spending as the dependent variable.26
Following recent studies, we also estimated multivariate regressions with spending expressed in dollar units rather than logarithmic units.11, 24, 27 These studies state that in large samples, linear models generally outperform other statistical models at estimating population averages, even though they can be less precise at estimating the tails of a spending distribution.26, 28,29,30 The standard errors for all regression estimates were clustered by physician organization across years to account for heteroscedasticity.31, 32 An estimate is considered to be statistically significant if a two-sided t test yields a p value ≤ 0.05.
Given that previous studies observe that physicians may align with hospitals in order to bill more services in the higher reimbursement outpatient setting, we examine cases where clinicians might exercise this behavior. For BETOS categories where we find a significant difference in spending by physician practice ownership type, we searched for current procedure terminology (CPT) codes (in professional claims) that could be cross-walked to revenue codes (in outpatient claims). We then compare the number of claims per enrollee between patients treated by hospital-owned versus physician-owned practices, and we calculate the percent of claims billed in the outpatient setting.
The research protocol was determined to be exempt from review by Rice University’s Institutional Review Board, because all physician and patient identifiers were removed from the dataset made available to us for analysis.