Physician Consolidation and its Effect on Specialist Care: An Application with Machine Learning

Abstract:

A number of studies have documented steady increases in physician practice size. While policymakers hope that larger physician organizations may be in a better position to promote efficiencies, monitor population health and improve care coordination for chronic conditions and quality, consolidation raises concerns around greater physician market power and higher prices. The impact of physician consolidation into larger practices is not well understood, and, in particular, how it differs by type of community, organization, and specialty. This paper provides estimates of how physician consolidation affects the price of specialty medical care. By combing econometrics and machine learning techniques it tests for heterogeneous effects by community and provider characteristics.

We use private insurer claims data from HCCI to construct prices from insurer allowable amounts. We examine physician services in 10 specialties and limit our analysis to their primary locations of service and billing codes. To identify changes in physician organization we use SKA data and data from the American Hospital Association.

To test the impact of physician consolidation on specialty prices we must address three empirical challenges. First, the impact of consolidation almost certainly varies across physician and patient types and, consequently, our approach must allow for heterogeneous treatment effects. Second, the heterogeneity of patient types, their conditions and comorbidities, is very large and, as a result of this high-dimensionality, cannot be adequately addressed with conventional methods. Third, physician consolidation decisions are not random and thus our study design must take into account the possibility of unobserved selection bias.

In our base model we estimate the impact of consolidation using conventional differences-in-differences methods. In conventional modeling of heterogeneity one might rely on quantile regressions or on ad hoc subgroup analysis across a few baseline characteristics. These analyses would require ex ante assumptions about how patient comorbidities and provider characteristics affect prices. Instead we employ use generalized random forests to find the empirically relevant heterogeneity. Such machine learning approaches seek to fit complex and flexible functional forms to the data, while avoiding overfitting.

Our analyses seek to address policy challenges around provider concentration and promotion of integrated care delivery. By uncovering which specialties and patient characteristics are most impacted by provider consolidation, our analyses can focus attention where impacts are greatest.

Funding Sources: Robert Wood Johnson Foundation