Industrial Reorganization: Learning about Patient Substitution Patterns from Natural Experiments

Since the seminal work of McFadden (1981), discrete choice demand models have become one of the main tools used to model consumer preferences. For example, economists use these models to assess the substitutability of merging hospitals. However, the validity of these models has been questioned since quasi-experimental variation in choice sets is rare. Indeed, Angrist and Pischke (2010) have suggested that the field of industrial organization's heavy reliance on unvalidated models is so problematic that it should be rebranded as industrial disorganization.

We assess the performance of discrete choice models in the context of hospital demand. Discrete choice models of hospital demand have been used to answer a number of policy questions, including conducting hospital merger retrospectives, examining the health impacts of physician incentives, and understanding provider quality. Further, antitrust authorities regularly use these models to guide merger policy. Second, several skeptics have challenged the quality of hospital demand model predictions in real-world situations, arguing that in some cases the models' ex ante predictions appear inconsistent with how events turned out.

We assess discrete choice models using four cases in which hospitals were severely damaged or destroyed by a natural disaster, but the majority of surrounding areas were not. These disasters induced an exogenous change in the set of hospitals that individuals would be able to choose from for medical care.  Such exogenous change gives us the opportunity to assess the performance of discrete choice models using the kind of quasi-experimental variation that has been missing from the literature. These incidents occurred in four very different areas, including rural, suburban, and urban environments. This gives us more confidence that our findings are relevant beyond these particular settings.

Using this quasi-experimental variation in choice sets, we estimate demand models on pre-disaster data and test their performance in predicting patient substitution patterns afterwards. We focus on model specifications used in the recent academic literature.

Across all our analyses, we confirm the importance of flexibly accounting for patient heterogeneity in preferences for unobserved hospital quality.  In general, we find that the model that best predicts individual choices uses a semi-parametric approach that allows for flexible substitution patterns across small groups of patients.  In contrast, models that focus solely on observable hospital characteristics perform worse than models that allow for preference heterogeneity across product specific dummy variables.

We then examine whether the choice of model matters for policy counterfactuals. We compute the predicted percent change in the willingness to pay from a series of simulated hospital mergers. The standard deviation of this statistic across models is approximately 15% of the average for mergers that may receive greater antitrust scrutiny. These are differences that could be sufficient in magnitude to impact policy in borderline cases. Further, we find that, on average, consumer harm is higher when using better fitting models for such mergers. Therefore, to the extent that errors are made as the result of using worse fitting models, they will be in the direction of more lax antitrust enforcement.