Forecasting Health Care Spending: A Comparison of Nonlinear Econometric and Machine Learning Methods
Forecasting Health Care Spending: A Comparison of Nonlinear Econometric and Machine Learning Methods
Monday, June 24, 2019: 10:30 AM
Madison A (Marriott Wardman Park Hotel)
Discussant: Naomi B. Zewde
Health care spending in the USA is exceptionally concentrated, with the top 5 percent of spenders accounting for nearly 60 percent of costs. But high spending is not necessarily persistent over time. Aldridge and Kelley, in two recent papers, show that, of these highest-spending 5%, 44% have transiently high costs while 40% have persistently high costs. Other studies have demonstrated that as few as one-third of those with high spending continue to have high spending one year later.
It is critical to predict the population with persistently high costs so that appropriate interventions can be implemented to reduce costs and improve quality. For example, interventions to identify people with persistently high spending and co-occurring mental illness has led to programs that successfully targeted behavioral health supports to reduce utilization while improving quality. Yet, there are very few evaluations of statistical models for forecasting health care spending and costs. It is well known in the time series literature that models which are known to perform well to explain outcomes within samples and across cross-sections are not necessarily the best models for forecasting future values of outcomes.
In this paper, we subject a battery of alternative statistical models to performance measures of one step ahead forecasting. We consider a variety of generalized linear models, the log linear regression model. We use finite mixture models and regression trees and other machine algorithms for continuous outcomes to exploit heterogeneity that is difficult to specify a priori. For each of these models, we use data on health care spending and individual characteristics from one year to fit the models and evaluate the forecasts from those models on spending in the second year. We use cross-validation techniques to evaluate in-sample and forecast fit. Currently, we have results using data from the Medical Expenditure Panel Survey. We expect to have results from at least two other datasets, the HRS and from Medicare Claims by the time of the conference.
The results from MEPS show that models which exploit heterogeneity (not a prior specified) provide better forecasts than models that do not. In addition, the use of posterior classification available in the finite mixture models framework provide a substantial boost to forecast power that is not replicated with the other modeling strategies.
It is critical to predict the population with persistently high costs so that appropriate interventions can be implemented to reduce costs and improve quality. For example, interventions to identify people with persistently high spending and co-occurring mental illness has led to programs that successfully targeted behavioral health supports to reduce utilization while improving quality. Yet, there are very few evaluations of statistical models for forecasting health care spending and costs. It is well known in the time series literature that models which are known to perform well to explain outcomes within samples and across cross-sections are not necessarily the best models for forecasting future values of outcomes.
In this paper, we subject a battery of alternative statistical models to performance measures of one step ahead forecasting. We consider a variety of generalized linear models, the log linear regression model. We use finite mixture models and regression trees and other machine algorithms for continuous outcomes to exploit heterogeneity that is difficult to specify a priori. For each of these models, we use data on health care spending and individual characteristics from one year to fit the models and evaluate the forecasts from those models on spending in the second year. We use cross-validation techniques to evaluate in-sample and forecast fit. Currently, we have results using data from the Medical Expenditure Panel Survey. We expect to have results from at least two other datasets, the HRS and from Medicare Claims by the time of the conference.
The results from MEPS show that models which exploit heterogeneity (not a prior specified) provide better forecasts than models that do not. In addition, the use of posterior classification available in the finite mixture models framework provide a substantial boost to forecast power that is not replicated with the other modeling strategies.
