Can Diffuse Interventions Improve Population Health? A Study of the State Innovation Models Initiative
Can Diffuse Interventions Improve Population Health? A Study of the State Innovation Models Initiative
Wednesday, June 26, 2019: 1:00 PM
McKinley - Mezzanine Level (Marriott Wardman Park Hotel)
Discussant: Muzhe Yang
A major component of the ACA provides mechanisms and financing for innovation in health care delivery and payment mechanism to improve patient care, population health and reduce health care costs. In 2013, six states – Arkansas, Maine, Massachusetts, Minnesota, Oregon, and Vermont — were awarded a total of $250 million to implement and test comprehensive health system innovation plans over a period of 3.5 years. These states were chosen, among applicants, in part because they relied on already-existing CMS initiatives across public and commercial payers.
In this paper, we examine the effects of the SIM initiative on health outcomes in the six states that implemented their innovation plans as compared to 15 states that were not involved at all. We use data from the Behavioral Risk Factors Surveillance System (BRFSS) for the years 2010 - 2012 and 2014 - 2016 to examine the effects of treatment. We do not include data from 2013 because that was the year in which states learned of their awards and began implementation. We use a generalized interrupted time series approach instead of a difference-in-difference model to study the effect of the SIM initiative. These statistical designs are very similar but the interrupted time series approach is agnostic about trends in the pre-program or baseline period. Given the numerous changes to the health care landscape induced by the ACA, we prefer not to assume, a priori that the health trends in these six treated states was the same as those in the control states.
Health status is a notoriously difficult construct to measure in part because it is multidimensional and latent. Instead of estimating models for individual measures, we develop an econometric model that uses three measures of health status simultaneously, assuming that the regression model is for an underlying common latent measure of health status. The measures include an ordinal measure of health status, which we assumes follows an ordered logistic distribution, a count of days affected by physical health problems, and a count of days affected by mental health problems, which we assume follow Poisson distributions. These outcomes are jointly modeled as being driven by a common discrete latent factor, thus making our model similar to a Latent Class Model. We then assume that a set of observed covariates, including the treatment design
covariates and a set of demographic and socioeconomic characteristics, affect the latent factor directly.
The results of our analysis are remarkably consistent. Individuals in states that implemented SIM saw significant improvements in health status in each of the samples. Individuals ages 45 and older in SIM states saw 1.1 - 1.8 percentage point improvements in their (latent) health status. When the focus is on the lower income subsample, the effect sizes increase to 1.2 - 2.3 percentage points. The effect sizes are even larger, 2.6 - 3.6 percentage points in the older, Medicare sample. There is no evidence of differential pre-intervention trends.
In this paper, we examine the effects of the SIM initiative on health outcomes in the six states that implemented their innovation plans as compared to 15 states that were not involved at all. We use data from the Behavioral Risk Factors Surveillance System (BRFSS) for the years 2010 - 2012 and 2014 - 2016 to examine the effects of treatment. We do not include data from 2013 because that was the year in which states learned of their awards and began implementation. We use a generalized interrupted time series approach instead of a difference-in-difference model to study the effect of the SIM initiative. These statistical designs are very similar but the interrupted time series approach is agnostic about trends in the pre-program or baseline period. Given the numerous changes to the health care landscape induced by the ACA, we prefer not to assume, a priori that the health trends in these six treated states was the same as those in the control states.
Health status is a notoriously difficult construct to measure in part because it is multidimensional and latent. Instead of estimating models for individual measures, we develop an econometric model that uses three measures of health status simultaneously, assuming that the regression model is for an underlying common latent measure of health status. The measures include an ordinal measure of health status, which we assumes follows an ordered logistic distribution, a count of days affected by physical health problems, and a count of days affected by mental health problems, which we assume follow Poisson distributions. These outcomes are jointly modeled as being driven by a common discrete latent factor, thus making our model similar to a Latent Class Model. We then assume that a set of observed covariates, including the treatment design
covariates and a set of demographic and socioeconomic characteristics, affect the latent factor directly.
The results of our analysis are remarkably consistent. Individuals in states that implemented SIM saw significant improvements in health status in each of the samples. Individuals ages 45 and older in SIM states saw 1.1 - 1.8 percentage point improvements in their (latent) health status. When the focus is on the lower income subsample, the effect sizes increase to 1.2 - 2.3 percentage points. The effect sizes are even larger, 2.6 - 3.6 percentage points in the older, Medicare sample. There is no evidence of differential pre-intervention trends.
