A Bayesian Hierarchical Model to Estimate State-Level Support for Health Care Reform from National Opinion Data
*Richard Gonzales, Harvard University 

Keywords: Health care reform, public opinion, hierarchical models

Public opinion plays a prominent role in discussions about major policy issues, including the debate over health care reform in the US. However, sample sizes in public opinion polls are typically too small to directly estimate support for a policy at the state level. Given the importance of understanding local opinion dynamics for studying Congressional behavior on health policy, this study seeks to extend small area estimation techniques to national public opinion data to obtain state-level estimates of support for health care reform.

I examined a series of five national public opinion surveys conducted from July 2009 to March 2010. In approximately half of the states, 10 or fewer respondents in each poll expressed an opinion on health care reform. To estimate support for reform given such sparse data, I fit a hierarchical logistic regression model using age, sex, race, and education variables as individual-level predictors, and region indicators and previous presidential vote-share as state-level predictors. The model therefore incorporates both geographic and demographic information, which has been shown to perform better than demographic information alone. Weakly informative priors were assigned to all parameters to estimate the model in a fully Bayesian context. To obtain the final state-level estimates of public support for health care reform, simulations from the posterior parameter estimates were fit to each of 64 possible demographic categories in each state. The predicted response for each demographic group was then weighted to the size of that group according to 2000 US Census Bureau data. The weighted estimates were used to compute the mean level of support in each state along with 95% credible intervals.

Preliminary results suggest that while most Americans opposed passing the Affordable Care Act, a majority in several states supported reform throughout the entire debate. This geographic variation frequently gets masked when commentators focus on national averages when reporting on public opinion. The variation in support also points to challenges for state implementation of the ACA.

This analysis shows the usefulness of combining hierarchical modeling methods with auxiliary poststratification information to obtain state-level estimates of public opinion. The method can be applied to almost any national opinion survey and is easily implemented using existing software and data to better understand the role of public opinion in health policy.