the benefits to the ACA, the population of Americans using the exchanges was relatively small.
As the percentage of an overall population which uses a given policy declines, that policy becomes
“distant” to a large fraction of the electorate, meaning that relatively few people have direct
experience with it (Soss and Schram, 2007). That, in turn, makes citizens more reliant on media
portrayals and partisan cues in generating attitudes (Jacobs and Mettler, 2018).
Moreover, the ACA’s exchanges were one element in a complex package of reforms. Campbell
(2003), Soss and Schram (2007), and Campbell (2012) note that policy feedbacks are more likely
when the policy’s beneficiaries are concentrated in ways which encourage them to identify as a
coherent group and act on their shared interests. But the ACA’s complex, multi-faceted design
means that key beneficiaries have a stake only in specific, often disparate provisions (see also
B´eland, Rocco and Waddan, 2018; Chattopadhyay, 2018). This complexity may fragment the
beneficiary population, reducing its capacity to generate a cohesive identity.
Research Design and Data Sets
Prior research leads us to expect substantial selection bias, as those who anticipate higher health
care bills or are not Republican are more likely to enroll via the exchanges (Lerman, Sadin and
Trachtman, 2017). As a result, any straightforward comparison of people who did or did not use
the exchanges is almost certainly biased.
How, then, to evaluate the exchanges’ impact? Descriptive analyses serve as a useful start-
ing point, since we should expect causal effects to generate associations at a minimum. Cross-
sectionally, it is valuable to know whether those who use the exchanges or receive subsidies think
differently about the ACA than those without insurance or with insurance from elsewhere. Know-
ing which groups of Americans are more likely to use the exchanges also helps identify promising
discontinuities or other strategies for causal inference. In addition, our descriptive analyses allow us
to consider temporal variation: did attitudes toward the ACA among different sub-groups shift at
the time of its implementation? In places, we are able to use panel data to track within-respondent
changes before and after the ACA’s implementation.
Armed with a knowledge of various statistical associations, we then proceed to estimates that
are credibly causally identified. Instead of providing a single, decisive test, we employ multiple
tests with different samples, estimands, outcomes, strengths, and weaknesses. Below, we discuss
7
Electronic copy available at: https://ssrn.com/abstract=3366994