Some follow-up technical notes on State-based exchanges

When I reposted the last post as a DailyKos diary, a commentator engaged me a bit on a few ideas to improve the basic concept. He suggested for controlling for partisan control of the state government as a binary variable (cooperative == full Democratic control=1; non-cooperative==full Republican control=0). I thought a tertiary variable (0=GOP 1=split 2=Dem) might be a bit better.

Having an independent variable for partisan control might soak up some of the variation for cooperation with implementing the ACA, but the problem is that any partisan control variable would be extremely highly correlated with State-based exchanges. There were very few complete Democratic states without at least a shared exchange (West Virginia and Illinois jump to mind), and only one GOP controlled state (Idaho) with an exchange. I'm not sure how useful it would be in practice at sorting out variation not associated with exchanges since the two are so strongly correlated.

In any case, it was a thoughtful piece of feedback and definitely worth the brief conversation we had.

Another idea I thought of was to take quarterly data for each state to increase the number of observations for each state from one to six to track changes from the end of 2013 through the second quarter of 2015. Taking the data from cross sectional data to cross-sectional-time series in this manner would create 300 observations (vs. 50) and increase leverage dramatically, while allowing us to track over-time change in states setting up and taking down exchanges and or expanding Medicaid at different points.  Of course, as my old methods prof John Jackson used to stay "No good deed goes unpunished" and we'd have to control for the serial correlation in the states with either a fixed or random-effects model.  At least we wouldn't have to worry about panel-corrected standard errors, seeing that we're dealing with the universe of states, not a subsample.  And there would also be the problem of increase error within states for each quarter, since the Gallup sample would slip to alarmingly low levels for some states, increasing the margin of error around the uninsured rate for a given quarter.

And that's assuming that I could even get the more specific data out of Gallup.

I don't really have the time to try out either of these ideas right at the moment, but anyone in Internet land can feel free to try and report back. I'm rather interested.