Three months ago I mashed together a little data set in order to examine variance in state level public support for the ACA, and wrote about it here. This thing has magically morphed into a conference paper, so I’ve been kicking it around when time allows. As I present this as of yet unwritten gem in a little over a month, it’s time to lock the model in and move on to the writing bit.
I’ll shamelessly quote from July’s post on variable description:
The two key variables in the data set are state level support for the ACA and the percentage uninsured in the state. My source for the former is from a paper written by Richard Gonzales, a doctoral candidate at Harvard’s Department of Health Care Policy, which was discussed here at The Incidental Economist. Gonzales estimates state level support from national Gallup data over a six month period (September 2009 to March 2010). As this is an estimate, it does introduce an additional layer of uncertainty into the model, but it’s the best data I could find. The point estimates and error bands appear sound in terms of face validity; the highest levels of support are found in New York, Hawaii, and Vermont, with the lowest in Oklahoma and Wyoming. The public opinion data are old, but Monkey Cage suggests (and we all pretty much agree on) public opinion on the ACA has been relatively stable. Percentage uninsured by state level is courtesy of Gallup, which are available here.
A bi-variate relationship was the exact opposite of what we would expect: as percentage uninsured increased, support for the ACA declined. Believing that wasn’t an accurate story, I added percentage point change in uninsured from 2008-10, hypothesizing that a greater increase in uninsured is associated with an increase in support. I also added both the PVI of the state and the percentage voting for Obama in 2008. As the latter two are highly correlated (.98), I ran discrete models with each in turn. Discussion on these variables can be found in July’s post linked at the top.
Since then, I sourced and included a bunch of variables that I had hoped would be revelatory, including per-capita GDP, average unemployment from 2008-10, the rate of Medicaid recipients expressed as a percentage, the poverty rate (again, %), and the percentages black and Latino. Furthermore, on September 17 I added the ‘average poll rating’ from Silver’s 538 at the NYT. Again, that is highly correlated with both the PVI and the percentage Obama vote in 2008, requiring three separate models.
Some brief explanation first: Partisan X is the respective partisan lean variable for each of the three models in turn (1. 538 polling average; 2. PVI; 3. Obama % in 2008). Significance levels are sparsely needed; for the uninitiated, three stars represents a significance level < .001, + represents a significance level at or below .10. Effectively what this means in these cases is that the odds of finding that estimate by random chance (as opposed to observing a real relationship) is less than one in a thousand (***), or one in ten (+). One might argue that as I have the universe of data at hand, as these are not probability samples, significance levels are largely meaningless. I’d argue this very point if it didn’t go off on some irrelevant tangent, so we’ll stick with convention for the time being.
Apologies for the table formatting; my staff of interns have the day off today.
|M1 (538 poll avg)||M2 (PVI)||M3 (Dem % 2008)|
|GDP per capita||-0.02||-0.02||-0.001|
|Average uninsured 08-10||-0.28+||-0.14||-0.18|
|%change uninsured 08-10||-0.29||-0.19||-0.38|
|Average unemp. 08-10||0.41+||.17||.34+|
|Medicaid Rate (%)||-1.59||-2.77||2.53|
|Poverty Rate (%)||.04||-0.02||.14|
So, what do we have here? The model fit is strong across the board: between 89 and 93 percent of the variance in public opinion support for the ACA is explained. The model including the % vote for Obama in 2008 is the strongest, but it’s such a marginal difference I don’t want to read too much into this.
Only two variables are consistently significant: the partisan measure, and percentage black. Before discussing the others, a further comment needs to be made regarding conventional significance levels. The industry standard, however arbitrary, is .05; meaning, if there’s a 5% or less chance that the observed relationship is a product of random chance, then we accept the relationship as being significant. It is permissible to stretch this to .10 but only in cases where we have an a priori theoretical reason to not only expect a relationship, but also to expect the direction of the relationship. While I have average uninsured in bold and with a significance marker, I have a difficult time imagining a theoretical expectation that would predict the direction of this estimate: higher levels of uninsured people being associated with lower levels of support for the ACA, so I’m going to overlook this one for now (along with overlooking the bivariate figure I posted in July). As an aside, it is possible that predominantly red states have atypically huge pockets of poverty and uninsured citizens, but the model should account for this through the other variables. So what we have remaining is higher levels of unemployment being related with higher levels of support for the ACA, as well as a relatively stable estimate for Latino support, yet only significant in one model, both of which make sense.
I tried several methods of isolating the swing states (of which I included NV, NM, CO, OH, WI, IA, NH, VA, NC and FL). I ran models with those states only, which was a universal dud, but then an N of nine isn’t going to afford any real statistical power. I also ran interactive terms in the model isolating whether or not these states as a group had a unique estimate for uninsured %, unemployment, Medicaid and poverty rate, and generated nothing.
What lessons can we take away from this little exercise? I have several ideas. Politically, it’s evident why health care has not been a major issue in the campaign. First, obviously, Romney would find it difficult to run against something that he supported in Massachusetts (but, of course, this is Mitt Romney we’re discussing. He can be anything to anybody at any time). Second, less obvious, it’s not going to make a difference. Romney states are largely predisposed to oppose, while Obama states are in favor. The ten swing states identified above have a much tighter distribution (45%-53%) than the nation writ large (32%-63%), but adopting a strident anti-ACA strategy has risks. Iowa and Florida show 45% support in this dataset, so these would be good targets to attack the ACA, but you risk losing support in Nevada and New Mexico (both 53%), and the other six (all at 49% or 50% in favor). It appears to me that there’s more to lose than there is to gain by going after the ACA in the swing states.
Academically, this paper might have something to add to the literature, which is pretty barren on public opinion and health care reform. Tesler (2011) finds that race is a factor in determining support; given that President Obama is regarded as an African American, support for the ACA is mediated by racial attitudes amongst the white population. Relying on data no more current than 2004, Gelman, Lee, and Ghitza (2010) find opposition to be concentrated among those with higher incomes, and those over the age of 65.
That aside, I’m more convinced now than I was in July that judgment of the ACA as policy is mediated by both a pre-existing partisan lens as well as by race. Obama was never going to win Republicans over to it, regardless of how hard he tried and how much he gave away at the beginning, because they were never going to support it anyway. Electorally, it can be used to rally the base, but that’s it.
Of course, one must also suspend any concern that they might have with the ecological fallacy to make some of the inferences I make above . . .