Trump, the ACA, & the uninsured: A county-level analysis.

The Trumpiest counties are also the neediest of insurance.

During the election, I heard they discussed issues…

Not particularly often, but maybe for 5–6 minutes during the third debate? One of those, obviously, was the Affordable Care Act, more popularly known as OBAMACARE. And it’s an important issue — a manner of socialized healthcare access, which many see as a first, albeit very incomplete step to getting a measure of equitable healthcare access in this country. Some people’s health insurance premiums increased substantially from the previous year, which provided some good political fodder during the run-up to the election; on the other hand, about 20 million extra people have some form of healthcare coverage, & it’s hard to argue that that’s a bad thing. Nevertheless…

To Repeal & Replace?

Well, we’re all probably familiar with recent history. Republicans have made the repeal of Obamacare their top priority… well, parts of it anyway. Or maybe all of it? And they’ll replace it with… I think they’re still thinking about that.

Well, obviously this hasn’t gone down very well with many factions of the voting populous, including ACA enrollees. One of the primary arguments for the GOP to not scrap the law, is that that their constituents are often beneficiaries of the ACA, either through enrollment on the Marketplace or through Medicaid Expansion. In conjunction with this, there were two articles that caught my eye from the Atlantic and Vox, which dive into some detail on unpacking the circumstances and mindsets of voters and geographies that also happen to rely on the ACA and its provisions. So in light of the more anecdotal evidence, I thought it would be good to see some more aggregate data.

The Trumpiest is also the…

Step 1! I downloaded county-level ACA enrollment data for 2015 & 2016 from You can download the data as Excel spreadsheets or be fancy & use the API. Either way is easy. And then I married that up to election results (very thoughtfully compiled by Michael Kearney). And when you graph those two things side by side…

The red line is the number of plans (29, to be precise) per 1,000 population across the country.

Two things popped out to me: 1) there isn’t a negative relationship between votes for Trump and the number of plan selections (it generally looks flat); & 2) the counties with the highest number of ACA plan selections are also those that went strongest for Trump. Honestly, I expected to see a generally flat trend on this graph, which is the case for many of the counties. But the outliers on the right tail are a bit surprising.

But this pattern isn’t just for this year — the year over year increase in enrollments is even more pronounced.

Nationally, there was a 5% increase in total number of plan selections between 2015 & 2016.

So as counties go Trump, there tends to be more of an increase in the number of plan selections per year. It’s important to note that Trump took the cake on counties with sparser populations (i.e. more rural counties), which necessitates a consideration of regression to the mean — that less populous counties are less likely to increase their enrollments simply because of having fewer people. The fact that there is an upward trend in this graph at all is honestly somewhat surprising. This alone should give us pause to consider what the GOP has in store for the Replace part of this equation.

But what about the uninsured?

The ACA isn’t the full story, however. There is also the uninsured population to consider. The graph below shows the relationship between the adult uninsured population against the vote for Trump — and there’s a clear upward pattern starting around 60%.

Insurance status trumps demographics

However, all I have shown so far are a few bivariate relationships, and of course voters make decisions on multiple, complex & intertwined issues. And particularly in this election, we saw some stark social, class, & racial divisions brought to the national forefront. I wanted to see what kinds of other factors correlate with Trump’s rate of success. I pulled in some basic information about counties — unemployment data, age, gender, race, and county population — and used those, along with ACA enrollments, as covariates to see how the uninsured rate affected Trump’s vote. Indeed, I was surprised:

A few notes: I used quasibinomial regression with the proportion of vote for Trump as the response variable. Diagnostics such as variance inflation & residual plots generally “looked good”. However, I’m sure other things could be done better (including other types of data, interactions, etc). Also, I’m assuming that that all observations are independent — but clearly there is spatial autocorrelation between counties, which I didn’t treat here. I welcome all critiques & suggestions.

The uninsured rate is actually a pretty good predictor of a county’s vote for Trump. There’s one caveat, which is that the insurance estimates, which I took from the US Census’ SAHIE program, is from 2014. So the strength of association could potentially be weaker if the uninsured rate has decreased since then. For instance, say there was a healthcare law that expanded Medicaid coverage…

Another thing to point out —I’m only looking at demographic features here. This excludes other issues of electability, like candidates’ platforms (rofl!— that was a joke), the party of the previous administration, etc. As an aside, American University Professor Alan Lichtman has accurately predicted the outcome of nearly every General Election win in US history, including Trump’s, using a “fundamentals” model. I like it, & at a later date I’d like to incorporate some features from his model into this work. You can read a summary of his “Keys” model here.

Trumpian constructs

But going back to the idea that the factors shaping decisions are complex and intertwined — I also used Principal Components Analysis to find underlying constructs of counties and their voting patterns. PCA takes the variance in a series of variables and presents them such that each component is uncorrelated/independent of other components. Presented here are the first three components.

Now, I’m not a sociologist & this is a fairly basic set of inputs, but I’ll make a few liberal interpretations based on what I found. Again, I’ll point out here that I’m looking at mostly demographic factors, and that what we see here is probably more directional than deterministic.

Dimension 1: The Trumpian Dimension —The factors driving this dimension? The white population of a county, the vote for Trump, and the median age. On the other hand, a county’s minority (non-white) population and the unemployment rate were high negative contributors to the dimension. The uninsured rate & ACA enrollments were negative contributors, but not to a particularly great extent.

Dimension 2: Gender differences— ’nuff said.

Dimension 3: Prosperous & Populous — This dimension is defined by counties with higher populations & greater median incomes. Perhaps not surprisingly, the uninsured rate is negatively associated with the other two items on this factor.

Conclusion —this is actually isn’t the full story

The pattern that most stuck out to me in the few graphs above, was the relationship between public insurance enrollment, uninsured population & the highest proportion of Republican votes. I thought there would be a stronger relationship between Marketplace enrollments and Trump’s vote. In other words, I suspected there to be a substantial level of cognitive dissonance among Republican voters around the ACA, and this analysis disproved that idea, at least to an extent.

On the other hand, the uninsured population is a decent predictor of the vote for Trump. The results of the regression, along with the right side of those scatterplots, leads me to believe that many Trump supporters also need a concrete and coherent replacement for the ACA.

Finally, this analysis isn’t the whole story, because I didn’t look at another important part of the ACA: Medicaid Expansion. This is another subject that is worth a data-driven examination, because there have been numerous anecdotal reports of individuals hating “Obamacare” but signing up for Medicaid under the Expansion. Medicaid data are currently difficult to get at the national level, although there is a plan to implement a platform sometime in 2017. And when those data do become available, I’ll report back on my findings.

Note: all the data grabbing (of which there was much) & assumptions made for statistical tests (of which there were many) can be found on my github.