Wednesday, November 25, 2020

What Do County Voting Results Tell Us About Demographic Trends?

Not as much as you think! We've had a plethora of analyses of county voting results purporting to show us which groups really drove the election results. Partially this reflects some justified skepticism about the exit polls' version of events. As awareness of exit poll flaws has spread, it has perhaps encouraged a turn to scrutiny of voting patterns.

Perhaps even more important, data analytic capabilities are now far higher at most media outlets, so it's not too difficult to download and crunch the county voting results and look for patterns. In its simplest form, looking at county voting results from this election indicates that margin shifts in Biden's favor were overwhelmingly in suburban areas, so, since suburban areas = college-educated whites, therefore college-educated whites were responsible for the Biden margin swing.
But the suburbs are vast and diverse; even in the large suburban counties in our biggest metropolitan areas, white noncollege residents still outnumber white college residents by 44 to 25 percent. And in small metro, white noncollege residents are more numerous by 52 to 23 percent. So tying margin shifts in the suburbs to white college voters alone is quite a stretch.
More sophisticated than this, outlets like 538, the Economist and others have taken the county voting results and regressed these margins against various county-level demographic characteristics like percent white college, percent white noncollege, percent Hispanic, etc. Again, this is now quite easy for many places to do, involving no more than downloading the county voting results and also obtaining county level demographics from, say, the government's American Community Survey and running the statistical procedure.
You've no doubt seen the results in various graphics. A whole bunch of dots on a chart, with the margin shift on the y-axis and the demographic characteristic on the x-axis. The dots kinda sorta form a pattern, which is helpfully indicated by the regression line drawn through the dots, with the direction and steepness of the slope indicating the relationship between margin shift and level of the demographic characteristic. Typically no goodness-of-fit measures are reported; the line is the line.
So what does a chart like this really mean? For example, if there's a chart relating margin shift to percent white noncollege in a county and the line slopes a bit downward as white noncollege percent goes up, what does that mean? It literally means that at the county level, and only at the county level, we would predict higher white noncollege counties to have lower margin shifts. But we don't know from this kind of aggregate data exercise what the individual-level behavior of specific voter groups--white noncollege and others--within these counties actually was (to assume otherwise is the ecological fallacy). That is a different, more specific question for which this exercise may suggest hypotheses, but does not directly provide an answer.
But that does not stop writeups of these results from claiming that they have in fact shown how groups within the population moved. To stay with the white noncollege example, the downward sloping regression line is interpreted as meaning that white noncollege voters generally shifted against Biden. That could be. But that is only one possible story.
Another possible story is that white noncollege voters in rural areas, where concentrations of these voters are highest, have a distinctive and homogeneous culture that led them to keep or increase their Trump margins. But only about a fifth of white noncollege adults live in rural areas White noncollege voters in large suburban and small nonmetro counties, where three-fifths of this population lives (including in states like Michigan, Pennsylvania and Wisconsin) and in core urban areas are different economically and culturally and did shift toward Biden. A downward sloping line therefore but with the net margin shift across the demographic likely positive for Biden.
So is that the real story? We don't know that either, merely that it is possible and not at all contradicted by a aggregate plot of county data. And actually more consistent with the available individual-level survey data, whatever with those data may be.
One final illustration from a specific state. In PA, slightly less than a fifth of residents are in rural areas and slightly than a quarter of the white noncollege population. Theses areas had a pro-Biden margin shift of 1 point. In large suburban (58 percent white noncollege) and small metro (64 percent white noncollege) areas , where three fifths of the white noncollege population of PA resides there were, respectively, pro-Democratic margin shifts of 4 and 3 points. In the urban cores, the least white noncollege by far, there was no margin shift at all.
Is it possible that all these white noncollege voters all over PA shifted against Biden and were simply overwhelmed by shifting in the other direction by the relatively small white college population, including in the many heavily white working class counties? Absolutely that's mathematically possible. But we don't somehow "know" that by looking at a scatter plot and the slope of a regression line.
So the next time you see one of these county data charts, take it with a grain of salt. Eventually, we'll have properly modeled 2020 estimates of voter group partisan preference and turnout trends that integrate survey data with voting results and underlying demographics. But until that glorious day, be careful what you believe.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.