To what extent are Democrats inherently disadvantaged in congressional elections because of where they live? This was a hot topic of conversation just after the 2012 congressional elections, when the Democrats won a majority of the popular vote but a distinct minority of the seats in the House of Representatives. The subject has recently resurfaced in public debate, and the issues are important enough that they deserve some consideration.
First, some background. Jowei Chen and Jonathan Rodden (hereafter, C&R) have developed a computer program that draws thousands of simulated districts using precinct-level data and a handful of systematic criteria. They then compare the actual redistricting plan to the simulated ones. If they don’t differ substantially, then any bias in the district lines cannot be blamed on gerrymandering. Or so goes the argument.
This method has attracted considerable attention because it appears to show that Democrats suffer because they are concentrated in urban areas and “waste” votes on big wins, not because Republicans have drawn the lines that way. In other words, Democrats are “unintentionally gerrymandered.” C&R are not the only ones to make this claim (for a recent example, see Nate Silver’s post here), but those who make it commonly cite this research.
Michael McDonald, a respected expert on redistricting, recently offered a thoughtful critique of the C&R method, calling it unreliable and divorced from the reality of redistricting. He has pointed to the many fair and apparently usable maps drawn by the public in the last redistricting cycle—including in key states where C&R had claimed a strong bias against the Democrats—as evidence against the C&R conclusions.
It does not help C&R’s case that they recently updated their analysis for Florida—the first and most famous application of their method—with more recent presidential election data and found a very different result. Instead of an inherent bias against Democrats and no clear sign of gerrymandering, they found no inherent bias and a strongly gerrymandered outcome.
So what does this mean? Is the C&R method reliable, and is it telling us something useful about redistricting? McDonald argues that algorithms will often produce “nonsensical, spaghetti-like districts” that would never be part of a real plan. Worse, he argues we will never know how often an algorithm like C&R’s will make this mistake, so the product of the algorithm can’t be trusted.
I don’t doubt McDonald’s experience with these algorithms, but I think his criticism misses the point of the C&R approach. Unlike many past efforts at automated redistricting, this algorithm isn’t meant to replace human line-drawers by generating maps that would clear every legal and political hoop. Instead, it offers some sense of the range of unbiased possibilities given the realities of political geography. It asks, “What if we drew districts at random according to a few simple rules? Would the results be any different than the actual plan?” If the answer is “no,” then the humans drew lines as if they were unbiased computers, regardless of any other factors that led them to choose one set of lines in particular.
As for whether the algorithm produces weird districts, all of C&R’s simulated maps for Florida are posted on the web, and to the naked eye they don’t look much worse than many legally adopted plans in other states. And that’s without directing the algorithm to favor compactness. When C&R take that extra step, the results are downright boxy.
To be fair, C&R sometimes imply that in certain states it’s impossible to draw compact districts that are fair to both parties. I don’t think that’s quite right. Their method does not tell us what is possible—only what is likely in the absence of any partisan intent. By contrast, the authors of the publicly-drawn districting plans McDonald points to may have had a partisan intent—specifically, the intent to undo the underlying geographic bias. That suggests something important about redistricting reform: the map-makers may need to take explicit account of partisanship to avoid reproducing the inherent bias in the underlying political geography. That’s a little ironic, since redistricting reformers often propose that map-makers ignore partisanship completely. But it may be true all the same.
What about Florida? Doesn’t the complete reversal of C&R’s original finding suggest serious flaws in their method? Not necessarily. In fact, the same shift in political geography that C&R uncover is visible in the raw data. Democrats won 50% of the presidential vote but only 47% of the precincts in 2000, while in 2012 they won 51% of both. In other words, measured this way the bias disappeared. There are problems with using only these raw data, to be sure. A better method would correct for large differences in population between precincts, and would take into account how the precincts are related to each other in geographic space.* In other words, it would do exactly the things the C&R method already does. Nonetheless, the fact that multiple methods all point to a similar result should give us some confidence that the C&R method isn’t doing something screwy.
But if the C&R method is valid and useful, doesn’t the result in Florida then prove that “unintentional gerrymandering” isn’t a factor anymore? It certainly might, though it’s too soon to say. We need to apply the C&R algorithm to more recent presidential results in other states. There may have been similar change outside Florida, but given the continued concentration of Democrats in urban areas I’d be skeptical it has been too widespread.
What the Florida case does highlight, however, is a point that John and I have made before: redistricting doesn’t guarantee results. Consider how quickly the reality in Florida has shifted. In 2001, Republicans in Florida drew districts that reflected the state’s anti-Democratic geographic bias, but by 2008 (the year of the updated C&R numbers) this bias had disappeared and in 2011 Republicans were shoring up an eroding position. Could their efforts hold back the tide for the next 10 years? Maybe. But based the last 10, I wouldn’t be terribly confident.
Moreover, none of this considers the imperfect link between the presidential vote and the vote for Congress. While the two are more closely linked than ever these days, they are still different. Incumbents routinely outperform their party’s presidential candidate, and strong challengers or uniform partisan swings can overwhelm the most carefully drawn redistricting plan. John and I estimated that incumbency alone could account for the entire discrepancy between the national seat share and vote share this cycle.
Gerrymandering is certainly real, and it was egregious in a number of large states this last election cycle. Since even a single unfair outcome is one too many, independent commissions ought to be the preferred way to ameliorate the problem. But we should be cautious about expecting too much from such reforms. A neutral process will always carry the risk of a biased outcome, so we should either require plans to eliminate bias or accept that sometimes bias will happen without intent.
(Thanks to Jowei Chen for sharing precinct-level data from Florida. Naturally, he is not to blame for any errors in the manner they are used.)
*However, size was largely uncorrelated with Democratic vote share in the precincts, so its impact on the raw numbers presented here is probably modest. If anything, it may understate the degree of bias against Democrats since larger precincts were very slightly more Republican than smaller ones.
[Cross-posted at The Monkey Cage]