[EL] Milhiser article on partisan gerrymandering

Sun Dec 7 18:00:44 PST 2014

Jonathan,

I read your expert report in the Florida litigation long ago. For others' reference, it is here:

https://docs.google.com/file/d/0B0bHdAFS4MgqamUycEV1ZTlXUUE/edit

I base my comments about how your algorithm performs with respect to compactness from statements such as this on p.5 of your expert report:

"Our approach is to leave aside any data about partisanship or demographic characteristics other than population counts, and allow the computer algorithm to generate a complete and valid districting plan based purely on geographic contiguity and compactness."

Later, you call districts created by algorithm "reasonably compact" (p. 7), but provide no definition or evidence of what you deem to be "reasonable".

These statements echo what you wrote in QJPS, describing your algorithm as, for example, performing "...in a manner that preserves the geographic compactness of the districts" (p. 250).

These are strong statements asserting your algorithm produces compact districts. There is no qualifier to the effect that only most of the time your algorithm produces compact districts or what it may mean if it does not. Defendants did academia a favor by demonstrating that at least in some instances, your algorithm fails to produce visually compact districts. If one did not believe the operation systems literature on the problems with sampling from an NP-Hard graph (of which redistricting is a case, see "The Promise and Perils of Computers in Redistricting" http://informatics.mit.edu/publications/promise-and-perils-computers-redistricting), these counter-examples are proof that your algorithm does not indeed randomly sample from the set of compact redistricting plans. Your algorithm is biased by at least including some non-compact districts, and likely in other ways that we cannot know because of the nature of the problem. Thus, any *statistical* inferences drawn from it are suspect (but see below about what these plans reveal about the legislature's preferences). 

In the world of expert witness testimony where one is always best off carefully clarifying statements. Unlike the peer-review process, the opposing testifying and consulting experts will pick over your work looking for exactly these sorts of statements and contradictory evidence in your methods and data that can impeach your credibility.

Please point it out to me if I missed it, but in neither your expert report nor your QJPS article do your report the compactness of the districts you create. You simply assert your algorithm produces the desired result. You set a very low bar for what is compact if the baseline is the legislatures' non-compact plan (see the papers I linked to in my original email). If that is the bar for what constitutes "compactness" then you should revisit your QJPS article. The alternative maps from the public we analyze in our research demonstrate that not only is it possible to draw districts that are much more compact than the redistricting authorities', but have fair plans, too (and respect the VRA, respect local political boundaries, and create more competitive districts).

As for respecting the Voting Rights Act, I said "most" algorithms. I know your QJPS article does not respect the VRA, but your expert report attempts to do so. It is important to understand the ad hoc nature of your modification of your QJPS algorithm to your expert report as it affects your conclusions. You fix the VRA districts drawn by the legislature, and apply your algorithm to the remaining districts. The algorithm will thus almost certainly be unable to find alternative configurations of districts that will satisfy the VRA. For example, in our Virginia redistricting competition, William and Mary law school students discovered an alternative configuration of the Third Congressional District, such that the Fourth District became the state's majority-minority district and the Third became an influence district. This outcome could only be realized by shifting some portions of the Third to the Fourth District (Rep. Bobby Scott spoke favorably of this arrangement, Senate Democrats proposed a substantially similar arrangement in a later bill, and my racial bloc voting analysis for the successful litigation indicated that the Third drawn in this manner would not be retrogressive). In this example, your modified algorithm would not fully sample from the space of feasible compact districts that also respect the Voting Rights Act. While that may not be important for the expert report, it would be for future use of this modification in research.

I also have read Nolan McCarty's rebuttal report. In it he notes that you changed a component of your QJPS methodology in a manner resulting in a finding consistent with your client's argument. Since this change is consequential to your QJPS findings, which many find compelling, it is worthwhile to review this criticism.

Nolan notes that in your QJPS article you "normalized" the presidential vote shares within districts such that you simulate a hypothetical 50/50 electorate. In your expert report, you did not normalize the presidential vote, but simply used the vote shares without modification. When Nolan recalculated your expert report such that the statistics were consistent with your QJPS approach, he found that the legislature's plan could no longer be accepted as a partisan gerrymander.

Should we normalize presidential vote or not is an interesting academic question. Democratic presidential candidates have won a plurality of the national popular vote in 5 of the last 6 presidential elections. It strikes me that at least in presidential elections, which is the metric we have to use to compare districts nationally, the baseline presidential vote is probably slightly Democratic leaning. Your and Nolans' expert reports thus indicate that this seemingly-innocuous normalization procedure affects your substantive QJPS findings that Democrats gerrymander themselves where they live (setting aside my criticisms of the algorithm).

After presented with this evidence, if Judge Lewis thought highly of your method he did not reference your report or testimony in his ruling overturning two specific districts, but not the redistricting plan:

http://judicial.clerk.leon.fl.us/image_orders.asp?caseid=444437&jiscaseid=&defseq=&chargeseq=&dktid=102203318&dktsource=CRT

Finally, you misunderstand what crowd-sourcing can provide, although you offer no critique of the method of revealed preferences other than stating your method is superior. Much of what your expert report attempts to do is the method of revealed preferences, by comparing the legislature's plan to alternative plans on the metrics that the legislature claimed it cared about (in this case the state constitution). If your machine-generated plans do better than the legislature's, then that reveals something about the legislature's intent. Likewise, courts often require plaintiffs to submit at least one such alternative plan to rule against a redistricting authority or may examine plans introduced during the redistricting process to divine intent.

The original word "computer" is refers to a person that computes, not a machine. In many applications, human computers have capacities to see beyond machine computers. Our experience with the limitations of redistricting optimization algorithms, which you modified to your purposes in early versions of your work, led us to embrace what humans are capable of. Our thoughtful reflection on these limitations led us to reject the approach of "sampling" from the space of redistricting plans since such sampling algorithms by the very nature of the problem cannot be known to sample without bias. Humans, of course, also have biases. Still, all plans, whether they are drawn by humans or machines, may through the method of revealed preferences reveal the intent of a redistricting authority. We thus advocate for the generation of more plans during the redistricting process, by machines and humans, to better map out the space of feasible alternatives on multiple dimensions (Nagel first proposed this approach in 1965). These alternative plans that were rejected (the road not taken) allow courts to more critically assess plans adopted by a redistricting authority. Our paper on the Ohio redistricting by formula experiment and our current research into Mexico's experience with automation are perhaps the most intriguing, in that it is possible to quantify what values we may want from a redistricting plan and let the best scoring plan win, which in the case of Mexico, were plans drawn by humans, not computers.

Best regards,

-Mike

============
Dr. Michael P. McDonald
Associate Professor
University of Florida
Department of Political Science
234 Anderson Hall
P.O. Box 117325
Gainesville, FL 32611

phone:   352-273-2371 (office)
e-mail:  dr.michael.p.mcdonald at gmail.com                
web:     www.ElectProject.org 
twitter: @ElectProject
-----Original Message-----
From: Jonathan Rodden [mailto:jrodden at stanford.edu] 
Sent: Friday, December 05, 2014 11:47 AM
To: Michael McDonald
Cc: law-election at UCI.EDU
Subject: Re: [EL] Milhiser article on partisan gerrymandering

We are not sure where Michael McDonald got such inaccurate information about the Florida redistricting trial.  Perhaps he was unable to find our expert reports online and jumped to the wrong conclusions. The judge did not allow any written expert reports into evidence.  We presented our reports orally at the trial over the course of two days of direct testimony. The plaintiffs relied heavily on the simulation analysis both at trial and in post-trial briefs.  

Indeed, the defense invested heavily in unsuccessful attempts to discredit the simulations. Early in the process, they sifted through thousands of maps and called into question the compactness of a handful of our districts.  In response, we generated a large number of simulations that were equally or more compact than the legislature's plan.  This settled the issue. We also produced a large number of simulations that were equally or more respectful of county and municipal boundaries than the legislature's plan.  Most importantly, we presented a large number of simulations that respected every possible interpretation of the VRA that might conceivably apply in Florida.  

Regardless of the approach, the result was the same: the legislature's plan was an extreme outlier relative to the non-partisan simulations.

Concerns about the fact that we did not draw every conceivable valid plan were not taken seriously at trial.  

We remain puzzled by the claim that a handful of maps submitted online by interest groups or college students with time on their hands would be a better baseline against which to measure a proposed plan.  Nevertheless, we also applied McDonald's favored crowd-sourcing approach in one of our reports by analyzing all of the plans submitted by "the public" through Florida's online interface.   

This analysis revealed that motivated Democrats and Republicans can indeed generate highly partisan plans that are beyond the tails of the distribution of the simulations.  One such plan caught our eye for its truly impressive bias.  This led to the discovery that an individual's gmail account had been used as a conduit, via the online "open redistricting" platform, to get plans drawn by political operatives into the hands of legislators.  The wisdom of crowds indeed.  

Perhaps for good reason, Judge Lewis was ultimately more impressed by this type of testimony than by any of the social scientific reports.  But by no means does this indicate that simulations cannot be useful in the courtroom. Our experience in Florida was quite to the contrary. 

Going forward, the demand for a non-partisan baseline in cases like these will remain strong, and crowd-sourcing is most likely not the answer.    

Jonathan

----- Original Message -----
From: "Michael McDonald" <dr.michael.p.mcdonald at gmail.com>
To: "law-election at UCI.EDU" <law-election at uci.edu>
Sent: Thursday, December 4, 2014 8:38:18 AM
Subject: Re: [EL] Milhiser article on partisan gerrymandering

In addition to the deficiencies of most automated algorithms failing to adhere to all legal requirements like the VRA, systems science research indicates that given current computing limits we cannot be sure plans produced by automated algorithms are indeed random. Thus, we cannot apply statistical tests. Furthermore, as defense lawyers vividly demonstrated in the recent Florida litigation by displaying a series of maps to impeach Rodden’s testimony, Chen and Rodden’s algorithm does not necessarily produce compact districts. This evidence was so damning that plaintiffs dropped Rodden’s testimony and expert report from their arguments. Which is to say, there are theoretical and implementation assumptions that underlie automated algorithms that go underappreciated by their proponents, which can lead to consequences when a bright light is shined upon them by opposing experts. 
(Ironically, defendant’s expert for Republicans in Florida was Nolan McCarty, the journal editor who published Chen and Rodden’s QJPS article that found Republicans didn’t gerrymander, which contrasted with the expert report for plaintiffs.)

For these reasons, Micah Altman (if I may speak for him) and I believe that while automated algorithms can illuminate alternative plans -- an idea first proposed by Nagel in 1965 -- human beings approach redistricting in fundamentally different ways that can explore a greater range of legal plans than what computers are capable of producing. For example, during Mexico’s recent redistricting humans repeatedly beat an optimization algorithm using simulated annealing (from a paper still in working form, but willing to share on request). Because these human-generated plans still cannot guarantee to randomly sample from the feasible set of redistricting plans, we cannot apply statistical tests to them. However, by the weak axiom of revealed preferences, human and computer-generated plans can reveal preferences of a redistricting authority, in that if a plan exists that beats an adopted plan on all metrics that the redistricting authority stated it cared about, and it differs on another criteria, say is more “fair”, then we can infer that the redistricting authority didn’t prioritize fairness. 
This approach is routinely used by the courts when they ask plaintiffs to produce a plan to correct alleged constitutional defects. This idea, in part, motivated our creation of open-source redistricting software called DistrictBuilder (www.districtbuilder.org), which supported public mapping advocacy efforts across the United States and was used by a few redistricting authorities, too.

Some papers on our approach of crowd-sourcing redistricting, and what these plans reveal about the preferences of redistricting authorities:

Public Participation GIS: The Case of Redistricting

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2321870

The Promise and Perils of Computers in Redistricting

http://informatics.mit.edu/publications/promise-and-perils-computers-redistricting

Paradoxes of Political Reform: Congressional Redistricting In Florida

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2421908

Redistricting by Formula: An Ohio Reform Experiment

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2450645

A Half-Century of Virginia Redistricting Battles: Shifting from Rural Malapportionment to Voting Rights to Public Participation

http://informatics.mit.edu/publications/half-century-virginia-redistricting-battles-shifting-rural-malapportionment-voti-0

============
Dr. Michael P. McDonald
Associate Professor
University of Florida
Department of Political Science
234 Anderson Hall
P.O. Box 117325
Gainesville, FL 32611

phone:   352-273-2371 (office)
e-mail:  dr.michael.p.mcdonald at gmail.com
web:      <http://www.electproject.org/> www.ElectProject.org
twitter: @ElectProject

From: law-election-bounces at department-lists.uci.edu
[mailto:law-election-bounces at department-lists.uci.edu] On Behalf Of Nicholas Stephanopoulos
Sent: Wednesday, December 03, 2014 11:08 AM
To: Rob Richie
Cc: law-election at UCI.EDU
Subject: Re: [EL] Milhiser article on partisan gerrymandering

More comprehensive versions of the Duke study have been carried out by Jowei Chen and Jonathan Rodden (see here <http://www-personal.umich.edu/~jowei/florida.pdf> ) and Roland Fryer and Richard Holden (see here <http://www.stat.columbia.edu/~gelman/stuff_for_blog/fryer.pdf> ). None of these studies takes the VRA into account (or, for that matter, other common criteria such as respect for political subdivisions and respect for communities of interest). So I think it's fair to say that we don't know what the partisan implications of randomly drawn lawful plans would be. We know what the partisan implications are of randomly drawn and maximally compact plans, but these aren't the same thing as legally valid plans.
On the issue of gerrymandering more generally, Eric McGhee and I have a forthcoming article (see here <http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2457468> ) in which we introduce a new measure of partisan distortion and argue that it could serve as the basis for a legal test. We're responding not to Vieth but rather to the favorable comments by five Justices in LULAC about the concept of partisan symmetry.

Nick

On Wed, Dec 3, 2014 at 5:30 AM, Rob Richie <rr at fairvote.org> wrote:

The mathematicians' failure to acknowledge the Voting  Rights Act is just one indicator of how  such simulations are only  as good as the assumptions that govern them. Worth doing and discussing, but not to be accepted without an understanding of the impact of those assumptions.

Anther example in the article is that it is based on analyzing the actual votes cast in House races rather than the more indicative relative presidential vote. The 2012 elections included some over-performing Democratic candidates including one incumbent who received more than 20% of votes from backers of Mitt Romney -- an extraordinarily high crossover vote in today's partisan climate that didn't continue this year.

I'll also address one lurking  part of this discussion: the myth that section two of the Voting Rights Act is the cause of the Democrats' problems in the South. That certainly was true in the past,  but it is not today absent adoption of independent redistricting commissions. That is, the need to create majority-minority districts is the only reason that Republicans drew plans that allows Democrats to win any congressional seats at all in Alabama, Louisiana, Mississippi and South Carolina. Republican mapmakers also could have knocked Democratic representation down two seats in Georgia and one seat each in North Carolina and Tennessee. (We show this in some alternative southern congressional districts <http://www.fairvote.org/assets/Uploads/Redistricting-Reform-in-the-South.pdf>
map work we did earlier in the year.)

As it is, the combination of single winner districts, polarized voting and partisan  control of redistricting makes most of the South a true basketcase when it comet representative democracy in congressional elections. Note that when we issued our final November 2016 election projections <http://www.fairvote.org/research-and-analysis/congressional-elections/fairvotes-projections-for-u-s-house-elections-in-2016/>
earlier this month (on November 6th, more than two years before the 2016 elections), we projected winners in more than 85% of districts using a methodology that resulted in a single error in more than 700 projections in 2012-2014.

In the South, the numbers are  extraordinary -- setting aside Florida and Virginia, there are 126 districts in the remaining southern states and the border  states of Kansas, Missouri, Kentucky and West Virginia. We project winners in 125 of them; if all these seats were open, we'd still project 123 of them. The prospects of redistricting reform in any of these states is very slight without Congress asserting its power over the time, place and manner of elections -- something it onl  is using right now involving redistricting  to mandate gerrymandering (that is, mandate single-winner districts, as it required in 1967).

Rob Richie

On Wed, Dec 3, 2014 at 1:29 AM, Douglas Johnson <djohnson at ndcresearch.com>
wrote:
Such studies have been done for decades. In fact the difference between the party getting more votes and the party getting more seats was fairly central to many of the arguments (though not to the ruling) in the Tom Delay Texas re-redistricting case. There is no “genius” to such analysis, despite Millhiser’s use of that term.

The more simplistic versions of such studies (including this Mattingly and Vaughn study) completely ignore the requirements of the Voting Rights Act. 
In fact, the phrase “voting rights act” never even appears in the Mattingly and Vaughn paper. The only remote reference to it is “The two districts with largest African American representation had on average around 36% and 32% African American population, which compares favorably to the state wide percentage of 22%, but not to the current districts.” That sounds to me either like a total failure to understand the VRA, or a knowing attempt to distract reporters from recognizing that they’re deliberately ignoring the VRA despite the VRA’s significant impact on the redistricting plan in question.

Millhiser specifically condemns redistricting plans that “are almost certainly the product of a legislature that carefully designed the maps to produce a desired result.” But that is exactly what the VRA requires – on a racial/ethnic and language front. The difficulty is not in proving that redistricting plans are not random (or at least that they’re not the most-compact plan possible). The difficulty is in proving the plans are drawn for partisan purpose, as opposed to drawn ensure compliance with the VRA or to follow city, county, or other community of interest boundaries. 
The filings in Vieth  provide a look at a variety of potential tools for such analysis. Millhiser and Vaughn do not.

This is a topic in desperate need of good solid analysis and study. I often describe Vieth as the Supreme Court’s version of a “Call for Papers.” 
Unfortunately neither the article nor the study qualify as such.

-          Doug

Douglas Johnson, Fellow
Rose Institute of State and Local Government at Claremont McKenna College  <mailto:douglas.johnson at cmc.edu> douglas.johnson at cmc.edu  <tel:310-200-2058> 310-200-2058

 <http://electionlawblog.org/?p=68747> “If The Supreme Court Reads This Study, It Could End Partisan Gerrymandering Forever”

Posted on  <http://electionlawblog.org/?p=68747> December 2, 2014 8:10 am by <http://electionlawblog.org/?author=3> Rick Hasen  <http://thinkprogress.org/justice/2014/12/01/3597643/this-study-could-end-the-worst-kinds-of-gerrymandering-if-the-supreme-court-took-the-time-to-read-it/>Ian Millhiser writes.I was surprised to see no mention of Voting Rights Act requirements indrawing these districts._______________________________________________Law-election mailing listLaw-election at department-lists.uci.eduhttp://department-lists.uci.edu/mailman/listinfo/law-election_______________________________________________Law-election mailing listLaw-election at department-lists.uci.eduhttp://department-lists.uci.edu/mailman/listinfo/law-election--Nicholas O. StephanopoulosAssistant Professor of LawUniversity of Chicago Law Schoolnsteph at uchicago.edu(773) 702-4226http://www.law.uchicago.edu/faculty/stephanopoulos