(This post is intended for people who are following the DP-at-the-Census story. More background will come in another post, maybe. EDIT: My blog posts (part 1; part 2) hosted by BU’s Center for Reliable Information Systems and Cyber Security provides an inadequate introduction.)
Some folks in the differential privacy (DP) community have asked me who will win in Alabama’s lawsuit against the Department of Commerce over the use of DP in the disclosure avoidance system for the 2020 Census [0]. Here is why I’d guess that we are unlikely to get a ruling that the use of DP is per se illegal.
Greatly oversimplifying, Alabama is arguing that the use of DP in the 2020 DAS is illegal, based on the language of 13 USC 141 [1]. While confidentiality and data quality are central to the policy questions surrounding DP at the Census, Alabama’s lawsuit raises numerous other legal issues. To rule that DP in the DAS is per se illegal, the court would have to overcome a handful of legal issues that weigh against Alabama’s lawsuit while substantively expanding voting rights jurisprudence. Alabama addresses each of these points—some convincingly—but together they weigh against Alabama’s case. Considering that the case is likely to get appealed to a very conservative Supreme Court, I’d guess that this won’t happen.
Disclaimer: I am not a lawyer and I am very far from an expert in the many legal issues in this case. These are some thoughts of an outside observer who has no special knowledge or formal legal training. Take the below with an ocean of salt. Moreover, there are other ways a court could disrupt the Census’s use of DP without ruling that it is illegal. For example, they could issue a preliminary injunction, or rule that the Bureau did not follow the right processes in making its decision as required by the Administrative Procedures Act.
Leave a comment if you think I’m missing or misunderstanding anything, or with any questions / clarifications / mistakes.
The first few legal hurdles are all related to whether or not Alabama has standing—the capacity to bring a lawsuit in the first place.
Injury
Alabama must point to specific injuries that they will suffer if Census uses the 2020 DAS. They argue that funding formulas will probably get messed up, but cannot name a particular program that will lose funding. More importantly, they point to redistricting—arguing that the 2020 DAS will make it impossible to precisely balance the population of Congressional districts as required by law. But case law requires only a good faith effort to balance population using the best data available, and even that requirement is not absolute (see OPOV below). Moreover, Alabama’s own constitution provides for the possibility that the decennial census data will be unsatisfactory for apportionment and redistricting, granting the state legislature the power to conduct its own enumeration.
Redress
Alabama must show that enjoining the use of DP in the 2020 DAS will likely remedy the injury. But the Census Bureau claims that if it was forced to use the non-DP techniques used in 2010, the quality of the data would get worse instead of better. Moreover, enjoining the use of DP would require the whole disclosure avoidance system to be overhauled. Census states that this would significantly delay the data release. This is relevant because in the same lawsuit, Alabama is arguing that release of data has already been illegally delayed. So the remedy for the DP claim would likely exacerbate the injury in the delay claim. I have no idea how this plays out in a classroom, let alone a courtroom.
Statistical methods
Alabama is challenging the use of DP under a law that regulates the use of “statistical methods” in the decennial census [2]. Alabama contends that “It’s clear that differential privacy falls into this category.” After all, DP involves statistics, right? But “statistical method” means something specific. Roughly, the statute defines a statistical method as “representative sampling, or any other statistical procedure . . . [to adjust the] enumeration of the population as a result of statistical inference.” As used in the 2020 DAS, DP is not doing representative sampling nor statistical inference. The 2020 DAS is a two-stage algorithm: a DP stage and a post-processing stage. It’s conceivable that one could maaaybe argue that the post-processing stage uses statistical inference. But that’s a separate question that nobody has raised, and doesn’t even implicate DP itself. On the other hand, the Department of Commerce’s filings haven’t done a great job of challenging Alabama’s argument on this point.
Aggrieved party
Alabama’s challenging the use of DP under a law that allows any “aggrieved person” to sue [2]. But the statute defines aggrieved persons to include some flesh and blood natural persons, the House of Representatives, and the Senate. States are not on the list, as pointed out by one of the judges during the hearing [3].
One Person One Vote (OPOV)
Alabama argues that the use of DP in the 2020 DAS would make lawful redistricting impossible, and therefore would itself be unlawful. Most states, including Alabama, draw Congressional districts so that their populations—as reported by the most recent decennial census—differ by at most one. This practice comes from a reading of Karcher v Daggett [4] where the Supreme Court held that there is no de minimis population deviation. A state must show that “each significant variance between districts was necessary to achieve some legitimate goal.”
But the OPOV obligation isn’t to balance the actual population of voting districts. That’s impossible. Error in the Census is nothing new, and populations shift between April 1st of the decade and the time that districts are finalized. OPOV requires a “good-faith effort” at balancing the population as well as possible using the best data available, and even that requirement is not absolute. The best data available has always been—and will continue to be—the decennial census data. (Unless Alabama wants to conduct its own enumeration as its state constitution allows.)
Finally, existing OPOV case law imposes an obligation on the states, not an obligation on the Census Bureau. I believe that the question of whether OPOV imposes an obligation on the Census Bureau has never actually been raised, so this is a case of first impression. Holding that the 2020 DAS’s use of DP violates OPOV, would be making new law and strengthening voting rights. I think it’s unlikely that the current Supreme Court would do so.
Tabulations
Section 141(c) of Title 13 requires Census to produce “tabulations of population.” Alabama argues that “The plain meaning of tabulation of population . . . is fairly obvious: one counts the number of persons . . . and enters that number into a table.” But this view is hard to square with another part of the same law. In Sections 141(a) and 195 of Title 13 [5], Congress specifically authorizes the use of statistical sampling to produce these tabulations. Census is only prohibited from using sampling to apportion representation in Congress (i.e., how many representatives each state gets) [6]. Otherwise, sampling is fair game.
In it’s complaint, Alabama fails to distinguish between tabulation-by-sampling as authorized by Congress and tabulation-by-2020-DAS. If Census used sampling to create the redistricting data, it would require much more than counting the persons and entering the number into a table. It would involve mathematics as sophisticated as anything going on in the 2020 DAS.
Timeliness
Alabama walks a fine line in arguing that the use of DP in the 2020 DAS is ripe for review, but that their lawsuit was not unreasonably delayed. The Census announced its plans to use DP in late 2017 and has periodically released demonstration code and data since 2019. At the same time, Census is still tuning the final parameters of the 2020 DAS—the latest demonstration data came out April 2021. In the initial hearing on Alabama’s motion for a preliminary injunction, judges gave Alabama a hard time on this point. If Alabama’s challenging the use of DP per se, then why not bring the case years ago? If instead Alabama’s arguing that the implementation of DP in the 2020 DAS is problematic, then why is it ready to be reviewed?
Deference to the experts
Courts generally defer to agency expertise on the specific way they carry out their mandates. It seems that Alabama needs to convince the court to substitute its own interpretation of the Title 13’s confidentiality provisions for the Census Bureau’s, though I’m unsure what the appropriate standard is (not Chevron Deference). Alabama’s complaint addressed this issue only very briefly.