Monday, January 22, 2018

Double blind review: continuing the discussion

My first two posts on double blind review triggered good discussion by Michael Mitzenmacher and Boaz Barak (see the comments on these posts for more).  I thought I'd try to synthesize what I took away from the posts and how my own thinking has developed.

First up, I think it's gratifying to see that the the basic premise: "single blind review has the potential for bias, especially with respect to institutional status, gender and other signifiers of in/out groups" is granted at this point. There was a time in the not-so-distant past that I wouldn't be able to even establish this baseline in conversations that I'd have.

The argument therefore has moved to one of tradeoffs: does the installation of DB review introduce other kinds of harm while mitigating harms due to bias?

Here are some of the main arguments that have come up:

Author identity carries valuable signal to evaluate the work. 

This argument manifested itself in comments (and I've heard it made in the past). One specific version of it that James Lee articulates is that all reviewing happens in a resource-limited setting (the resource here being time) and so signals like author identity, while not necessary to evaluate the correctness of a proof, provide a prior that can help focus one's attention. 

My instinctive reaction to this is "you've just defined bias". But on reflection I think James (and others people who've said this) are pointing out that abandoning author identity is not for free. I think that's a fair point to make. But I'd be hard pressed to see why this increase in effort negates the fairness benefits from double blind review (and I'm in general a little uncomfortable with this purely utilitarian calculus when it comes to bias).

As a side note, I think that focusing on paper correctness is a mistake. As Boaz points out, this is not the main issue with most decisions on papers. What matters much more is "interestingness", which is very subjective and much more easily bound up with prior reactions to author identity. 

Some reviewers may be aware of author identity and others might not. This inconsistency could be a source of error in reviewing.

Boaz makes this point in his argument against DB review. It's an interesting argument, but I think it also falls into the trap of absolutism: i.e imperfections in this process will cause catastrophic failure. This point was made far more eloquently in a comment on a blog post about ACL's double blind policy (emphasis mine). 

I think this kind of all-or-nothing position fails to consider one of the advantages of blind review. Blind review is not only about preventing positive bias when you see a paper from an elite university, it’s also about the opposite: preventing negative bias when you see a paper from someone totally unknown. Being a PhD student from a small group in a little known university, the first time I submitted a paper to an ACL conference I felt quite reassured by knowing that the reviewers wouldn’t know who I was. 
In other words, under an arXiv-permissive policy like the current one, authors still have the *right* to be reviewed blindly, even if it’s no longer an obligation because they can make their identity known indirectly via arXiv+Twitter and the like. I think that right is important. So the dilemma is not a matter of “either we totally forbid dissemination of the papers before acceptance in order to have pure blind review (by the way, 100% pure blind review doesn’t exist anyway because one often has a hint of whom the authors may be, and this is true especially of well-known authors) or we throw the baby out with the bathwater and dispense with blind review altogether”. I think blind review should be preserved at least as a right for the author (as it is know), and the question is whether it should also be an obligation or not.

Prepublication on the arXiv is a desirable goal to foster open access and the speedy dissemination of information. Double blind review is irrevocably in conflict with non-anonyous pre-print dissemination.

This is perhaps the most compelling challenge to implementing double blind review. The arXiv as currently constructed is not designed to handle (for e.g) anonymous submissions that are progressively blinded. The post that the comment above came from has an extensive discussion of this point, and rather than try to rehash it all here, I'd recommend that you read the post and the comments. 

But the comments also question the premise head on: specifically, "does it really slow things down" and "so what?". Interestingly, Hal Daumé made an attempt to answer the "really?" question. He looked at arXiv uploads in 2014-2015 and correlated them with NIPS papers. The question he was trying to ask was: is there evidence that more papers uploaded to the arXiv before submission to NIPS in the interest of getting feedback from the community? His conclusion was that there was little evidence to support the idea that the arXiv had radically changed the normal submit-revise cycle of conferences. I'd actually think that theoryCS might be a little better in this regard, but I'd also be dubious of such claims without seeing data.

In the comments, even the question of "so what?" is addressed. And again this boils down to tradeoffs. While I'm not advocating that we ban people from putting their work on the arXiv, ACL has done precisely this, by asserting that the relatively short delay between submission and decision is worth it to ensure the ability to have double blind review.

Summary

I'm glad we're continuing to have this discussion, and talking about the details of implementation is important. Nothing I've heard has convinced me that the logistical hurdles associated with double blind review are insurmountable or even more than inconveniences that arise out of habit, but I think there are ways in which we can fine tune the process to make sense for the theory community. 

Tuesday, January 09, 2018

Double blind review at theory conferences: More thoughts.

I've had a number of discussions with people both before and after the report that Rasmus and I wrote on the double-blind experiment at ALENEX. And I think it's helpful to lay out some of my thoughts on both the purpose of double blind review as I understand it, and the logistical challenges of implementing it.

What is the purpose of double blind review? 

The goal is to mitigate the effects of the unconscious, implicit biases that we all possess and that influence our decision making in imperceptible ways. It's not a perfect solution to the problem. But there is now a large body of evidence suggesting that

  • All people are susceptible to implicit biases, whether it be regarding institutional status, individual status, or demographic stereotyping. And what's worse that we are incredibly bad at assessing or detecting our own biases. At this point, a claim that a community is not susceptible to bias is the one that needs evidence. 
  • Double blind review can mitigate this effect. Probably the most striking example of this is the case of orchestra auditions, where requiring performers to play behind a screen dramatically increased the number of women in orchestras. 
What is NOT the purpose of double blind review? 

Double blind review is not a way to prevent anyone from ever figuring out the author identity. So objections to blinding based on scenarios where author identity is partially or wholly revealed are not relevant. Remember, the goal is to eliminate the initial biases that come from the first impressions. 

What makes DB review hard to implement at theory venues? 

Theory conferences do two things that are different from other communities. We
  • require that PC members do NOT submit papers
  • allow PC members to initiate queries for external subreviewers. 
These two issues are connected. 
  1. If you don't allow PC members to submit papers, you need a small PC. 
  2. If you have a small PC, each PC member is responsible for many papers. 
  3. If each PC member is responsible for many papers, they need to outsource the effort to be able to get the work done. 
As we mentioned earlier, it's not possible to have PC members initiate review requests if they don't know who might be in conflict with a paper whose authors are invisible. So what do we do? 

There's actually a reasonably straightforward answer to this. 


  • We construct the PC as usual with the usual restrictions.
  • We construct a list of “reviewers”. For example, "anyone with a SODA/STOC/FOCs paper in the last 5 years” or something like that. Ideally we will solicit nominations from the PC for this purpose.
  • We invite this list of people to be reviewers for SODA, and do this BEFORE paper submission
  • authors will declare conflicts with reviewers and domains (and reviewers can also declare conflicts with domains and authors) 
  • at bidding time, the reviewers will be invited to bid on (blinded) papers. The system will automatically assign people. 
  • PC members will also be in charge of papers as before, and it’s their job to manage the “reviewers” or even supply their own reviews as needed. 
Any remaining requests for truly external sub reviewing will be handled by the PC chairs. I expect this number will be a lot smaller.

Of course all of this is pretty standard at venues that implement double blind review. 

But what if a sub-area is so small that all the potential reviewers are conflicted

well if that's the case, then it's a problem we face right now. And DB review doesn't really affect it. 

What about if a paper is on the arXiv? 

We ask authors and reviewers to adhere to double blind review policies in good faith. Reviewers are not expected to go hunting for the author names, and authors are expected to not draw attention to information that could lead to a reveal. Like with any system, we trust people to do the right thing, and that generally works. 

But labeling CoI for so many people is overwhelming.

It does take a little time, but less time than one expects. Practically, many CoIs are handled by institutional domain matching, and most of the rest are handled by explicit listing of collaborators and looking for them in a list. Most reviewing systems allow for this to be automated. 

But how am I supposed to know if the proof is correct if I don't know who the authors are. 

Most theory conferences are now comfortable with asking for full proofs. And if the authors don't provide full proofs, and I need to know the authors to determine if the result is believable, isn't that the very definition of bias? 

And finally, from the business meeting....

Cliff Stein did an excellent job running the discussion on this topic, and I want to thank him for facilitating what could have been, but wasn't, a very fraught discussion. He's treading carefully, but forward, and that's great. I was also quite happy to see that in the straw poll, there was significant willingness for trying double blind review (more than the ones opposed). There were still way more abstentions, so I think the community is still thinking through what this might mean.


Sunday, January 07, 2018

Report on double blind reviewing in ALENEX 2018

+Rasmus Pagh and I chaired ALENEX 2018, and we decided to experiment with double blind review  for the conference. What follows is a report that we wrote on our experiences doing this. There are some useful notes about logistics, especially in the context of a theoretically-organized conference on experimental algorithms.

ALENEX 2018 Double Blind Review

For ALENEX 2018, we decided to experiment with a double blind review process i.e one in which authors and reviewers were unaware of each others’ identity. While double blind review is now almost standard in most computer science conferences, it is still relatively uncommon in conferences that focus on theoretical computer science and related topics.

The motivation

In the original argument we presented to the ALENEX Steering Committee, we presented the following reasons for why we wanted double blind review:
1. Eliminating bias.
Andrew Tomkins did an experiment for WSDM this year and wrote a report on it: https://arxiv.org/abs/1702.00502

One particular observation:

"Reviewers in the single-blind condition typically bid for 22% fewer papers, and preferentially bid for papers from top institutions. Once papers were allocated to reviewers, single-blind reviewers were significantly more likely than their double-blind counterparts to recommend for acceptance papers from famous authors and top institutions. The estimated odds multipliers are 1.66 for famous authors and 1.61 and 2.10 for top universities and companies respectively, so the result is tangible”

2. Common practice.

Virtually every CS community except theory is doing double blind review, including most of ML (NIPS, ICML), DB (VLDB, SIGMOD), Systems (NSDI), etc. While theory papers have their own idiosyncrasies, we argued that ALENEX is much closer in spirit and paper structure to more experimental venues like the ones listed.

3. Limited burden on authors and reviewers for an experiment

There was no real logistical burden. We were not blocking people from posting on the arXiv, or talking about their work. We’re merely requiring submissions be blinded (which is easy to do). For reviewers also, this is not a problem - typically you merely declare conflicts based on domains and that takes care of the problem of figuring out who’s conflicted with what paper (but more on this later).

4. Prototyping.

While theoryCS conferences in general do not make use of double blind review, ALENEX is a small but core venue where such an experiment might reveal useful insights about the viability of double blind overall. So we don’t have to advocate changes at SODA/STOC/FOCS straight up without first learning how it might work.

5. PC submissions.

We are allowing PC members to submit papers, and this has been done before at ALENEX. In this case double blind review is important to prevent even the appearance of conflict.

The process

Before submission: We provided a submission template for authors that suppressed author names. We also instructed authors on how to refer to prior work or other citations that might leak author identity - in brief, they were asked to treat these as any third-party reference. We also asked authors to declare conflicts with PC members.
After submission/before reviews: We recognized that authors might not be used to submitting articles in double blind mode and so went over each submission after it was submitted and before we opened up bidding to PC members to make sure that the submissions were properly blinded. In a few cases (less than 10/49) we had to ask authors to make modifications.
During review process: The next step was to handle requests for subreviewers. Since PC members could not determine CoIs (conflicts of interest) on their own, all such requests were processed through the PC chairs. A PC member would give us a list of names and we would pick one. (so more private information retrieval than a zero knowledge protocol!)

Issues

A number of issues came up that appear to be unique to the theory conference review process. We document them here along with suggested mitigations.
  1. Managing the CoI process: In theoryCS conferences, subreviewing happens outside the umbrella of the PC. PC members have the power to request any number of subreviewers for papers, and this process happens after the papers are submitted. In contrast, in other venues, subreviewers essentially function as members of the PC - they are invited to be reviewers ahead of time, and are listed when the author declare conflicts of interest. This means that under the process we used, PC members cannot determine for themselves whether a subreviewer has a CoI with a paper, whereas in the alternate process, this is taken care of automatically. One possible mitigation is to ask PC members to list potential reviewers ahead of time and have them registered in the system for authors to declare CoI with. While this might generate a long list of subreviewers for authors to review, this process is customarily handled by a) allowing authors to declare conflicts by affiliation (domain name) and then b) presenting them with a filtered set of reviewers to mark conflicts with. Domain-based filtering is probably the most effective method for handling conflicts based on current or recent affiliation: it allows for reviewers to be added after the fact, and systems like Microsoft’s CMT allow for it.
  2. The difficulty of hiding identity based on prior work: In experimental work, a group will often write a series of papers that builds on infrastructure that they’ve developed. The relative difficulty of building such infrastructure also means that groups become “known” for working in certain areas. This made it a little difficult for authors to honestly blind their identity, because their papers clearly built on software that wasn’t publicly available and therefore had to be part of their group. The solution of blinding that reference itself does not always work because then it is hard to evaluate the quality of the work described in the paper. We note that this problem occurs in other, more experimental parts of CS. The typical solution is to continue with the blinding effort in any case, and make an effort to release code publicly, so anyone could have used the code being built on. In our view, this is a less significant problem than the first point. To this end, here are some guidelines from CHI and CSCW (both ACM conferences).
  3. Is the paper provably double blind? A common complaint about double blind review is that it is not perfect -- that it’s possible with some effort to determine some subset of the authors with some degree of certainty. The response that we gave when asked, and that is usually given, is that the goal of double blind review is not to provide a zero knowledge protocol, but to prevent the immediate implicit bias that comes from directly seeing the author names prior to reading the paper. We note that this is a common complaint from people in the theory community: however our experience with double blind review in other venues has been that after a while, one gets accustomed to reviewing papers without attempting to first determine the authors and the process works as intended.

Feedback

We also solicited feedback from the program committee after the review process was complete. We asked them three questions:
  1. What did you like (and what worked) about the double blind review process instituted this year for ALENEX?
  2. What in your opinion did NOT work?
  3. Is there any other feedback you'd like to provide about the double blind review process?
The responses we got for question 1 were uniformly of the form of “I’m glad that we did it, and felt that papers got a fairer shake than they would have otherwise”. One PC member even said unequivocally that the double blind review process made it more likely that they would submit to ALENEX in the future.  
On question 2, PC members brought up the issues we raised above, recommending that we make it clearer to authors how they need to blind their submissions and also mentioning the difficulty of assigning subreviewers.  
On question 3, the feedback was uniformly positive in favor of continuing with double blind review, inspite of the issues raised.

Summary

On balance, our experience with double blind review was positive. While there were logistical issues, many of these can be resolved by the methods we describe above. Further, there is now a wealth of knowledge accumulated in other areas of computer science that we can learn from. SIGPLAN has built a comprehensive FAQ around this issue that answers many of the questions that arise.
We recommend continuing with double blind review for at least the next two years at ALENEX, firstly because this brings us in line with common practice in most other parts of computer science, and secondly because many of the logistical issues people face with DB review will go away with habituation. At that point, the potential inconvenience of the process reduces and will be outweighed by the benefits in transparency and lack of bias.


Disqus for The Geomblog