A Community's Perspective on the Status and Future of Peer Review in Software Engineering

06/22/2017 ∙ by Lutz Prechelt, et al. ∙ University of Stuttgart Freie Universität Berlin Technische Universität München 0

Context: Pre-publication peer review of scientific articles is considered a key element of the research process in software engineering, yet it is often perceived as not to work fully well. Objective: We aim at understanding the perceptions of and attitudes towards peer review of authors and reviewers at one of software engineering's most prestigious venues, the International Conference on Software Engineering (ICSE). Method: We invited 932 ICSE 2014/15/16 authors and reviewers to participate in a survey with 10 closed and 9 open questions. Results: We present a multitude of results, such as: Respondents perceive only one third of all reviews to be good, yet one third as useless or misleading; they propose double-blind or zero-blind reviewing regimes for improvement; they would like to see showable proofs of (good) reviewing work be introduced; attitude change trends are weak. Conclusion: The perception of the current state of software engineering peer review is fairly negative. Also, we found hardly any trend that suggests reviewing will improve by itself over time; the community will have to make explicit efforts. Fortunately, our (mostly senior) respondents appear more open for trying different peer reviewing regimes than we had expected.



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

For our purposes, peer review is the practice by which a publication venue sends an article to several expert colleagues (the peers) for review before it is accepted for publication (or not). Although a few venues recently started trying out a different approach (e.g., Faculty of 1000 Ltd. (2017); ScienceOpen (2017)), this basic model of pre-publication peer review is usually considered a cornerstone of quality assurance in the scientific process, in software engineering and beyond (Mulligan et al., 2013).

This article attempts to understand what is currently working well or not-so-well about peer review in software engineering (SE) and how this might change in the next 20 years.

1.1 Variants of Peer Review

The acceptance decision may be made after just one round of reviewing (single-stage peer review111authors may be allowed to comment on the reviews: rebuttals), typical for conferences, or after multiple rounds with improvements of the work (multi-stage review222always with rebuttals), typical for journals.

Usually, the authors do not know the identity of the reviewers (blind review). Reviewers might know the identity of the authors (single-blind review) or not (double-blind review). Only rarely do the authors get to know the names of reviewers (non-blind review, zero-blind review) or does the public get to see the content of the reviews (open review, public review).

1.2 Issues with Peer Review

Informally, researchers often criticize peer review as not doing its job properly and indeed the practice has various inherent problems, for instance:

  • Reviewers will not always be competent to properly review a particular work, and often provide inconsistent reports (Baxt et al., 1998; Smith, 2006).

  • Reviewers will sometimes be biased against certain aspects of the work: methods, technology, goals, etc. (Armstrong, 1997).

  • Reviewers may, protected by their anonymity, abuse their power to inhibit the publication of lines of work that compete with their own (Smith, 2006; Hettyey et al., 2012).

  • Reviewing can be viewed as contributing little to the reviewer’s reputation and so reviewer motivation can be lacking and reviewing be done rather sloppily (Zaharie and Osoian, 2016).

Because of issues like these, other fields (most prominently in the biomedical realm) have long worked to understand the status of peer reviewing and how to improve it (Nature, 2006). For instance, such research has produced strong evidence that double-blind reviewing will lead to results that are less biased than with single-blind reviewing, e.g. Budden et al. (2008), a fact that is now also being picked up in software engineering (Bacchelli and Beller, 2017). But beyond that, software engineering venues are not, so far, particularly prone to experimentation with possible improvements to the peer reviewing regime. In light of the above issues, this might be a pity.

For instance, the high-class health journal The BMJ (acceptance rate 7%) not only performs reviews zero-blind (that is, reviewers sign their reviews), they also publish the reviews along with accepted articles (open reviewing, BMJ (2017)); there is no comparable software engineering venue doing anything as radical.

1.3 Research questions

Our perspective is understanding and then improving the peer review process. We designed our survey along the following research questions. Results and discussion will be structured mostly into one section per research question.

Section 5: What do authors and reviewers perceive to be the purposes of peer review? Which are more important than which others?
Section 6: How well do they perceive peer review to work today (in the sense of producing valid and helpful reviews) and why?
Section 7: How much should reviewers and authors be blinded?
Section 8: Which aspects of reviewing should be public?
Section 9: Should reviewers be compensated for their work? How?
Section 10: What changes to the current reviewing regime should be performed?
Section 11: How might the answers to each of the above questions change in the next few decades?

1.4 Research contribution

Our article makes two research contributions: First, it characterizes the attitudes of mostly senior members of the ICSE333International Conference on Software Engineeringauthors-and-reviewers community with respect to the research questions. Second, it predicts how these attitudes will likely be different for a similar sample of people in the future, several decades away.

1.5 Structure of this article

After reviewing related work (Section 2), we will present our method: The survey population (Section 3.1), the survey instrument (3.2), the execution of the survey (3.2), our data analysis techniques (3.3 and 3.4), and the resulting public data archive (3.5). Then, we discuss the respondent demographics (4) before presenting the results structured according to our list of research questions (Sections 4 to 11). We then discuss our study’s limitations (Section 12) before we conclude (Section 13).

2 Related work

We organize this section along the research questions from Section 1.3. What sets our study apart from other survey work in the area is the use of open questions and qualitative analysis. While we reference various related work, we consider two large scale surveys of peer reviewers attitudes across many disciplines as our baseline background material upon which we frame our study primarily: First Mulligan et al. (2013) with 4037 respondents, second Ross-Hellauer et al. (2017) with 3062. The latter, organized by OpenAIRE, an Open Access collaboration project, is special in that 76% of respondents reported to have participated in open reviewing previously; an unusual population. We found only one reviewing study in the software engineering literature (Bacchelli and Beller, 2017), also a survey.

Purpose of peer review: Weller (2001, p.xii) proposed a concise characterization: “The valid article is accepted, the messy article cleaned up, and the invalid article rejected”. The Mulligan et al. (2013) survey found the main perceived purposes to be (in this order): to improve the quality of published papers; to determine their originality; to select the best possible manuscripts for a journal. Our work will ask the question also beyond predefined answer categories and ask for elaboration.

How well does peer review work today: The Mulligan et al. (2013) survey had 69% of respondents report high or very high satisfaction. When asked what aspects of their articles were improved the most through peer review, respondents mentioned the introduction most (90%) and statistical methods least. Our work will ask about percentages of good, mediocre, or bad reviews received and about specific positive and negative peer review experiences to provide a more detailed picture.

Blinding: Much discussion has happened lately on how much anonymity should be in the peer review process (David Pontille, 2014; Jubb, 2016). Empirical research has found interesting effects from double-blind reviewing. For instance, Budden et al. (2008) found that more articles of female researchers were accepted after the journal Behavioral Ecology adopted double-blind review (but not in other journals that did not). Laband and Piette (1994) found for a sample of economics journals (and controlling for several confounding factors) that articles accepted after single-blind review were cited less often than articles accepted after double-blind review. As for software engineering, Bacchelli and Beller (2017) survey how double-blind peer review is perceived by the ICSE community and find that about half of the respondents believe all software engineering venues should switch to double-blind reviewing. SeeBac17 investigate bibliographic data from 71 of the 80 largest computer science conferences of 2014 and 2015 and find evidence that newcomers (people who have not previously published at that conference) get a smaller share of a conference when single-blind reviewing is used compared to conferences using double-blind reviewing.

The Mulligan et al. (2013) survey respondents did not like the prospect that their names be made visible to the authors (8% more likely to be willing to review under such circumstances, 51% less likely) or to the readers (18% and 45%). In the OpenAIRE survey, 67% of respondents believed zero-blind reviewing would make reviewers less inclinced to provide a review and 44% believed it would improve review quality; 65% believed it makes strong criticism less likely (Ross-Hellauer et al., 2017). Our study will ask for degrees of agreement with double-blinding and zero-blinding.

Publicness: Support for the review reports to be published alongside the accepted paper was low (11% more likely and 58% less likely) in Mulligan et al. (2013). Similar percentages were found for the possibility of disclosing names to authors only (8% and 51%) and for having the reviewer names only published alongside the article (18% and 45%). Even in the OpenAIRE survey, 52% of respondents expect reviewers to become less inclined to review, although 65% expect published reviews to be useful for readers, 60% expect an increase in review quality, and 45% expect authors to become more inclined to submit to such journals.

Some venues such as F1000Research (Faculty of 1000 Ltd., 2017), ScienceOpen (ScienceOpen, 2017), or The BMJ (BMJ, 2017) require public reviews, and initiatives such as Publons (Publons, 2017) or Academic Karma444http://academickarma.org promote them for the rest of the scientific publishing world. Our study will ask for degrees of agreement with publicness of reviews.

Reviewer compensation: Overall scientific publication rates are increasing by 8-9% each year (Bornmann and Mutz, 2015). As a result, there is a reviewer fatigue syndrome (Breuning et al., 2015): reviewers decline review invitations more and more often (Mulligan et al., 2013). Warne (2016), a study specifically on reviewer compensation, reports mean agreement of 4.0 (on a 1-to-5 scale, based on 3000 surveyed researchers) with the statement “I would spend more time reviewing if it was recognised as a measurable research activity”. 51% of the Mulligan et al. (2013) participants would more likely review for a venue that compensated them somehow, only 15% less likely. Our study asks for degrees of agreement and for specific compensation ideas.

Useful reviewing regime changes: About 30% of the Mulligan et al. (2013) respondents believed that the current status of peer review is the best we can have, but the study did not ask the other 70% for improvement suggestions. Several such suggestions come from viewpoint articles. For example, Ralph (2016) recommends for Information Systems research to provide editorial review only for empirical articles and to desk-reject many of those based on checklists. Ferreira et al. (2016) recommends to demand a rate of reviews for each scientist, standardize peer review through training in academic curricula and workshops, and decoupling peer review from journals. Our study asks for any improvement idea, plus elaboration.

Future change: We are not aware of any study that goes beyond reporting current attitudes to explicitly extrapolating them into the future in a data-based manner. Our study will do so based on regression modeling with demographic variables.

We will refer to specific similar or contrasting results of related work as appropriate when we discuss our results.

3 Methods

Our results are based on a mixed quantitative/qualitative survey of software engineering authors and reviewers.

3.1 Survey population

As our base population, we pick the set of all authors and reviewers of recent instances of the International Conference on Software Engineering (ICSE 2014, 2015, and 2016), because it represents software engineering research broadly across most topic ranges and at a high level of quality. We collected the author email addresses from the published articles and the reviewer addresses via the program committee web pages or from lists provided by the program committee chairs. Reviewers include the members of program committee (each year), review committee (2015 only), and program board (2014 and 2016 only).

This results in a set of 966 people. Of these, 642 (66%) have been an author in only one year, 99 (10%) were authors in multiple years. Further, 156 (16%) served as reviewer in one year, and 68 (7%) served as reviewer in multiple years. Of 34 people (3.5%), we could not produce an individual email address (e.g., because no author address was given at all or all authors shared one address), resulting in an actual base population of 932 people.

3.2 Survey instrument and execution

Our questionnaire was built from scratch and had 19 questions. They were a mix of closed or quantitative ones on the one hand and open ones for qualitative analysis on the other. Most closed questions used a 10-point disagree/agree scale, the others are numeric or binary. We will often provide specific wording from the questionnaire along with the presentation of the results. The questionnaire is openly available (see Section 3.5) We sent out an invitation email to the base population in August 2016, stating “We kindly ask you to participate in a small survey on the future of peer review. Your participation, by answering 19 questions that take about 15 minutes of your time, will broaden the understanding of peer reviewing specifically in software engineering: (1) How are current reviewing practices perceived? (2) How could the peer review process be improved?”.

The invitation contained one link to the questionnaire and another by which a recipient could tell us s/he had left software engineering research and would not reply for that reason.

We left the survey open for 14 days (this was mentioned in the email) and sent no reminder. We received 74 bounce messages from email addresses that had meanwhile become invalid; these will mostly belong to junior authors. We received 45 out-of-office autoreplies, 13 of which pointed to an absence of the recipient of more than one week. These 74+13 cases reduce our effective base population to 845.

The “no longer a researcher link” was used by 19 people, reducing it further to 826.

The survey had 241 respondents, giving a 29% response rate. 167 respondents (69% of 241) worked through all pages of the questionnaire. As all questions were optional, each single question has a lower (and varying) number of responses. The time between opening the survey and answering the last question ranged from 4 minutes to 1 day, 22 hours; the first and third quartile were 12 minutes and 27 minutes, respectively. The median completion time was 17 minutes.

3.3 Quantitative data analysis methods

For the quantitative data, we mostly report percentages relative to the respective number of responses and sometimes visualize it with box plots and Likert plots. We use linear modeling for the extrapolation into the future.

3.4 Qualitative data analysis methods

For the open questions, we applied a rough version of open coding Strauss and Corbin (1990, II.5) to derive a reasonable post-hoc classification of the responses so that we can quantify the frequency of the most common types of response. We kept the coding process simple: There was no pre-specified granularity goal or semantic styles goal for the codes, nor did we align codes across different questions. Many codes occurred only rarely, we will therefore not present, define, or even mention all codes.

3.5 Data-and-materials archive

To increase the transparency of our study and its reproducibility, we disclose all the instruments used and the data analyzed in an online open science archive including:

  • a README file,

  • the questionnaire,

  • the raw collected data (including the answers to the open questions),

  • the results of the open coding (annotations and codebooks),

  • the statistical and plotting routines,

  • and the resulting plots.

The archive can be found in the open science repository of the present paper  (Prechelt et al., 2017).

4 Results: Demographics

Compared to the population, our respondents ( for this question, which is treated as 100% for this question) are extremely senior. 52% identified themselves as tenured professors, 15% as non-tenured professors, and 22% as being on the post-doctoral level or “industrial researcher” level.

Of those who provided a gender (, 100%), 15% identified themselves as female, 85% as male.

Of those who provided an age (, 100%), 14% were in their twenties (minimum: 23), 43% in their thirties, 21% in their fourties, 17% in their fifties, and 4% beyond (maximum: 67).

Those who stated their country of affiliation (, 100%), come from 32 different countries, the most common being USA (33%), Germany (12%), and Canada (6%), and all others below 5%.

Respondents said they had published 6.2 peer-reviewed articles in the past twelve months and 1.6 articles at the three ICSEs in question, on average. They had been reviewers at 0.8 of those three ICSEs on average. For the base population, the latter value is 0.3; another indication of our respondents’ strong seniority.

In the results below, we will report on different subgroups where appropriate.

5 Results: Purpose of peer review

We asked respondents how much they agree with each of the nine suggested purposes of peer review shown in Figure 1.

Figure 1: Purposes of peer review. Answers range from strong disagreement (1) to strong agreement (10).

All nine purposes receive more than 50% of replies on the “agree”-side of the scale, six of them even more than 75%. The, by far, most popular answer is to ensure the validity of the research, the core of peer review’s gatekeeping function. The runners-up are to make sure the article is well written, limitations are properly discussed, and the research is relevant. Relatively least popular are detecting plagiarism and protecting the reputation of the venue.

The question was followed by three open text slots to add additional concerns in the form of open answers.

Open coding of the open-ended answers found 13 categories. The top two (each occurring in 17 of the responses) refer to ensuring the novelty of the results and to ensuring scientific progress, respectively. Some of the novelty-related answers stressed specific aspects, such as “[…] not just in the ICSE community but in the broader research community”555Our response quotations are usually verbatim excerpts, except for spelling and punctuation corrections. If we applied changes to wording (for comprehensibility or anonymization), these are indicated by square brackets. or “Ensure that innovative, but possibly incomplete, ideas are injected into the community to stimulate discovery and innovation.” Ensuring progress was characterized in many different ways, from general ones (“assessing contribution to the field”) down to rather specific aspects such as “To ensure that the reporting allows for reproducibility and replicability”.

The third-most frequent code (occurring 13 times) represents checking that articles make proper use of related work, relate themselves to the state of the art, and provide appropriate theoretical framing of their research design.

Most of the other codes (occurring between 10 times and 2 times) echo concerns already represented in the categories of Figure 1, but the respondents added detail or emphasized a sub-aspect. Most popular among those were ensuring “quality” (10 occurrences) or “soundness” (e.g. of method execution or result interpretation, 9 occurrences), improving writing (10 occurrences), and “selecting” among articles (e.g. “grain from chaff”, “top contributions”, or “To balance acceptances across topic areas”, 9 occurrences).

Two of the other codes, however, are new: Learning (from other reviewers or about current research, 5 occurrences) and ensuring impact on SE practice (2 occurrences).

Mulligan et al. (2013) found that the purpose of peer review is, respectively, to improve the quality of published papers (94%); to determine their originality (92%); and to select the best possible manuscripts for a journal (85.5%). We are not aware of studies exploring the purpose of peer review in an open-ended way, as we did. Our results offer a base for future studies on the purpose of peer review.

6 Results: How well does peer-review work?

6.1 Quantitative estimate

It is problematic to ask for the net effect of peer reviewing unless the population consists exclusively of editors or PC chairs. So instead we asked “As an author, what percentage of the reviews that you receive is good, reasonable, unhelpful or grossly faulty.” and elaborated as follows:

  • “By ’good’ we mean a review that is helpful for the acceptance decision and for the authors and that substantiates all of its important points of criticism or praise. It may be quite critical and even propose rejection.”

  • “By ’reasonable’ we mean a review that is ’good’ in some respects, but lacks detail in others.”

  • “By ’unhelpful’ we mean a review that largely or completely lacks substance.”

  • “By ’grossly faulty’ we mean a review that misunderstands or ignores key aspects of the article, leading to wildly exaggerated praise or criticism; this covers only questions of fact, not of opinion or weighting.”

Figure 2: Answer distributions for the frequency of four quality levels of peer reviews. Whiskers show minimum/maximum, the fat dot is the median, the M the arithmetic mean.

The results are shown in Figure 2. On average and roughly speaking, one third of reviews is considered good, one third reasonable, and one third either unhelpful (20%) or grossly faulty (10%). However, there is considerable diversity in the opinions: The most optimistic quarter of respondents believes 50% or more of all reviews are good, while the most pessimistic quarter believes only 21% or less are good. (For comparison, the Mulligan et al. (2013) survey found high or very high review quality satisfaction for 69% of respondents.)

We perceive these responses as balanced (rather than cynical). They do not paint a rosy picture of getting one’s work reviewed in SE: In a typical set of three reviews, one has to expect that only one of them will be as thorough and helpful as they all should be, while the other two are not. As a result, acceptance decisions will be highly noisy.

6.2 Why are the faulty reviews faulty?

We asked “In your opinion, what were the main reasons for unhelpful and/or grossly faulty reviews (if any)?” and received 136 answers. In those, our open coding found 25 different reasons mentioned and 276 mentions (100%) overall.

Two reasons stand out: Reviewers not allocating enough time (24%) and reviewers being insufficiently familiar with the topic of the work (22%). Some of these answers sounded cynical (even sarcastic) or sad, but most were matter-of-fact; we made no attempt to code emotional quality. A few answers captured a lot of their issue succinctly: “lack of time or effort”; “In cases it is simply because the reviewer did not do his/her job, or accepted to referee a paper for which he/she was not qualified. But when you submit to good venues, with good PCs, that happens less frequently.”.

Six other reasons were mentioned at least a dozen times: The reviewer does not care to make a good review (10%), the reviewer is biased towards some type of research content or method (8%), misunderstandings (5%), generally low reviewing skill (5%), inappropriate priorities set by the reviewer (5%), and exaggerated expectations (4%).

So at least one third of mentioned reasons (lack of time and lack of care) ought to be repairable. We are not aware of other studies asking respondents why faulty reviews become faulty.

6.3 Worst peer review experience as author

We asked “As an author, what has been your worst experience with peer review?” and received 123 answers. In those, our open coding found 37 types of experience mentioned and 158 mentions (100%) overall.

The most common topic was a lack of justification in a review: Unjustified rejection (13% of mentions), unjustified individual points of criticism (7%), or a discrepancy between the decision and the review text (6%). The idea of our question was to collect anecdotes and indeed many respondents provided such stories. Some of them included evidence that the issue with the reviews was not merely imagined, like this one: “A paper being rejected with very short reviews that gave no indication as to the reasoning behind the decision. While everyone has a horror story about a rejection, I had a paper that was submitted to a journal and rejected without review by the editor: I submitted the paper to another journal and it was fast-tracked into the next available issue and now has over 300 citations (Google scholar).” Or this one: “I got one paper rejected because it “didn’t even cite [XYZ]”. The reviewer accused us of having no clue about the field and not knowing even the most elementary works in the field. Therefore, he refused to review the paper any further, i.e., the review was just a couple of sentences long. Interestingly, one of the authors of [XYZ] […] was also an author of the paper that got this crappy review. We of course [were] fully aware of [XYZ], but did not find it relevant for what we presented.”

Next in line (and in fact related) are “lazy reviewers” (9%) and overly short reviews (7%). Examples: (1) “The review was one line: You failed to convince me this is an interesting idea.”, (2) “A 10,000 word manuscript fetching a 250 word review, out of which 200 words are spent in summarizing the manuscript”, (3) “Reviewers […] stop reading the paper after the abstract”, (4) “Reviews which were not only misguided in their criticism, but entirely indecipherable due to reviewers’ evident off-the-cuff writing (fragmented sentences, lack of clarity, reference to misspelled terms). This is particularly galling because a.) it is not possible to extract valuable criticism when the reviewer seems not to have read the same paper as you wrote; b.) it is particularly insulting when a rejection is so obviously unconsidered that the reviewer hasn’t read it to themselves.”.

6% of mentions (that is, 7% of respondents) state they never had a particularly bad reviewing experience. On the other hand, 4% report direct insults or criticism addressing the researcher rather than the work (argumentum ad hominem). Examples: (1) “At [XYZ], I have seen a considerable amount of things like name calling. My students have been told they are schizophrenic, directly in peer review, for the bizarre crime of running empirical studies and reporting the data. […] I am consistently amazed at the total lack of accountability in reviewing – even 3rd grade level name calling is, somehow, allowed in a venue like that.”. (2) “a reviewer made personal attacks on one of my co-authors. Fortunately, my co-author took it with good humor, but I felt that it was unacceptable. I reported it to the PC chairs though there is no way to confirm that any word ever made it to the reviewer. I hope it was just a momentary lapse of judgment on the part of the reviewer, but it definitely reduced my opinion of [XYZ] as a venue considering that the PC chairs did not even acknowledge it.”. Further anecdotes revolved around other types of suspected abuses of reviewer anonymity: (3) “A review process for [journal XYZ] where a well-established author on the same topic did not want a new actor around, pretending that everything was already done by him and his research group, which was clearly false.”. (4) “an expert disagreeing because the material presented showed that the results of one of their earlier articles were flawed. It makes no sense to suppress articles that attempt reproduction, certainly not articles that provide a detailed basis why earlier results are flawed. This is also a failure of the committee in general, for not seeing the conflict of interest.”. Reviewer anonymity will be a topic of Section 7.

Four other categories have 6 or 7 mentions (4%) each. They speak about unqualified reviewers, unconstructive criticism, reviewers unjustifiedly pushing their own work, or reviewers severely lacking an understanding for the nature of empirical work. Examples of the latter: (1) “I did not have 100% response [rate in] a survey”, (2) “Why did you not just measure and see what the result is?”, (3) “Desk-rejected qualitative research: There are no numbers!”.

The rest is a long list of rare problems (just 1 or 2 mentions) that includes anything from administrative problems over a decision based on only a single review to receiving a review that was obviously written for a different submission. Several of these relate to various types of pickiness and one of them is particularly worrying: “I, sometimes, feel that if I’m too honest about the limitations of a technique, reviewers will simply pick on it. Of course, if the limitations are too large to render the technique useless, then I agree that it is a big issue. However, what I often find is that if I hadn’t mentioned that limitation in the first place, the reviewers wouldn’t have picked up on it.”

In order to end on a more positive note, we quote this participant: “All this sounds like complaining, I am sure. I accept that the review process is a human process and therefore filled with problems. But I hope to make it as good as possible.”

We are not aware of other studies asking respondents to openly report their worst experiences with peer review; our results therefore complement quantitative results on preconceived problem areas such as those provided by Mulligan et al. (2013).

6.4 Worst peer review experience as reviewer

We also asked the same question from the other perspective: “As a reviewer, what has been your worst experience with peer review?” and received 111 answers. In those, our open coding found 53 different types of experience mentioned and 142 mentions (100%) overall.

The most frequent answer (17 times, 12% of mentions) is that reviewers never had a particularly bad experience, e.g.: “Not too many, typically I’ve met my peer-reviewers, so we all behave pretty civilized.”

Among the rest, four of the types stick out, at 11 to 13 mentions each (8% to 9%; the next-lower one has only 4%). At the top of the list are poor-quality submissions. Here are some variants of that: (1) “Having to read papers which should have been desk-rejected as unreadable.”, (2) “Articles that are so bad and unreadable that they are an insult to the time reviewers voluntarily and freely spend on this process.”, (3) “Not particularly horrible, but once I had to review a paper that was too abstract and general. There was nothing that could be criticized about it, it was a vision paper and the authors were established members of the community. None of the three reviewers had anything to critique about the paper, but it was also clear that there was not much of substance in it. The paper ended up marked borderline/weak accept and eventually being published.”.

Second in line are authors that do not make the necessary improvements to their article, as in these stories: (1) “Spending a lot of time on reviewing a conference paper and discovering conceptual flaws when such flaws are then met with apathy by the authors and other reviewers. It is sad when such flaws are not documented in the final paper version (or explained why they are in fact not flaws). To me this is a big drawback of the conference publication model, since a journal editor can enforce that authors respond to such criticism.”, (2) “In a journal or conference with a revise-resubmit process, I think finding authors who dismiss or ignore feedback is particularly insulting, especially if I’ve spent quite a bit of time to thoroughly read their work and think about my feedback.”, (3) “I don’t know whether [ignoring my improvement request] was because he considered my request unreasonable, because he had lost the raw data, or because [fixing the problem] showed that his results were not very strong. This is particularly bad because it shows that authors can selectively present data that strengthens their claims, and the review process is not strong enough to guard against it”, (4) “Journal A review: I recommended major revision, and really major it would have to be. My review contained about three dozen issues. New version of the article comes in: the two smallest issues have been addressed, none of the important ones are. I state this and reject the article. The editor rejects the article. This was on a Thursday. On the following Monday, journal B queries me for a review. It turns out it is the same article again, in exactly the version rejected by journal A on Thursday.” One of the respondents remarked on a similar story as follows: “Hmm, is double-blind reviewing going to make such behavior more common? That would be horrible.”.

Third is what reviewers perceive as inappropriate behavior (including passivity) of editors, PC chairs or other powers-that-be, for example: (1) “Encouraging PC chairs […] to get papers accepted, because the acceptance is too low.”, (2) “The worst experience was at [XYZ], where several of the decisions PC members came up with after long and careful discussions have been overruled […] without substantial arguments and without asking back.”, (3) “[I rejected an article] that used students as subjects because it was in violation of the basic rules of ethics. The other reviewers accepted the paper on the grounds that the results were good, despite the fact that the non-compliance with ethical norms could have introduced serious threats to validity in the results.”, (4) “I think physical PC meetings reward fast thinking and good communication skills, without analyzing in depth the issues in the paper. I think online discussions work better than PC meetings in this respect. Moreover, PC meetings tend to be too far away from the time when papers are read and reviewed, especially if authors are granted a chance of rebuttal, which extends the reviewing time line.”. On the other hand, remarks elsewhere show that reviewers may have more influence than some of them may believe, like here: (1) “It took a lot of dialog with the editor to sway him/her from applying a simple vote.”, (2) “I decided not to participate as a member of the program committee in the future.”.

Rank 4 belongs to the first of many types that echo all of the issues brought up from the author perspective, just this time the reviewers criticize their co-reviewers (and sometimes themselves). For example: (1) “The other three reviewers wrote bland, nothing-type reviews and were impressed by a lot of statistical mumbo-jumbo which was actually badly flawed. The paper claimed to be a how to do it type article and therefore particularly dangerous. It took a lot of dialog with the editor to sway him/her from applying a simple vote.”, (2) “Papers from a completely unfamiliar area where I could not validate the results or determine their significance.”. We split the bad co-reviewers issues into many types, such as the reviewer being lazy, dogmatic, inflexible or the review being too short, unbalanced, unjustified, too critical, uncritical. Had we collected all of these under a single type, it would have ranked at the top by far with 28 mentions (20%).

Some of the remaining (rare!) issues concern cases of power abuse on the side of reviewers or editors. Examples are: (1) “We once had a reviewer from [country XYZ] on our PC who would give all papers from [country XYZ] the very best grades, even if everybody else did not like the paper. He then did not even show up at the PC meeting. All his reviews were canceled, and we PC members were in for a night shift.”, (2) “Reviewers pushing papers of well known authors (or authors who are their friends) to get the paper accepted.” (3) 

“Seeing other reviewers writing just 1 or 2 line without saying anything on the paper (and probably not reading it), and still fighting to accept/reject the paper.”


However, please remember that the most frequent reply type was the no-major-problems type, e.g. “Not much.”

Similarly to Section 6.3, we are not aware of other studies asking respondents to openly report their worst experiences with peer review as reviewers. Our open-ended exploration offers insights that previous quantitative studies have not provided.

7 Results: How much blinding is appropriate?

Background: ICSE has traditionally used a single-blind reviewing regime; ICSE 2018 is the first to switch to double-blind reviewing. Past ICSEs (to our knowledge, from 1987 to 1991) have required that one of the reviewers be listed at the foot of the title page of accepted articles as having “recommended” the work.

Figure 3: How much blinding is appropriate? Answers range from strong disagreement (1) to strong agreement (10).

We had three agree/disagree items on this topic, which all received answers (100%). The items’ wording and the response percentages are shown in Figure 3

There is a two-thirds majority agreeing that reviewers should not know author names (in practice, this means the “double-blind” regime). In Mulligan et al. (2013), 76% of their cross-discipline participants considered double-blind peer review the most effective method; in Bacchelli and Beller (2017), 46% of their software engineering participants were in favor of all software engineering venues to go double-blind.

As for blinding reviewers with respect to the names of the co-reviewers, sometimes called “triple-blind”, respondents are split half-and-half.

A potentially surprising result arises for the third question, zero-blinding: About one third of respondents say they believe reviewers ought to give up their anonymity and sign their reviews. The OpenAIRE survey had not asked a “should” question, but even for its very openness-minded participants only 44% had agreed zero-blinding would improve review quality (Ross-Hellauer et al., 2017).

8 Results: Should drafts and reviews be laid open?

Figure 4: Should drafts and reviews be laid open?

We had three agree/disagree items on this topic, which also all received answers (100%). Their wording and the response percentages are shown in Figure 4.

Respondents are split half-and-half about whether reviews should be published along with an accepted article. There is also some limited support for the more radical ideas of publishing article draft and reviews also for rejected articles (31%), or even publishing the article draft immediately upon submission (28%). These sentiments are more positive than those found by Mulligan et al. (2013) and only moderately less enthusiastic than those in the OpenAIRE survey (Ross-Hellauer et al., 2017).

9 Results: Should reviewers be compensated?

9.1 Monetary or quasi-monetary compensation

We asked “Reviewers should receive a (quasi-)monetary compensation for their work (e.g. memberships, subscriptions, registration discounts, or money payment). If so, which?”. 41% of respondents agreed (to varying degrees) and 33% provided a free-text comment on the issue. In those, our open coding found 14 different types of suggestion mentioned and 142 mentions (100%) overall. 16% of those suggest monetary compensation, nearly all of the others suggest variants of the other ideas mentioned in the question, the most popular being conference registration discounts (43%) and waivers on society memberships (7%) or subscriptions (10%).

These results reflect a much more honor-based attitude towards reviewing than those from Mulligan et al. (2013), where 41% of participants showed inclination towards monetary and 51% towards quasi-monetary compensations.

In contrast, a few of our respondents even made critical remarks on the for-profit culture in much of the scientific publishing system, like this one: “It is ridiculous we are doing free work that will ultimately result in more money for Elsevier/IEEE. Willing to debate if money should go to person or institution”.

9.2 Showable proof of good work

We also asked “Reviewers should receive showable proof for good reviewing work (e.g. public visibility of their review texts, or a reviewing quality certificate). If so, which?”. 71% of again respondents agreed and 51% provided a free-text comment on the issue. As for the monetary compensation question, these comments were heavily primed by the examples given in the question, but contain a number of further ideas as well. In the 81 comments, our open coding found 31 different types of suggestion mentioned and 116 mentions (100%) overall.

The most common suggestion was indeed handing out a certificate (suggested by 51% of mentions). Some of these were more specific, for example they proposed that only the best reviewers should get a certificate (16%) or the certificate should state the quality of the reviews (11%). Various ideas amounting to other forms of transparency or a reputation system, when taken together, represent another 27% of mentions, so that these two categories sum to 78% of mentions, making everything else minor. Interesting specialized points made by only one or two respondents in these two or other categories include: publish certificates centrally, use Publons, mimic StackOverflow, don’t forget the subreviewers, blacklist bad reviewers.

The only Mulligan et al. (2013) equivalent is “Acknowledgment in the journal”, which 40% of respondents found attractive. The OpenAIRE survey does not report on this issue. Warne (2016) is specifically about compensation and reports mean agreement of 4.2 (on a 1-to-5 scale) with “Reviewing should be acknowledged as a measurable research output”.

Summing up, there is a lot of support for issuing some kind of reviewing certificate to reviewers and some support for various forms of quasi-monetary compensation. Overall, our respondents are far more welcoming to showable proof (71% agreement) than to monetary or quasi-monetary compensation (41% agreement).

Three initiatives are already pursuing goals of the “showable proof” type on a general level: Publons666http://publons.com (for journals only) counts reviews and also allows publishing them, Academic Karma777http://academickarma.org aims at making the content of all reviews and review responses public and allows signing reviews, Review Quality Collector888http://reviewqualitycollector.org (RQC, currently for conferences only, later also for journals) issues certificates based on an explicit quantitative review quality assessment.

10 Results: How should the reviewing regime change?

At the end of our survey, we asked our respondents “If you could change the current review practices at will, what would you consider the most valuable improvements (and why)?”. This was a free-text question only (no predefined categories at all) and it received 118 responses. In those, our (somewhat over-eager) open coding found an enormous 57 different types of change suggestion mentioned with 162 mentions (100%) overall.

Among these, two stick out by a far margin (with 17% and 15% of mentions, respectively): introduce double-blind reviewing and introduce open reviewing. Open reviewing in this sense is a combination of publishing the review (and perhaps the draft submission) and attaching the reviewer’s name to it, but the respondents provided very different amounts of detail in their description so not all of them may actually have meant all of these elements. One could actually use both suggestions in one process: prepare the initial review under a double-blind regime, perhaps even have a discussion with the authors still in double-blind mode, and then lift anonymity on both sides and publish the reviews (for accepted or all submissions) and perhaps the article drafts as well. 5 people indeed mentioned both together, which is logical if one subscribes to “The most important thing is to have a symmetrical reviewing process (i.e., either blinded or unblinded).”. Most proponents of one of these ideas, however, do not favor the other, with attitudes like this one for the open reviewing camp: “So many [reviews] are so insightful, everyone should be able to learn from their critiques!” or this one against:

“Making reviews more visible opens up a can of worms that will, ultimately, not be helpful. People are vindictive, if you haven’t noticed. Even the best researchers can have their moments.”

and this one for the double-blind camp: “There is evidence of bias in scientific reviewing and evidence that double-blind reviewing can reduce it. Experience with light double blind reviewing in related fields suggests that it is successful, low-cost, and has few drawbacks.” or this one wary of it: “Double blind has good justifications, but it is unfortunate it also lowers ability of established researchers to push the envelope: the low accept rate makes blinded highly novel papers have low chance of acceptance without a level of experimental evidence […]”.

Following these top suggestions are three with 4% of mentions each: reward reviewers, rate reviewers, decrease reviewing load. Also with 4% of mentions comes a category ’novel process’ with sketches of radical ideas such as this one: “Reviewers rank all papers. Authors decide whether or not they accept to present their work. If and only if they present their work then their rank is published with a review summary. Author presentation time is proportional to their rank position.”

The long tail contains a number of straightforward suggestions such as reducing reviewing load, avoiding sub-reviewers, or introducing rebuttals as well as a few more far-reaching suggestions we found remarkable: (1) Elect (rather than appoint) program committees. (2) “Build some sort of reviewer rating system. Reward good reviewers and warn and eventually punish bad reviewers. Build a culture of valuing good strong reviews.”, which becomes most interesting in combination with “People who publish (including co-authors) should be required to review a comparable amount.”. (3) Get rid of publishing papers at conferences: “In Computer Science, [we should move] away from a model in which excellent research is routinely rejected only because there was another paper at the same conference that the PC considered to be of equal quality but higher excitement level”. (4) Physical PC meetings are an important practice for some (“there are several best practices that help maintaining high standards, most notably physical PC meetings”) and a threat to good peer review for others (“they are a drain on everyone and on the environment, and they tend to favor outcomes advocated by strongly vocal members.”.

We are not aware of similar results from other studies.

11 Results: How these results may change in the future

The responses to our questions regarding blinding, publishing reviews and drafts, or compensating reviewers represent attitudes. How will these attitudes change in the future? We asked our respondents for their age, so we can (and now will) look for age-related trends in our data.

11.1 Theoretical assumptions, approach

If we set aside the case of disruptive changes and look only at trends already represented in our dataset, we see two possibilities:

  • Hypothesis G: There is a generational trend; the attitude of a person is largely stable. If only G were true, the same attitude of our now-younger respondents today would largely be the attitude of then-older respondents in twenty years.

  • Hypothesis S: There is a seniority trend; the attitude of a person changes with experience and/or role. If only S were true, the attitude of our now-older respondents today would likely be the attitude of then-older respondents in twenty years as well.

Obviously, we should expect a mix of both effects. But is one of them dominant? Our data cannot provide a definite answer, but can provide a strong clue, because we have a good proxy for seniority, experience, and role in our data: The current professional position a respondent holds.

We will build linear models of attitudes using age and seniority as predictors and see whether one or both are statistically significant and how large their coefficient (i.e., the respective effect) is. Where only age is significant, this indicates G is dominant. Where only seniority is significant, this indicates S is dominant. Where both are significant, this indicates both effects mix. Where none of them is significant, this indicates time trends are weak or non-existing.

There is a problem: age and seniority correlate strongly. Therefore, we should not expect the linear models to be highly stable999

The instability induced by predictor collinearity is commonly measured by the variance inflation factor (VIF). Values under 4 are generally considered totally unproblematic

(OBrien07). The VIFs in our models range vom 1.39 to 1.82. Therefore, we will only be able to say which of G or S is dominant if the difference between the two effects is large.

11.2 Predictor variables

We will use the following predictor variables in the models:

  • ageD: age in decades. We use decades rather than years to make the coefficients larger and easier to read.

  • prof: whether or not the respondent is a tenured or non-tenured professor; this is a proxy of seniority.

  • tenured: whether or not the respondent is a tenured professor; this is an alternative proxy of seniority.

Each model will have age as a predictor plus zero or one of the seniority measures, plus possibly the interaction of age and seniority. For the latter, we will use the notation of R in the coefficient table, e.g. ageD:profFALSE and ageD:profTRUE. prof and tenured, being only binary, tend to have less predictive power, giving the G effect a head start, which we need to keep in mind for the discussion.

11.3 Dependent variables

We try each of the following dependent variables in the models:

  • Pgood: the percentage of “good” reviews.

  • Punhelpful: the percentage of “unhelpful” reviews. (The “reasonable” ones appear less interesting.)

  • Pfaulty: the percentage of “grossly faulty” reviews.

  • AknowR: whether authors should know who their reviewers are. This, as all of the other attitude variables below, is measured on the 10-point disagree/agree scale which we always interpret as a difference scale here and represent it by evenly-spaced fractional numbers in range .

  • RknowA: whether reviewers should know who their authors are.

  • RknowR: whether reviewers should know who their co-reviewers are.

  • openness: the average of the above three.

  • opennessChg: ditto, but with the sign of the latter two components reversed. This represents the attitude towards change relative to the single-blind regime that was most common in software engineering (in particular: used at ICSE) in the timeframe we asked about.

  • pubreviews: whether or not reviews should be published along with accepted articles.

  • publicness: the average of all three publicness-related questions we asked.101010We do not use the other two separately because they were formulated in a manner that makes their solo interpretation ambiguous.

  • monetary: whether or not reviewers should receive monetary compensation for their work.

  • certificate: whether or not reviewers should receive “showable proof” of good reviewing work.

11.4 Model selection method

For each of the 12 dependent variables, we will consider five different models as follows (60 different models overall) and present only the most convincing one from each block – or none, if none is convincing at all. A convincing model needs all coefficients to have statistical significance and a high . Each candidate model has theoretical plausibility, so we do not consider this procedure to constitute “fishing for significance”. Nevertheless, we will use a low significance threshold of for each coefficient to reduce false positives.

We will show this procedure by spelling the process out for one of the dependent variables; we will only present end results for the remainder. Consider Table 1: Each block of rows represents one model, numbered in the leftmost column. The second column describes the predictors used in the model: age and prof separately (A+P); age-and-prof interaction (A:P); ditto for tenured (A+T; A:T); or age only (A). The third and fourth column show the coefficients in the model; the fifth the corresponding value; the final column shows adjusted

for the model: The fraction of variance explained after deducting the random-chance component for each degree of freedom used by the model.

type name coeff.
36 A+P (intercept) 2.48 0.001 0.070
ageD -0.60 0.003
profTRUE 0.03 0.950
37 A:P (intercept) 2.50 0.003 0.070
ageD:profFALSE -0.60 0.018
ageD:profTRUE -0.60 0.002
38 A+T (intercept) 2.47 0.002 0.070
ageD -0.59 0.009
tenuredTRUE 0.01 0.981
39 A:T (intercept) 2.61 0.006 0.070
ageD:tenuredFALSE -0.64 0.024
ageD:tenuredTRUE -0.61 0.003
40 A (intercept) 2.46 0.000 0.077
ageD -0.59 0.001
Table 1: All five candidate models for dependent variable opennessChg. Our model selection rules suggest to use model 40 for this variable. Columns and model selection are explained in the main text.

Model 36 (this will be the model number in our overall models list) in Table 1 tells us that agreement with reviewing regime change (opennessChg) is 2.48, halfway between total agreement and a neutral stance, if the respondent is a baby (0 decades old) and not a professor (profTRUE is 0). By the age of 20, agreement will have fallen to 1.28 and by the age of 50 to -0.52. However, the seniority effect (profTRUE) has has a non-significant coefficient, so the whole model is not meaningful. Model 38, which replaces prof by tenured, is very similar; both cannot be used.

Model 37 uses the interaction of ageD and prof instead of using them side-by-side; all three coefficients are significant, so this could be a useful model. However, both coefficients of the interaction are practically the same, so this is not a meaningful model either. Its cousin, model 39, behaves similarly. The coefficients are a bit more different, but the first interaction coefficient is no longer significant, so this model also cannot be used.

Model 40, the simplest of them all, using only age for prediction, is the only convincing model and hence the one we select. Much like model 36, it says agreement is at 1.28 for the average 20-year-old respondent and -0.49 for a 50-year-old.

11.5 Results

type name coeff.
25 A:P (intercept) -3.93 0.000 0.028
RknowA ageD 0.61 (0.026)
30 A (intercept) -3.05 0.011 0.041
RknowR ageD 0.76 0.009
40 A (intercept) 2.46 0.000 0.077
opennessChg ageD -0.59 0.001
52 A:P (intercept) 0.99 0.000 0.050
monetary ageD:profFALSE -0.17 0.007
ageD:profTRUE -0.14 0.003
57 A:P (intercept) 1.09 0.000 0.015
certificate ageD:profFALSE -0.11 (0.046)
ageD:profTRUE -0.07 (0.072)
Table 2: Best model for each dependent variable where a convincing model was found. The interpretation is explained in the main text.

We discuss all dependent variables in order. Where a convincing (or semi-convincing) model was found, it is shown in Table 2.

Models 1–15. Neither of the variables Pgood, Punhelpful, Pfaulty has any model at all with all-significant coefficients, so models 1 to 15 are all missing from the table. This tells us that the perception of review quality appears to be a timeless phenomenon.111111

More precisely: Age and seniority effects are too weak to show up in a dataset of the size we have. This also suggests the estimates are largely unbiased.

We will use this term, timeless, for similar cases of no-good-model-at-all below.

Models 16–30. AknowR, the practice of reviewers signing their reviews, is timeless: Our respondents are all similarly skeptical. Strictly speaking, RknowA is timeless as well, but it has one model, 25, close enough to significance that we include it here for information: Older respondents appear much less adamant that authors should be hidden from reviewers. RknowR model 30 shows a relatively strong age effect: Younger respondents tend to prefer hiding reviewers’ names from each other, ones over the age of 40 no longer think so.

Models 31–35. AknowR, RknowA, RknowR all represent a form of openness (transparency) in the reviewing process, so averaging them describes attitudes towards openness in general. But none of the models 31 to 35 is convincing; the openness attitude as a whole is timeless.

Models 36–40. Given that it is currently the norm in software engineering (at conferences) that reviewers know authors and each other, we can reverse the sign of variables RknowA and RknowR and compute an “inclination towards change from the current openness regime”. This is the variable we have discussed in the example in Section 11.4: Young respondents are inclined to change, older ones much less so (model 40).

Models 41–50. How about some other form of transparency: publishing the reviews? Both variables, pubreviews and publicness, are completely timeless: None of the models can explain more than 0.4% of the variance.

Models 51–60. Finally, there is the issue of compensating reviewers. Model 52 is the first two-variable model that comes out as most convincing. It states that young respondents tend to think (if only with a rather weak majority) that reviewers should be compensated (quasi)monetarily. Older ones believe this less, reaching the zero point at the age of 59 if they are not professors and 73 if they are.121212The more exact coefficients are 0.9947, 0.1674, and 0.1361. The corresponding model 57 for non-monetary compensation behaves in the same manner, except its coefficients are not fully significant and the age effect becomes so weak that only 100-year-olds131313Professors even need to wait until they are 153. stop believing-at-least-a-bit that issuing certificates to reviewers would be worthwhile. We include this model for comparison.

11.6 Interpretation

Overall, there is not a single A+P model with a significant profTRUE coefficient and not a single A+T model with a significant tenuredTRUE coefficient. This tells us that seniority effects, if they exist at all, cannot be strong. Therefore, the age effects found are likely mostly generation effects (hypothesis G), not seniority effects (hypothesis S).

Summing the results up, we see that while there are generational effects here and there, they are not very strong. We should not expect reviewing to change drastically just because the now-younger generation will advance to positions of power. Explicit change initiatives will likely be required instead.

12 Limitations

Despite our relatively good response rate of 29%, our sample is obviously not representative of the base population, as is easy to see from the demographics in Section 4. By assuming age and status distributions for our base population we could in principle correct for this distortion, but we consider this too unreliable and so do not do it. So the study is limited in that our characterization of what population it represents remains imprecise.

As with any survey, the truthfulness and well-reflectiveness of the answers is not certain, but we saw no signs of distorted responses and our base population can be considered as a serious one. Therefore, we expect this problem to be negligible. The same is true for accidentally wrong inputs.

The evaluation, both the statistical and the qualitative one, is mostly straightforward, so we do not expect grave mistakes to have gotten seriously in the way of the correctness of our reported results. In any case, other researchers can check based on our fully disclosed data (see Section  3.5).

The strongest threat to validity concerns Section 11: Our trend analysis requires the assumption that attitudes such as those surveyed here tend to be stable. There is evidence that this is the case (Schwarz and Bohner, 2001), but it still remains an assumption. Fortunately, no strong conclusions arise from that analysis and need to rest on that assumption, so the actual threat to validity is small.

13 Conclusions

Our survey of perceptions of and attitudes towards contemporary peer review in software engineering research among a rather senior subset of the ICSE 2014, 2015, 2016 reviewers and authors brought the following major findings:

  1. The respondents agree with a multitude of purposes that could be ascribed to peer review. The strongest agreement (at 96%) is with the purpose of ensuring the validity of the research in question (Section 5).

  2. The respondents are skeptical regarding the quality that software engineering reviews typically have today: On average, they deem only one third of all reviews they receive to be of good quality, while another third is either useless or grossly faulty (Section 6.1). If these perceptions are correct, software engineering reviewing is severely broken and a lot of time, nerves, and goodwill of all involved get wasted. Given the key role of peer review in the scientific process, we should make efforts to improve this situation.

  3. When asked for the reasons why the grossly faulty reviews are faulty, the respondents offered several dozen possibilities. The top three of these, however, cover half of the answers: Reviewers not investing enough time (24% of mentions), reviewers not knowing enough about the subject area (22%), and reviewers not caring to prepare a good review (10%) (Section 6.2). None of the three is insurmountable: Time investment is a matter of priorities, lack of expertise can be accommodated by not taking on the review in the first place, and a lack of care stems from a modifiable (if not easily) attitude. Improvement efforts can be successful in principle.

  4. We asked respondents for their worst peer review experience from the author perspective, which sheds some light on how reviews are broken when they are considered to be very broken. The top few categories account for 42% of mentions: A lack of justification of something important (26%), overly short reviews (7%) or some other form of apparent reviewer laziness (9%) (Section 6.3). Even if many reviewers indeed lack the skill to notice which statements in their reviews require justification, this is something that could be taught and trained; the other problems are even more straightforward to avoid – again: in principle; if one wants to.

  5. Our question on the worst peer review experience from the reviewer perspective shows that authors and editors or PC chairs also have their share of responsibility for the bad state of peer review, but the manners and reasons are more varied here (Section 6.4).

  6. As for blinding, two thirds of respondents agreed author names should be hidden from reviewers (double-blind reviewing), half agreed co-reviewer names should be hidden, and one third agreed reviewer names should not be hidden from authors (zero-blind reviewing). So, although few software engineering reviewers currently appear to sign their reviews, not so few appear to be willing to go the route opposite to the current trend towards double-blind reviewing and go fully transparent instead (Section 7).

  7. Half of the respondents also agree that review texts should be published along accepted articles (Section 8).

  8. 41% of our respondents agreed that reviewers should receive monetary or quasi-monetary compensation for their work, the latter being preferred (Section 9.1).

  9. 71% agreed that reviewers should receive showable proof of good work as a compensation. Half of the specific suggestions in this regard amount to some kind of certificate (Section 9.2).

  10. When asked how they would change the current reviewing regime if they could, our respondents produced a broad set of suggestions. Two of them were more popular than the rest: Introducing double-blind reviewing, covering 17% of mentions, or introducing open (zero-blind and published) reviewing at 15%. (Section 10).

  11. We investigated how many of the above estimates and attitudes depend on age and/or seniority (i.e. experience and role) and found almost no seniority effects and only a few weak or modest age effects (Section 11). This means one should not expect the reviewing regime to change quickly just because the current senior researcher generation retires; an explicit effort to change the attitude of the software engineering community as a whole will likely be required.

Existing (non-SE) venues such as F1000Research, ScienceOpen, or The BMJ show that “radical” solutions like non-anonymous, public reviewing are possible; many of the current issues with peer review quality likely shrink to modest proportions under such conditions. Furthermore, initiatives such as Publons and Review Quality Collector provide ideas how even anonymous regimes can be improved. How about some experimentation?


We thank our survey respondents for their sometimes quite elaborate, yet helpful answers. We thank all reviewers involved indirectly for participating in the respective reviewing processes. We also thank the anoymous reviewers of the present article, in particular the one who semi-identified him- or herself as a member of the ICSE steering committee. Daniel Graziotin has been supported by the Alexander von Humboldt (AvH) Foundation.


  • Armstrong (1997) Armstrong, J.S., 1997. Peer review for journals: Evidence on quality control, fairness, and innovation. Science and Engineering Ethics 3, 63–84. doi:10.1007/s11948-997-0017-3.
  • Bacchelli and Beller (2017) Bacchelli, A., Beller, M., 2017. Double-blind review in software engineering venues: The community’s perspective, in: Proceedings of the 39th International Conference on Software Engineering Companion, IEEE Press, Piscataway, NJ, USA. pp. 385–396. doi:10.1109/ICSE-C.2017.49.
  • Baxt et al. (1998) Baxt, W.G., Waeckerle, J.F., Berlin, J.A., Callaham, M.L., 1998. Who reviews the reviewers? feasibility of using a fictitious manuscript to evaluate peer reviewer performance. Annals of Emergency Medicine 32, 310–317. doi:10.1016/S0196-0644(98)70006-X.
  • BMJ (2017) BMJ, T., 2017. Open peer review. http://www.bmj.com/about-bmj/resources-reviewers. As of 2017-05-02.
  • Bornmann and Mutz (2015) Bornmann, L., Mutz, R., 2015. Growth rates of modern science: A bibliometric analysis based on the number of publications and cited references: Growth rates of modern science: A bibliometric analysis based on the number of publications and cited references. J Assn Inf Sci Tec 66, 2215–2222. doi:10.1002/asi.23329.
  • Breuning et al. (2015) Breuning, M., Backstrom, J., Brannon, J., Gross, B.I., Widmeier, M., 2015. Reviewer fatigue? why scholars decline to review their peers’ work. APSC 48, 595–600. doi:10.1017/S1049096515000827.
  • Budden et al. (2008) Budden, A.E., Tregenza, T., Aarssen, L.W., Koricheva, J., Leimu, R., Lortie, C.J., 2008. Double-blind review favours increased representation of female authors. Trends in Ecology & Evolution 23, 4–6. doi:10.1016/j.tree.2007.07.008.
  • David Pontille (2014) David Pontille, D.T., 2014. The blind shall see! the question of anonymity in journal peer review. Ada: A Journal of Gender, New Media, and Technology 4. doi:10.7264/N3542KVW.
  • Faculty of 1000 Ltd. (2017) Faculty of 1000 Ltd., 2017. F1000research. https://f1000research.com/about. As of 2017-05-02.
  • Ferreira et al. (2016) Ferreira, C., Bastille-Rousseau, G., Bennett, A., Ellington, E., Terwissen, C., Austin, C., Borlestean, A., Boudreau, M., Chan, K., Forsythe, A., Hossie, T., Landolt, K., Longhi, J., Otis, J., Peers, M., Rae, J., Seguin, J., Watt, C., Wehtje, M., Murray, D., 2016. The evolution of peer review as a basis for scientific publication: directional selection towards a robust discipline. Biol Rev Camb Philos Soc 91, 597–610. doi:10.1111/brv.12185.
  • Hettyey et al. (2012) Hettyey, A., Griggio, M., Mann, M., Raveh, S., Schaedelin, F., Thonhauser, K., Thoss, M., van Dongen, W., White, J., Zala, S., Penn, D., 2012. Peerage of science: will it work? Trends Ecol Evol 27, 189–190. doi:10.1016/j.tree.2012.01.005.
  • Jubb (2016) Jubb, M., 2016. Peer review: The current landscape and future trends: Peer review landscape. Learned Publishing 29, 13–21. doi:10.1002/leap.1008.
  • Laband and Piette (1994) Laband, D.N., Piette, M.J., 1994. A citation analysis of the impact of blinded peer review. JAMA 272, 147–149. doi:10.1001/jama.1994.03520020073020.
  • Mulligan et al. (2013) Mulligan, A., Hall, L., Raphael, E., 2013. Peer review in a changing world: An international study measuring the attitudes of researchers. Journal of the American Society for Information Science and Technology 64, 132–161. URL: http://dx.doi.org/10.1002/asi.22798, doi:10.1002/asi.22798.
  • Nature (2006) Nature, 2006. Nature’s peer review debate. http://www.nature.com/nature/peerreview/debate/index.html.
  • Prechelt et al. (2017) Prechelt, L., Graziotin, D., Méndéz, D., 2017. Open science repository for ’on the status and future of peer review in software engineering’. figshare URL: https://doi.org/10.6084/m9.figshare.5086357, doi:10.6084/m9.figshare.5086357.
  • Publons (2017) Publons, 2017. Recognising peer review. speeding up research. https://publons.com. As of 2017-05-02.
  • Ralph (2016) Ralph, P., 2016. Practical suggestions for improving scholarly peer review quality and reducing cycle times. Communications of the Association for Information Systems 38, 13. URL: http://aisel.aisnet.org/cais/vol38/iss1/13.
  • Ross-Hellauer et al. (2017) Ross-Hellauer, T., Deppe, A., Schmidt, B., 2017. http://doi.org/10.5281/zenodo.570864. Zenodo URL: http://doi.org/10.5281/zenodo.570864. preprint.
  • Schwarz and Bohner (2001) Schwarz, N., Bohner, G., 2001. The construction of attitudes, in: Tesser, A., Schwarz, N. (Eds.), Blackwell Handbook of Social Psychology: Intraindividual Processes. Blackwell Publishers Inc., Malden, Massachusetts, USA, pp. 436–457. doi:10.1002/9780470998519.ch20.
  • ScienceOpen (2017) ScienceOpen, I., 2017. Peer review guidelines. ScienceOpen URL: http://about.scienceopen.com/peer-review-guidelines/.
  • Smith (2006) Smith, R., 2006. Peer review: a flawed process at the heart of science and journals. J R Soc Med 99, 178–182. doi:10.1258/jrsm.99.4.178.
  • Strauss and Corbin (1990) Strauss, A.L., Corbin, J.M., 1990. Basics of Qualitative Research: Grounded Theory Procedures and Techniques. SAGE.
  • Warne (2016) Warne, V., 2016. Rewarding reviewers–sense or sensibility? a wiley study explained. Learned Publishing 29, 41–50. doi:10.1002/leap.1002.
  • Weller (2001) Weller, A.C., 2001. Editorial Peer Review: Its Strengths and Weaknesses. Information Today Inc.
  • Zaharie and Osoian (2016) Zaharie, M.A., Osoian, C.L., 2016. Peer review motivation frames: A qualitative approach. European Management Journal 34, 69–79. doi:10.1016/j.emj.2015.12.004.