From cases to principles: reflections on ethics in forensic peer review
2,743 words · 14 min read
Last week I attended a workshop on Ethical Issues in Forensic Peer Review on Lake Como alongside a range of other folks from within this space including self-identified sleuths, some who have bemoaned the vigilantism of these sleuths (or at least how they sometimes conduct themselves), representatives from publishers who handle research integrity, and those who had been on the receiving end of scrutiny as authors1. I had previously participated in a qualitative semi-structured interview component of the broader project that this workshop formed and was invited by the organisers because of my experiences with error detection and possible research fraud. The most prominent of these was the investigation into the work of Matheus Barbalho and colleagues, which resulted in a white paper documenting a series of improbable data patterns and a more accessible explanation of how the investigation developed. The subsequent process was long, involved several journals and institutions, and led to multiple retractions.2 You can also see some previous reflections on this process in a Q&A session I did for STORK.
These cases provided the immediate background to my invitation, but one of my main reflections from the workshop was that it is very easy to become absorbed in the particulars of an individual case. The details can be fascinating, troubling, and sometimes almost unbelievable. Yet the more useful question is usually not simply what happened in this case? It is what does this case illustrate about the scientific system more generally?
From a case to a class of problems
A case can expose vulnerabilities that are otherwise difficult to see. It can reveal how much trust is placed in co-authors, how poorly journals and institutions sometimes communicate, how unclear the routes for raising concerns can be, or how much unpaid labour is required to correct the record. A case may also show that the people involved do not occupy simple moral roles. The same person may be a whistleblower in one setting, a co-author who missed something in another, a reviewer trying to make a fair judgement, or a journal or institution deciding on what the consequences should be in a particular situation.
But a vivid, and often emotionally charged, case is not automatically representative. It may help us generate questions or identify mechanisms without telling us how prevalent a problem is. Moving from a particular case to claims about error and fraud in “science”, a specific discipline, or researchers in general, whether in terms of prevalence, severity, or even what the “right” or “wrong” way of confronting this is therefore requires additional inferential steps. This distinction became important in our discussion of the relationship between research sleuthing and metascientific questions of error and fraud more generally.
Sleuthing, metascience, and the target of inference
One suggested distinction was that research sleuthing resembles detective work, whereas metascience looks more like traditional scientific research. I am not convinced that this marks a deep methodological divide. Hans Reichenbach’s famous distinction between the context of discovery and the context of justification came to mind and seemed to me to provide a good lens through which to view this.
Taking the context of discovery for example… The motivations behind “sleuthing” and why particular errors are discovered and examined came up as a key point of contention between participants. A sleuth might examine a particular body of work because of a tip-off, an inconsistency noticed while reading, unusual results, or background knowledge that makes a claim seem implausible or that a particular lab or author(s) have had errors in their work previously. Sometimes it may even be motivated by intellectual or personal rivalry. A metascientist interested in prevalence of such errors may instead define a population of papers and select a random or otherwise systematic sample. The route by which the object of inquiry, errors or possible fraud, is selected is importantly different.
In the context of justification, however, both activities can and do draw on the same broad resources to corroborate or falsify the claims being made: logic, mathematics, statistics, and relevant background knowledge. I gave the example of Sherlock Holmes during the workshop to try to highlight this similarity. Holmes as detective applies the tools of logic, often abductive and deductive reasoning, to a unique case where he develops and tests theories about it’s particulars. I imagine if he were asked whether the way he goes about his investigations is “scientific” or not he’d respond in the affirmative. The difference between so called “detective” work and “science” is often less about the inferential tools employed than about the target and scale of the inference.
A sleuth may ask whether the data or claims in one paper, one dataset, or one connected body of work can be reconciled with the stated methods and relevant background knowledge. A metascientific study may ask how frequently a feature occurs across a defined literature, how it varies between fields, which incentives predict it, or even the sensitivity and specificity of particular detection methods themselves3. A purposively selected case by whatever means it is selected can be highly informative about the former almost by definition, and almost silent about the latter. Random sampling matters when the intended inference is to a population; it is not a requirement for learning anything at all from a particular case.
This suggests that “sleuthing” and “metascience” need not name rival kinds of inquiry. They can instead name inquiries with different selection processes and different inferential targets. Both still have to make their reasoning public, expose it to criticism, distinguish anomaly from explanation, and calibrate conclusions to the evidence available.
But, there may also be a third context worth separating from discovery and justification which arises from the social nature of the scientific enterprise: we could call it a context of action. Once we have discovered something that may be an error and have justified grounds for concern, what should we do? Who should be contacted, in what order, and with what level of confidentiality? When should uncertainty be made public? What would count as correction, repair, or a proportionate outcome? These are not settled by the discovery, nor the justification that an error is indeed an error, alone.
Ethics before guidelines
Much of the workshop understandably concerned applied questions: how both sleuths and authors should conduct themselves, how journals should respond, where responsibilities lie, and what boundaries should constrain public criticism. Yet before laying down guidelines or rules, it seems necessary to ask what ethical framework, or combination of frameworks, these are even meant to express.
Virtue ethics focuses attention towards the kind of person or community we should try to be prioritising character, habits, and dispositions as what “good”. In this setting, relevant virtues might include honesty, courage, humility, fairness, patience, and a willingness to correct oneself. A deontological approach instead focuses upon the duties or obligations we should have and imposes moral rules as to what is “good”. For example, this might include a duty to report serious concerns, to represent evidence accurately, to protect confidentiality, to disclose conflicts, to offer a right of reply, or not to make claims that exceed the evidence and criticise only the work and not the person. Consequentialism considers the outcomes of different practices as defining the “good”. We might consider outcomes such as correction of an unreliable literature, harms caused by false or disproportionate allegations, effects on that hold back or further promote criticism, reputational damage, impacts on public trust, and the opportunity costs of prolonged investigations.
I don’t think that these ethical traditions need to be mutually exclusive, particularly in this context. A workable ethics of forensic peer review may require all three to be considered. The character needed to act well under uncertainty, rules that protect people against predictable abuses, and evidence about the consequences of the procedures we adopt.
However we combine them though, we still have to state the values that they are intended to serve. David Hume’s is–ought problem is relevant here… we cannot derive an ought from an is. No collection of empirical facts can by itself tell us what should to be valued. Values have to enter as accepted premises before the logical apparatus can get moving.
Deciding on the set of values for which we would like particular ethical guidelines (I guess more virtue leaning… “good” people act in this way) or rules (more deontological… you “should” act this way or have a “responsibility” to do so) to be in service of is a required starting point to begin to logically derive what they should be. The consequentialism then begins to come into play once these are agreed in that we can begin to ask empirical questions about whether our “theories” so to speak about how best to act in service of these values obtain.
Structurally we might follow a simple approach such as:
- We agree A is an important value (particularly if that is some outcome we desire).
- And we agree that if practice B reliably advances value A then we ought to do it.
- The proposed practice B does, or does not, reliably advance value A.
- Therefore, there is, or is not, reason to adopt that practice.
The first premise is normative. The third is descriptive. Once the value has been made explicit, we can ask scientific questions about our ethical procedures. In that sense, consequentialist reasoning brings empirical testing back into the picture. If we claim that practice B should be adopted because it produces outcome A, evidence that B does not produce A undermines an important premise of the argument. Even here, a simple modus tollens falsificationist picture will not quite do though as is the case in science more broadly. Any inference about the impact of practice B from empirical observation relies on auxiliary assumptions about implementation, measurement, timescale, context, and unintended effects. This is the familiar Duhem–Quine problem, or the underdetermination of theory by evidence, and what folks such as Lakatos and Meehl grappled with in terms of theory appraisal. A failed practice may be defended as badly implemented rather than intrinsically ineffective, much as a scientific theory can be protected by adjusting its auxiliaries. That does not make evaluation impossible, but it does mean that the values, causal assumptions, and escape routes should all be made visible.
A provisional set of values?
Before attempting to develop guidelines or rules, I would start by asking whether the underlying assumed values are shared by those for whom the guidelines or rules are intended. Without this then it is too easy to find people talking past one another about the inferences that can be drawn from the empirical observations we make about those guidelines or rules.
I’m not going to attempt to lay out a provisional set of values here. I’ve written about some of my accepted scientific values and virtues in some of my previous work4. My point here is just to highlight that the values we take as assumptions need to be explicit, and ideally agreed upon, to be able to progress forwards in applied ethical matters. I say ideally agreed upon because even in the case that they are not, making them explicit at the very least allows us to identify where we disagree on fundamental normative moral premises as opposed to descriptive empirical matters. To be clear as well, our assumed values might change over time and we should be cognisant of this and how it impacts what we consider to be appropriate means of acting in their service.
Choosing where to put my attention
One personal consequence of the workshop was that I came away more comfortable with my general approach and current position… if only in terms of preserving my own sanity. I cannot really see myself as a “sleuth” in any intentional sense of the term. If I encounter something that appears wrong, I will try to address it if it feels worthwhile to do so whether personally or by making others aware of it. But deliberately searching for and pursuing case after case feels Sisyphean to me, especially when we only have limited time to spend in this life. “Whack-a-mole” was an analogy that came up at the workshop, including the feeling that sleuths have in sleuthing and for the work faced by research-integrity staff at publishers. Each resolved case leaves the system that generated it largely intact, while another case appears somewhere else.
That is not a criticism of people who choose to do this work. Their efforts can be, and usually are, extraordinarily valuable. It is instead a recognition of the scale and reactive character of the task, and of the personal costs of taking responsibility for problems that the scientific system routinely produces but only inconsistently corrects.
The expanding use of generative large language models, and especially agents able to produce and coordinate work at scale, may make that imbalance still worse. The cost of generating plausible-looking text, analyses, images, data, and submissions is falling5. Even if these tools also help with detection, an ethics built around inspecting and resolving one questionable output at a time risks being overwhelmed by volume. The whack-a-mole problem may simply speed up.
My own response is not to withdraw from responsibility. It is to keep doing my work as well as I can, to respond when I find a problem, and to make my reasoning and methods open to scrutiny. Increasingly, and particularly now that my position in industry really supports me doing so, I also want to do that work at some distance from the incentive structures of academia6 rather than optimising it for the rewards those structures offer. I cannot personally police the entire literature nor do I ever want to even try. What I can do is try to produce work whose care, openness, and seriousness are apparent by comparison… and hope that people can tell the difference.
For now, my main conclusion is that forensic peer review is not merely the application of clever tools to suspicious data. It is disciplined inference conducted in service of some set of values, whether explicit or not. How a case is found affects what can be inferred from it; how an inference is justified determines what may responsibly be claimed; and the values we adopt shape what should happen next. Making those three stages explicit may be a useful first step towards guidelines that are both principled and capable of empirical improvement.
Thanks
I’ll end just by reiterating my thanks to the organisers and all the other attendees. Not only was the location on Lake Como incredible, the workshop was really stimulating and I got to meet and talk with so many wonderful people about interesting topics. Thanks again!
Footnotes
Chatham house rule was in operation so I won’t be saying who was in attendance nor who said what exactly.↩︎
Retraction Watch also covered the accumulating retractions and investigations. My experiences in this area have not all placed me on the outside looking in. I was also a co-author on the papers affected by the Barbalho case, and I have written about how I “dropped the ball” as a co-author of a subsequently retracted meta-analysis. In another case, I encountered a retraction during the second round of searching for a systematic review and followed up with the research group; Retraction Watch reported the eventual finding of data fabrication. These different positions matter when thinking about the obligations of sleuths, authors, reviewers, editors, and institutions.↩︎
I guess you could say this is a manner in which an approach taken in the context of discovery is submitted to the context of justification. It feels somewhat akin to the way in which Faust and Meehl framed much of their work on cliometric metatheory.↩︎
See the positionality statement example I have provided in the section on Axiology in my introductory chapter Foundations of Scientific Research in the Sport and Exercise Sciences.↩︎
Whilst some will be using these tools responsibly and transparently, others might not be. For an amusing example of how easy it is to produce scholarly slop with them check out Spurious Scholar from Tyler Vigen who also created the original Spurious Correlations website.↩︎
I will say that I brought up Tal Yarkoni’s old blog post, No, it’s not The Incentives—it’s you more than once too. Whilst I don’t deny that the incentive structures do in fact influence people’s choices, I do take the position that, no matter how difficult it may be (and the specific context of course would affect my judgement of the choice made), the choice of whether to give in to them or not is ultimately down to you.↩︎