My colleague Kevin Gross and I have a new preprint up on the arXiv.
Just for fun, rather than a simple text explainer, a thread with some slides for a talk I'm giving at https://www.icssi.org/ tomorrow.
Here's the paper itself: Rationalizing risk aversion in science. https://arxiv.org/abs/2306.13816
A couple of years ago, we published a PNAS paper that tackles the researchers' side of the story, explaining why grant review panels may be unlikely to fund risky studies.
https://www.pnas.org/doi/10.1073/pnas.2111615118
The present paper addresses the funding agencies' side, and looks at why researchers may be reluctant to take on high-risk projects even if they are funded.
To get at this, we have think about the incentives that academic researchers face.
Because it's very difficult to monitor the effort that researchers put in, academic scientists are rewarded almost exclusively for their research output.
Rewards come in the form of jobs, promotions, salary, and prestige, for example. We'll refer to these all as wages.
Researchers might be willing to take on risky projects if they could be insured against that risk, with wages that didn't depend on the vicissitudes of scientific fortune.
But you can't completely ensure against the failure to get results, because bad luck is indistinguishable from loafing and you need to somehow incentivize effort.
A social planner might induce effort and risk by paying large bonuses for major results.
But given any budget to allocate among researchers, this is inefficient from the researchers' perspective because risk aversion means that to increase the utility of a high earner even modestly, you have to take a lot of wages away from low earners at high utility cost.
But here's the thing: science has no external social planner.
Scientists themselves determine what results are considered worthwhile, and how worthwhile. They decide who is hired, promoted, given wages, is awarded prizes, and garners esteem. In other words, they set their own wages, of course subject to a budget constraint.
What we prove in this paper is that (in a simple model of the scientific enterprise), when left to their own devices scientists will choose for themselves a wage scheme that results in below-optimal risk-taking.
This is an inevitable consequence of the unobservability of effort and risk, coupled with the lack of an external social planner.
In other words, the rewards the science collective offers are not sufficient to motivating socially optimal levels of scientific risk.
@ct_bergstrom
Fascinating paper and thread, thank you.
Are all scientists equally capable / appropriately trained to design or execute high risk research?
Is the long run utility of high risk research undoubtedly superior to incremental progress? (cf training early career researchers and students; managing potential societal harm and other inadvertent consequences; understanding the 'signal' in fields with high methodological uncertainty).
@ct_bergstrom this all makes a lot of sense to me, and i think may help guide changes. One criticism is the assumption that higher risk level equates with higher reward. You can never know something will be a paradigm shift going in. Breakthroughs, by their nature, have to happen organically.
I would argue that a well designed study is reasonably low risk at producing scientifically useful information. It just may not match the hypothesis. Which is usually the more interesting path
@ct_bergstrom Re: "science's social planner".
I'm wondering at a glance, the role that major prizes might play in this, such as the X-Prize for the HGP or similar.
@ct_bergstrom
Incentivize effort, or incentivize results.
Effort: Punch clock at the lab. No non-work websites, minimum 20 pages worth of papers published per semester.
Results: Sum citation index of last years publications for an "index/day" metric. How, where or why the publications were produced is irrelevant.
Both approaches suck. Just like they suck in industry, and have been largely abandoned in favour of psychologically better grounded management methods.
Is that always so? I think a diligent researcher who gets off on the wrong track should be able to come back with a coherent, possibly insightful, explanation of what went wrong where a lazy one wouldn't.
@elduvelle @neuralreckoning @ct_bergstrom
People get all worked up over negative results, but a real negative result is actually really hard to show. The problem is that a negative result can be negative for lots of reasons, most of them boring.
For example, maybe the DREADD didn't affect rat behavior because there was no DREADD in the virus. (That happened to us once.) Or maybe the human subject pool was tasked wrong because they didn't understand the instructions. (That happened to us once.)
To get a real negative result, you have to have positive controls to show that all of the techniques are doing what you think they are and that the negative result is not a consequence of a trivial outcome.
Yes, you need controls for positive results as well, but it's easier to determine what those controls are, and reviewers tend to demand those controls. People who try and fail to publish negative results almost never have the right controls for those negative results (which are not the same at all as the controls you need for the positive results). A well-structure negative result experiment should be very publishable.
@ct_bergstrom
One structural issue is that research papers are used for two incompatible purposes: Disseminate research results; and evaluate the researcher who wrote them. The spread of citation indexes makes it far worse.
You do need to evaluate or rank researchers. For that, a robust way independent of published results is really necessary. Of course that's a tall order; I have no idea how to do it.
@ct_bergstrom
In my (limited) experience part of the problem is that if you *do* propose a high-risk project, the reviewers come back with "Interesting, but unlikely to work". and scores it lower than safer projects, so it doesn't get funding.
And since a proposal submission is expensive - in time, in opportunity cost (can often submit only one proposal) - you need to optimize for the best likelihood of getting funded. Submitting a low-chance proposal hurts you all by itself.
@ct_bergstrom I did some work with Thomas Sinkjær when he was at the Villum Foundation to try and address the funding side of this with the ‘Villum Experiment' grant scheme.
https://veluxfoundations.dk/en/technical-and-scientific-research/villum-experiment
Short anonymous applications, ‘golden ticket' for reviewing panel members so you don't need consensus. We talked a lot about the problem of researcher incentives - but didn't come up with a way to fix it.
Carl T. Bergstrom
Joel Truher
Anastasia Christakou
Michael Reese
jonny
Marc Kaulisch
Dr.Implausible
Janne Moren
Bernard Forgues ☮︎
Ferles 
PhDog 🇮🇪
Dan Goodman
El Duvelle
Redish Lab
James J Malcolm
Andrei Kucharavy
John Quiggin
JohnMashey
Tom von Alten
Steven