metascience

Kate Shaw enters a report in the science section of Wired on a paper that modeled decision-making in animal groups: "How ignorance could improve group decisions." The paper itself by Iain Couzin and colleagues was in Science [1]. Jevin West and Carl Bergstrom authored a companion perspective piece in Science explicating some of the paper's findings. Here's the paper's abstract:

Conflicting interests among group members are common when making collective decisions, yet failure to achieve consensus can be costly. Under these circumstances individuals may be susceptible to manipulation by a strongly opinionated, or extremist, minority. It has previously been argued, for humans and animals, that social groups containing individuals who are uninformed, or exhibit weak preferences, are particularly vulnerable to such manipulative agents. Here, we use theory and experiment to demonstrate that, for a wide range of conditions, a strongly opinionated minority can dictate group choice, but the presence of uninformed individuals spontaneously inhibits this process, returning control to the numerical majority. Our results emphasize the role of uninformed individuals in achieving democratic consensus amid internal group conflict and informational constraints.

In other words, they have generated an agent-based model where each individual may have a marginal effect on group behavior relative to an "intensity" parameter, and the group's decision is dictated by the collective center of gravity. "Uninformed" in the title is a misnomer; the study examines what happens to the group as individuals with little or no bias (that is, low "intensity") are added to the group. Adding a high fraction of such individuals tends to reduce the influence of a small minority of individuals with an intense bias. In other words, the addition of low-bias individuals in the model skews the group decision in favor of a low-intensity majority.

Not surprising; the model generates the results expected given the assumptions. The study's most interesting aspect is its application of the model to a problem of schooling in fish.

So why do so many of the press accounts of this study claim that it shows that "ignorance" can support democratic decisions? The "uninformed" individuals in the model lack information only in the sense that they lack a strong preference. In no other sense does information come into the model, other than as a measure of decision bias. This is "information" more or less in the sense that someone can "ask the audience" on "Who Wants to Be a Millionaire?" The contestant basically has to hope that a minority of the audience knows the correct answer, while the majority has no bias toward a false answer.

I've been thinking a lot about modern human origins lately, as I'm finishing up an e-book project. For many years I have worked very closely with the details in this area, really pushing up my sleeves and getting into both the genetic and morphological aspects of the problem. What always struck me was how the "scientific consensus" emerged in favor of an Out-of-Africa replacement hypothesis, and against the multiregional evolution hypothesis. The protagonists on either side of the modern human origins controversy didn't really change very much over time. There were few high-profile "defections" from one side to the other. Initially, many of the best-known paleoanthropologists and geneticists sat on the fence. After all, few people work so directly with the evidence of modern human origins, when there is a debate it is reasonable to be cautious. But over a decade, like wildebeest waiting to jump into the crocodile-infested river, paleoanthropologists and geneticists whose work has little to do with modern human origins began to tip in favor of the replacement scenario.

They jumped too soon.

But why? I've been tackling this question: How did the majority of paleoanthropologists and human geneticists get this one so wrong?

I feel so fortunate to have been engaged in this problem, because it says so much about the process of science. Science is always a process where progress requires an opinionated minority to recruit support among peers who are not specialists in the same area. Such a minority may forge consensus through consistent and repeated demonstration of facts. More likely -- as in the case of modern human origins, where new evidence was often equivocal -- a motivated minority will apply a broader range of rhetorical strategies. Over the years, I saw people pull out every trick in the book to persuade the uncommitted to their point of view.

Rapid visual processing allows schooling fish to signal and comprehend the direction of movement within milliseconds. Scientists signal each other through their publications, grant reviews, and the press. It takes a lot longer for scientists to school.

References

Razib Khan conveys a list of suggestions from a recent paper by Joseph Simmons and colleagues [1], concerned with reducing reporting biases in research papers. The article is directed toward psychology research, but many of the observations hold true in paleoanthropology and genetics.

The central point is that every research paper is a product of a course of inquiry that may come to include many kinds of questions, most of which are unanswered or answered negatively by results and data. When scientists report results, they focus on those that meet some statistical threshold. The threshold ostensibly makes results "significant" but the actual probability of seeing such a result depends on how many things the scientists looked at, not only on those they choose to report:

In this article, we show that despite the nominal endorsement of a maximum false-positive rate of 5% (i.e., p ≤ .05), current standards for disclosing details of data collection and analyses make false positives vastly more likely. In fact, it is unacceptably easy to publish “statistically significant” evidence consistent with any hypothesis.

The culprit is a construct we refer to as researcher degrees of freedom. In the course of collecting and analyzing data, researchers have many decisions to make: Should more data be collected? Should some observations be excluded? Which conditions should be combined and which ones compared? Which control variables should be considered? Should specific measures be combined or transformed or both?

It is rare, and sometimes impractical, for researchers to make all these decisions beforehand. Rather, it is common (and accepted practice) for researchers to explore various analytic alternatives, to search for a combination that yields “statistical significance,” and to then report only what “worked.” The problem, of course, is that the likelihood of at least one (of many) analyses producing a falsely positive finding at the 5% level is necessarily greater than 5%.

"Researcher degrees of freedom" sounds erudite, but all they're really describing is tinkering. When the data lead you to a result, they do so by leading you along a drunkard's path of new analytical biases.

Razib has presented the authors' suggestions for researchers and reviewers, to try to reduce the tinkering bias. I think if we followed those suggestions in paleoanthropology, our discipline would be stronger in some ways, weaker in others. For example, the authors suggest rejecting any paper with fewer than 20 observations in a cell of a test of association. Clearly, if we rigidly enforced such a rule, we'd have a lot more work done on comparative collections, and that would be a good thing. On the other hand, we'd have a lot more papers like the ones I like to write, about how the data are insufficient to test a hypothesis. Our science would shift even further toward description, which would benefit some kinds of research and punish others.

One may object that there are many cases in paleoanthropology where a single observation is fundamentally important. I would just point out that such cases are most evident where the single observation is a many-sigma outlier to some pre-existing hypothesis. If we have a new radiocarbon date that's a three-sigma outlier above previous dates, it will either cause us to change our hypothesis or challenge the date's accuracy. There are biases even so -- for example, when we find outlier radiocarbon dates on otherwise-uncontroversial things, we tend to just ignore the outliers.

What I most liked about this paper was that the authors anticipated various objections. For example, many researchers would claim that a Bayesian statistical approach would eliminate or reduce the bias from "researcher degrees of freedom". Here's the authors' response:

Although the Bayesian approach has many virtues, it actually increases researcher degrees of freedom. First, it offers a new set of analyses (in addition to all frequentist ones) that authors could flexibly try out on their data. Second, Bayesian statistics require making additional judgments (e.g., the prior distribution) on a case-by-case basis, providing yet more researcher degrees of freedom.

That's my observation as well. Researchers adopt Bayesian methods for more ways to tinker. I also appreciate this comment:

We are strongly supportive of all journals requiring authors to make their original materials and data publicly available. However, this is not likely to address the problem of interest, as this policy would impose too high a cost on readers and reviewers to examine, in real time, the credibility of a particular claim. Readers should not need to download data, load it into their statistical packages, and start running analyses to learn the importance of controlling for father’s age; nor should they need to read pages of additional materials to learn that the researchers simply dropped the “Hot Potato” condition.

Furthermore, if a journal allows the redaction of a condition from the report, for example, it would presumably also allow its redaction from the raw data and “original” materials, making the entire transparency effort futile.

All in all, the article is a good reminder of Feynman's first principle, "You must not fool yourself, and you are the easiest person to fool."

References