From "The Economist," September 27, 1997|
SCIENCE AND TECHNOLOGY
Applying the scientific method to the processes of science can be illuminating
The claim of science to be a superior route to the truth rests on its procedures. Hypothesis suggests experiment and experiment hypothesis in a never-ending virtuous circle. Ideas that turn out to be incorrect are ruthlessly discarded. Individual scientists may be fallible, but their weaknesses are inevitably exposed. And so forth.
The success of many scientific disciplines suggests there is much to this claim. But that does not exempt it from examination. And one way to do that is to apply to science the same method that it applies to everything else, in order to see if its everyday practices are, themselves, scientific enough. With his colleagues, Christopher Martyn of Southampton University, in Britain, has done just that. He has looked at the process known as peer review, which is supposed to filter the torrents of scientific papers that pour into editors' offices and identify those worthy of publication. And having looked, he has found it wanting.
Dr. Martyn was one of some 300 people at the recent Conference on Biomedical Peer Review, in Prague. The paper he presented made disturbing listening. Peer review involves the editors of scientific journals forwarding the papers they have received to experts for assessment. The assessments are passed back to the authors, but anonymously, the idea being to encourage honest appraisal. Papers are accepted, with or without modification, or rejected, largely on the say-so of these expert referees.
Peer review has long had its critics. The referees are usually busy people and are rarely paid for their trouble, so the process is often slow. There is also a feeling that, despite the anonymity, an old-boy network operates in some fields (an idea reinforced by a recent study in Sweden which showed how discrimination in favour of acquaintances, and against women, operates in the related area of reviewing grant applications). But some people may think that what Dr. Martyn has found is even more troubling: that even well-meaning peer review is of disturbingly low quality.
In 1995 his research group sent a paper about the risk factors for death in the elderly to the British Medical Journal. After normal peer review, it was accepted. But then, in collaboration with the journal's editors, he deliberately introduced eight errors into his paper.
The modified manuscript was sent to 420 potential reviewers from the BMJ's database. Of the 221 who responded, none identified all eight mistakes and few caught more than two or three of them. Nonetheless, the reviewers tended to be free with their suggestions. For example, one neurologist who described himself as "unqualified to comment" because he lacked the requisite training in epidemiology or statistics, wrote "having said all this, the paper is clearly rubbish . . ."
Obviously the latter problem is partly the BMJ's fault for sending the modified paper to inappropriate reviewers (though all were deemed suitable by the database). That shows the importance of picking reviewers carefully but does not undermine the whole concept of peer review. The main conclusion of the study-that even appropriate referees fail to spot mistakes-is, however, more damning. It suggests that a fundamental overhaul of the review process is called for.
A partial solution may be more use of electronics. In a virtual version of what sometimes happens at conferences, a paper could be published first to a limited audience. Criticisms would be invited and then incorporated into the final version-or not, as the case may be.
Later this year, the Medical Journal of Australia will begin an experiment along these lines. It will post research articles on the World Wide Web and give an expert group of reviewers a password that will allow them to comment. After a period, it will then give a more broadly based group of practitioners the same access, so that the version of the paper which is ultimately printed will reflect both of these points of view. How such editing-by-committee will work in practice remains to be seen-but if the experiment fails, the theory behind it can, of course, be rejected like any other failed hypothesis.
The scientific method can also be applied to assessing the papers themselves. In another conference presentation Simon Wessely of King's College, London, argued that the nationalities of authors, and their areas of specialisation, can be the enemies of objectivity. These are serious allegations. That specialists might be a little myopic is understandable, but one of science's claims to elevated status is that it is blind to such trivia as nationality. That was not, however, what Dr. Wessely found.
He deliberately chose a controversial topic-a debilitating illness commonly known as chronic fatigue syndrome (CFS), the cause (and even the existence) of which is a subject of much discussion. He and his colleagues analysed 89 overview articles published about CFS in English-language journals between 1980 and 1996.
That, in itself, might be thought a rather narrow approach to a project that was examining cultural bias (though, to be fair, even journals from non-English-speaking countries are often published in English these days). But even in the monoglot world the researchers had chosen, national biases were apparent. Studies originating in America, for example, rarely cited British research, while British reports on the subject tended to give information from the other side of the Atlantic short shrift. This suggests that the two countries' CFS researchers were, to a large extent, ignoring each other. (For good measure, different groups of experts, such as those who study infectious disease and those who study mental health, did, indeed, ignore each other as well.)
Worrying as all this is, a third issue addressed by the conference was still more disturbing. It is an axiom of the scientific method that any data being analysed in a study must be a representative sample of reality. But publishers generally prefer research that has a positive result. Papers showing that something happens are more likely to be printed than those which show that it does not, even though such "negative" results can be important.
The problem is particularly acute in the field of clinical trials for new medical treatments. The bias is so strong that many researchers do not bother to report trials in which the new treatment is no better than an existing one. At first sight, that may seem to make sense. But in this area negative results are particularly significant. That is because few trials are large and clear enough to be decisive. To overcome this, a new science, known as meta-analysis, has grown up over the past few years. Meta-analysis is a way of extracting statistically meaningful information from lots of small trials, even if they have been conducted in ways that make them difficult to compare "by eye". Its conclusions, however, are only valid if the negative trials are included as well as the positive ones. Leave out the negatives and the results may be too optimistic.
Around 500,000 controlled clinical trials are thought to have been carried out since the method was devised in 1948, and it is estimated that at least 10% of these have languished in unpublished obscurity simply because they failed to demonstrate that "A" was better than "B". In this case, however, there is good news. To try to overcome the effects of publication bias, more than 100 journals around the world made an announcement coinciding with the conference that they are declaring an "amnesty" for such unpublished work (something of a cheek, since they helped to cause the problem in the first place), and are opening a register to receive it (firstname.lastname@example.org). They have asked for researchers who have conducted a study that was never published in full, or who know of others who have conducted such studies, to come forward. Better late than never.