EigoPro 1級 読解 — The Confirmation We Never Print: Science's Replication Crisis

For most of the twentieth century, the reproducibility of an experiment was treated as the quiet bedrock of science—so self-evident a virtue that it scarcely required defending. That complacency has not survived the past decade. When large collaborative efforts set out to repeat landmark studies in psychology, cancer biology, and economics, a disconcerting share of celebrated findings simply evaporated under scrutiny. The episode has been christened the "replication crisis," yet the word crisis is in one sense misleading, for it implies a sudden rupture rather than the slow surfacing of a chronic condition. It is tempting to attribute the malaise to fraud, and outright fabrication does occur. But fraud is a minor tributary; the main current runs through practices that lie in a grey zone between rigour and deceit. Foremost among them is so-called p-hacking: massaging the data, or slicing it in ever finer ways, until an effect clears the talismanic threshold of statistical significance. Few who do this think of themselves as cheating. They are responding, rationally enough, to the pressures of a profession that rewards publication above almost everything else. The result is not a moral failing of individual researchers so much as the predictable output of a system whose incentives are quietly misaligned with the search for truth. Academic journals compound the distortion. Editors prize the arresting, the counter-intuitive, the positive; a dull confirmation rarely earns a place in print, however salutary its lesson. Studies that find nothing are consigned to the file drawer, vanishing from the published record and leaving behind a literature systematically skewed toward improbable claims. Science is supposed to be self-correcting, but when confirmation goes unrewarded and null results go unseen, that correction comes far too slowly to inspire confidence. Reformers have not been idle. The pre-registration of hypotheses, the rise of "registered reports" in which journals commit to publishing a study before its outcome is known, and the routine sharing of data and code are all encouraging first steps. Yet they will founder unless the deeper architecture of reward is rebuilt—unless tenure committees and grant funders learn to prize reliability as highly as they prize discovery. The crisis, in the end, is not a verdict on the scientific method but a reminder that even the most disinterested of human enterprises is shaped, for better or worse, by the incentives we attach to it.

新規登録 または ログイン が必要です