Ethics and Statistics Andrew Gelman

An ethics problem arises when you are considering an action that (a) benefits you or some cause you support, (b) hurts or reduces benefits to others, and (c) violates some rule. Other definitions are possible; there is a vast literature on professional ethics that I will not discuss, instead focusing here on my own perspective as a statistician. Any ethical dilemma can be transformed into a close call by complicating the costs and benefits, making the rules violations less clear, and adding uncertainty. Ethical subtleties are often explained to children through questions such as, “Is it wrong to steal from a store?” Maybe not if the store is a drugstore that is closed for the night and you might need a certain drug to save a life right now. Similar twists can be given to ethics problems in science. Consider Mark Hauser, the Harvard psychologist who was found responsible for scientific misconduct after his research assistants became convinced, in the words of the Chronicle of Higher Education, that he was “reporting bogus data.” What if a researcher knows, simply knows, a certain theory is true—but, annoyingly, other researchers in the field disagree. Many scientists have had this experience: We make our point clearly, our reasoning is evidently correct, but others persist in believing the opposite. Is it ethical, then, to fake one’s data? I would say no, but one might argue that shortterm fakery is the best way to advance scientific truth as he sees it—and isn’t truth more important than silly rules? In other settings, behavior we would describe as unethical is considered by others to be simply part of the game. For example, a friend of mine, an academic statistician who occasionally does legal consulting, told us of a case in which he submitted an innocuous report on a random sample he had collected. Later, my friend learned the consultant on the other side of the case had written a rebuttal attacking his competence. The attack was baseless and my friend easily refuted it, but to just attack solid work, presumably for no other reason than you are being paid to do so, seems unethical to me. I suspect that consultant, however, merely saw this as standard practice, no less ethical than bluffing in a poker game or starting with a lowball offer in a negotiation. Although any ethics violation can be framed to be ambiguous, this does not, and should not, negate the importance of ethics. Statisticians should be able to appreciate the necessity of decisionmaking under uncertainty and ambiguity. In future columns, I would like to explore many dimensions of ethics, including those that arise in clinical research (e.g., concerns about randomly assigning patients to a control condition believed to be less effective than an available treatment) and statistical analysis (e.g., practices such as fishing with regressions to get statistical significance or, from the other direction, slicing data into small parts so as to lose significant comparisons in the noise) to problems involving probability and uncertainty (e.g., regulations that aim for an unrealistic zero tolerance for risk), as well as more general concerns such as plagiarism and misrepresentation of research findings. Ethical challenges arise from many sources, including conflicts of interest, imprecise rules, uncertainty, and tradeoffs in values and consequences. As statisticians, our greatest contribution here may ultimately come from quantifying tradeoffs, as in evidencebased medicine and evidence-based social policy. Before attempting any sort of quantitative treatment, however, I will tell some stories. The story for the present column concerns the ethical imperative to share data.