Statistical techniques have traditionally grown sidewise. Robustness now faces us with a new challenge: a need to grow by replacing old methods by new ones, at least as the default methods in many areas of application. In deciding to do this we should not wait until we know as much about the new methods as we pretended to know about the old ones. In particular, we need to change with a clear understanding that, perhaps after a decade of gaining further knowledge, we will want to change again. We need agreement to move ahead, so there will have to be compromises. Change is expensive; so we will have to take large enough steps. It is easy to suggest at least 10 areas into which robustness needs to be introduced in parallel, but not synchronous, waves. The view that “robust methods are dead” is discussed briefly, as are the difficulties associated with having equated least squares with common sense and the fact that not all initial analysis should be robust. It seems possible that the frequent use and poor understanding of regression may make it the easiest place for robust techniques (adequately enhanced by graphical diagnosis and procedures to help choose expressions) to penetrate deeply. Our failure to agree on such enhancements for classical regression compounds our difficulties. In converting to a world of repeated change, we will have to tell our clients to do things that we expect to be superseded in due course, or that a “real expert” could do better now. This will offend some consciences. We need to go ahead, and can, if we only will.
Unable to display preview. Download preview PDF.