In article <[EMAIL PROTECTED]>, Rich Ulrich <[EMAIL PROTECTED]> wrote: >On 26 May 2003 09:26:39 -0500, [EMAIL PROTECTED] (Herman >Rubin) wrote:
>> In article <[EMAIL PROTECTED]>, >> john v verkuilen <[EMAIL PROTECTED]> wrote: >> >Rich Ulrich <[EMAIL PROTECTED]> writes: >RU > >> >>A final word about assumptions: "Do the analysis both ways." >> >>When you get different results, THEN you can worry about >> >>what assumptions have gotten in the way. >jayv > >> >This is darn good advice. IMO the main contribution of robust estimation is >> >to provide systematic alternate methods, even if one intends to use standard >> >ones. >HR > >> This is darn BAD advice. There are usually dozens of ways >> of "doing the analysis", and plugging data into a cookbook >> without knowing the probability assumptions is not the way >> to get good answers. >Here, perhaps, is the difference between data that Herman sees, >and the biostatistical applications that concern jayv and me -- >Herman can say that there are "usually dozens of ways" of >doing the analysis. By contrast, Jay and I have users who >will be educated, and sometimes pleased and impressed >at the broader picture that is provided by TWO ways of doing >the analysis. It is often the case that the two ways of doing the analysis both essentially use unstated assumptions which the investigator cannot even understand, and might well not be willing to approximately accept if understood. >By the way, >If there are dozens of ways available of doing the analyses, >well, the naive readers will be better served by insisting on trying >all the dozens, and then improving their educations by studying >and contrasting them ... than by flipping a coin to see which >ONE to run. Actually not. Unless there is a huge amount of data, using a dozen methods of fitting is likely to find one which seems to fit, even if all are nonsense. And if one happens to be the right model, another one may well fit better by accident. >Further "by the way", >It does seem to me that Herman seems to give advice as if >his clients can sprout a healthy knowledge of "probability >assumptions" and their consequences without ever looking >at an example. By contrast, I find examples to be useful. >And, if two models do NOT have different consequences that >will show up in *some* computer output, then it is not worth >distinguishing between them. One does not get to know how to make probability assumptions except by first getting a fair amount of conceptual probability (NOT computing with equally likely, or making things normal), and also learning how to use probability in the subject field. As a statistician, I can attempt to teach you the former, but not in a single posting, and at most can question you about the latter. As for not distinguishing between them, suppose one tells you to water and use certain fertilizers, and the other tells you the opposite? >[ snip, useful comments about robustness; symmetry] -- This address is for information only. I do not claim that these views are those of the Statistics Department or of Purdue University. Herman Rubin, Deptartment of Statistics, Purdue University [EMAIL PROTECTED] Phone: (765)494-6054 FAX: (765)494-0558 . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
