In article <[EMAIL PROTECTED]>, Dennis Roberts <[EMAIL PROTECTED]> wrote:
>>e role of the statistician is to advise the person with >>the statistical problem on how to treat it, often devising >>new algorithms in the process. It is also to question the >>investigator about the assumptions, and to point out what >>assumptions make little difference. >you sure put the role of the statistician down towards the bottom of the >barrel ... sometimes in fact, the stat person is a co equal participant in >the research or questions being explored ... stat people can have ideas too They can have ideas, but AS STATISTICIANS they must not push their ideas to the detriment of the problem. Those who help calculate the trajectories of space probes can have ideas, but this does not mean that they override the knowledge of physics and chemistry of the ones using the space probes. >you make it sound like the statistician is merely a passive ingredient ... The statistician may well have to devise a procedure to treat the user's problem. I have sometimes done this, and this does not mean that I can use a canned program. But it was still the USER'S problem. >> >You can't expect every person who has a legitimate interest in "data" to be >> >a "true" statistician ... just like we can't expect every person who has a >> >valid interest in weather phenomena ... to know all the intricate facts and >> >"history" about weather ... >>I never claimed this. The investigator, however, is the one >>who must make the probability assumptions about the data. >well, i would reinterpret this as ... it is the investigator who is >primarily responsible for trying to outline and detail as much as possible >... WHAT he/she is trying to accomplish ... what he/she is primarily >interested in exploring ... BUT, the stat person should not just sit idly >by ... they should try to HELP the party in seeing what might be reasonable >and unreasonable ways to define/collect/examine the data ... Helping decide how to collect data does not involve making assumptions about how the collected data will behave. In fact, just how to go about it depends on the assumptions, which need to be known before the collection. When it comes to examining the data, the statistician can say what is there and what is not, based on the user's assumptions. >>It is often the case that the assumptions are even more >>important than the data. It can even be the case that for >>some aspects the data is more important, for others the >>assumptions. >ok ... but, the statistician can help in trying to weigh in on what is the >case in that particular situation ... The statistician can certainly provide information on what is, and what is not, important among the assumptions. But the statistician needs to know this, and I have yet to see a single situation where it is reasonable to attempt to get the distribution of any data even approximately normal. I have also pointed out that normalizing the data makes the inference problem much harder; regression coefficients are much better behaved than correlation coefficients. >>As a coauthor, it is still important not to use statistical >>convenience for making assumptions which affect the results, >>rather than going back to the subject matter field. >i certainly agree with this ... BUT, this does not mean the statistician >has to sit closed lipped and not try to help the content person frame the >ideas/questions in ways that analysis is appropriate (and not) for ... Yes, the statistician needs to point out the assumptions which the user is unconsciously making. Non-normal stochastic processes with continuous independent increments do not exist; in fact, do processes with continuous independent increments exist in nature? It is possible to look at the data and spot features of it which make what should be a reasonable analysis wrong. >should not the penultimate goal be to help the client ... do the best >he/she can? or, be honest and say to the client ... this is a mess too >large to salvage ... Yes, but not by plugging into inappropriate formulas. >>My "commandments" bear repeating. >> I am often requested to repost my five commandments. These are >> posted here without exegesis. >> For the client: >> 1. Thou shalt know that thou must make assumptions. >ok ... makes sense >> 2. Thou shalt not believe thy assumptions. >then why make assumptions? Because it is necessary to approximate. Without assumptions, no data can be analyzed. The history of the universe is a stochastic process cannot be disproved. >> For the consultant: >> 3. Thou shalt not make thy client's assumptions for him. >well, the consultant should not do the client's project ... THAT i agree with >> 4. Thou shalt inform thy client of the consequences >> of his assumptions. >but, in #2 ... you said for the client not to believe their assumptions ... >so, does the client make assumptions or not? if the client makes a stupid >assumption ... that you said he/she must make ... shouldn't the consultant >speak up and help the client reshape or consider assumptions that seem to >make more sense? See some of the above comments. The consultant certainly should question the client on assumptions, and not too many do this enough. >>There are, unfortunately, many fields in which much of the activity >>consists of using statistical procedures without regard for any assumptions. >sure, this boils down to doing stuff poorly ... not just bad assumptions ... No, it is bad assumptions. The assumption made by those in education that all children are of roughly equal learning ability and that therefore they should all be taught in essentially the same way (this is even an understatement) is how they have acted for decades. There are few among investigators who seem able to recognize that the point null hypothesis is rarely even tenable except as an approximation, and that a difference not known to be statistically significant can be important. >> >I do agree that software makes "output" easy to come by but, that is a >> >totally different matter ... >>Not really. The oracle only answers the questions asked, >>not the real questions. >how can we know ad hoc ... that this is the case? i would say that many >questions are about as fully answered by standard output ... as one can get >... others, not so ... >one can't make a blanket statement that applies in all (or even necessarily >the majority) cases >> One does not just approach the >>altar, pay the priest his fee, present the sacrifice, and >>expect a miracle to provide the state of the universe. >depends on the religious convictions you hold ... There is an article entitled, IIRC, "Statistics, the religion of medicine." The normal distribution, and the statement that if there is a difference between the performance of groups, there must be discrimination, seems to be the religious convictions of educationists. -- 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/ . =================================================================
