RE: Applying Corresponding Regressions to automotive mileage data setSteve,
I have sent four messages now to this list that were not printed and I assume this one will not either. But I will try again. this post is mostly for others but relates to your efforts to test CR. As I read what you have said, it makes sense that you would come to the conclusions that you did and that you would reject CR on the basis of your interpretation. I think, however, that there is more to the D score than you may know right now. In the years that followed it I considered the meaning of D=rde(y) - rde( in greater detail, partly because of confusion encountered when interpreting the results of the 1991 paper. Rde is one of several possible indices of polarization but it works ok on its own. In recent post concerning the calculation of CC from differences and sums of two variables, I presented another. I like the latter (CC) because it is easier to do but it is not necessarly more obvious. You might consider using it in the future rather than D. D uses the polarity indices (rde) from two variables pitted against one another. If both rde values are positive, it is possible (as I recall) to get a negative D. This is misleading. I worked in my studies except when I used simulations because I averaged over so many values that the the ++ cases did not contribute much to the variance. In my 200 paper I calculated polarity by multiplying the polarity indices (rde) instead of subtracting them. That way, opposition was necessary to get the negative score that suggests polarization. Earlier, I consider some real data analyses in the 1991 paper and suggested that some of the results were uninterpretable because of rde patterns that did not match the -+. One possible pattern among the four is - -, which I found in my 2000 paper in simulating reverse causes. If I recall correctly, your mileage date contained two ++ rdes and is thus not representative of a causation. But note that -- would not work if we multiplied the rde estimates, since a positive polarity would result. There may be no alternative but to simply calculate the rde estimates independently from one another and interpret them at that level, avoiding the use of indexes such as D and Polarity. Look for polarization. Dennis Roberts full comment is not included in your post but it seems characteristic of what I have come to expect from him. He seems to be looking for any excuse to dismiss CR, probably because of his contempt for me personally. As I said to you and repeat now to all others, I admire those who test ideas and if any of you can disprove CR then GREAT!. I do have other things to do. But lets do it fairly. Every multiple regression and SEM study models causation as the combination of independent variables. Think about the equation y=a +b1x1 +b2x2 etc. Most people who bother to test CR admit that with simulations it does detect linear combinations. Gottfried even thinks this so obvious as to be trivial. So, given the CR does detect combinations of variables, and assuming that Steve's interpretation was not clouded by the limited knowledge of the limitations of D, then why wouldn't Dennis Roberts scratch his head and say "Wow, now why didn't CR work for that data?" Seting aside the possibility that mileage is confounded with some other more meaningful candidate for cause, real scientists are always more interested in why something does or does not work than simply whether or not it works. We saw this same kind of unscientific thinking in Gottfried the other day. He knew very well that I had demonstrated in 1991 that CR loses its sensitivity to nested variables eventually. I said it in the paper. But he presented his replication of my work as though it were novel and as though I had tried to hide this fact. Furthermore, by kissing up to old boys on SEMNET, he missed the chance to ask important questions, such as what happens when we do not define the extremes versus midranges by quartiles but by more extreme and moderate criteria, say, the top and bottom 10% versus the middle 20%. With the help of a friend who knows calculus, we discovered that in fact the power of CC to detect the polarization went up sharply, reaching an asymptote of perfect polarization of r=+1.00 versus r=-1,00,as the extremes vs midrranges were more extremely/moderately defined. Thus CR can be made a lot more powerful when the data is there to allow for this method. These are the kinds of discoveries that real scientists make. They do not abandon good ideas just because a mob finds them inconvenient. They ask why and what if. I am a scientific psychologist. I approach statisticians for insights and to offer insights so that they might better refine ideas of promise. Over 99% of the statisticians I have contacted show absolutely no sign of trainning or talent for scientific inquiry. They are essentially dull minded dogmatists who repeat what they have been told coonverning calculations. When repeating dogma does not work, they attack personally. This is very disappointing, especially when I think about how many psychology students have been forced to take statistics from statisticians. So many are quick to jump to conclusions. We have seen such behaviors several times on this newslist, from people who have marketed themselves as great teachers. Shameful. But all is not lost. I do not know what Steve will find or conclude, but I suspect that he is at least approaching the matter as a scientist. I hope he finds the truth and when his decisions are made as to the validity of CR, I hope he will explain the logic of the math and science to us all. I suspect that Dennis will threaten to ban me from this list. Such are the ways of bad teachers and narrowminded people. But try to hang in here for awhile. I have also been working on some of the data Steve suggested, specifically the pollution and mortality data. The results are interesting and will ostensibly support the method of CR. I say ostensibly because as a theoretician at heart, I do not much trust the sort of data we are analyzing. I trust logic most, simulations second, experiments third survey type data, fourth and tea leaf reading a close fifth. Nontheless I will present the data. One more thing. My system was crashed this weekend and I lost tons of work, including email addresses etc. A couple of you have been working with me a lot lately. I would appreciate an email so I can get back in touch with you. Your addresses went with that microsoft educational software that killed my system. Best, Bill Chambers "Simon, Steve, PhD" <[EMAIL PROTECTED]> wrote in message E7AC96207335D411B1E7009027FC284902A9B27B@EXCHANGE2">news:E7AC96207335D411B1E7009027FC284902A9B27B@EXCHANGE2... Dennis Roberts writes: > what? this is a perfectly legitimate data set ... if the > model can't handle it ... let's don't blame the data I might be inclined to agree with you, but until I accumulate examples with more and different data sets, I want to be conservative in my assessment. My current suspicion is that the method is very sensitive to distributional assumptions and not very robust. I ran some more data sets with mixed results, but Dr. Chambers tells me that there is a better measure than D. When I get the details, I'll see if I can program it and then see how the new measure performs. It is kind of fun to try out all these datasets with traditional regression models and the corresponding regressions approach. Steve Simon, [EMAIL PROTECTED], Standard Disclaimer. The STATS web page has moved to http://www.childrens-mercy.org/stats. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
