My model is Austin, who left the board 20 years ago. He said "goodbye" and that's the last we heard from him. Austin was one of the best at understand regression analysis.
There's a lot good to say about Austin. (Before he left the forum in 2012 he was fairly prolific in his posts) You can still find all his posts. Search member="Austin," and topic="regression". - Aside from me he was probably the most advanced ever on this forum in term on understanding regression and making adjustments. Yet, nonetheless, he was a long way from understanding advanced regression. He definitely had the idea. But, he couldn't do MARS. And, no, certain things he just didn't understand:
Austin: "Over the years I have debated the subject of “matched pairs” at nauseum on this forum. Some years ago I published two articles on multiple regression analysis and spent years of research on this subject.
Being as brief as possible let me explain why “matching pairing” (and making adjustments) is a hoax and why it should never stand up in court or at a board hearing. If one accepts “matching pairing” one is flying into the face of math and science or in other words: If match pairing is a viable concept then the laws of math and science are wrong. That is why I refer to the way most people using a sales analysis method as “voodoo.”
Going to the crux of the argument is a statistical problem called “MULTIPLICOLLINEARITY.” What that words means is that when you have a regression equation the dependent variable (Y-value) is determined by value influencing factors (independent variables like property features and characteristics). The appraisal sales analysis approach is in essence a regression problem solved by various methods (mostly bogus methods like match pairing, market surveys, rules of thumb or boiling chicken bones in vinegar).
Here is how multicollinearity and match pairing butt heads. Multicollinearity is what happens in a regression equation (marketing grid) when a number of property features like age, condition, size, etc., all correlate with value.
If they didn’t correlate they should not be used. For example, generally larger houses sell for more than smaller houses.
What happens in a regression solution with a number independent variables is that when the value influencing factors are correlated to value then by simple logic the independent variables correlate to each other. If the slope of the size trend line goes up and quality of construction increases with size then the two trend lines must be correlated to not only value Y but to each other. The effect of this is that you can have a good regression fit from the equation but the predicted values of what are generally called adjustments (independent variables) are totally bogus. That means you cannot take a
regression system like the sales analysis grid and make adjustments.
It is simply because it is mathematically impossible to separate out the contribution of any one independent variable which is exactly what “match pairing” attempts to do.
In essence, making adjustments is a hoax because each adjustments results from match pairing. The only solution to this problem is reconciliation in the aggregate. I have developed a method of doing this and tried to post an example but the forum won’t let me up load PDF’s and demonstrate the method.
If you want a second opinion I will post this link as a source. Read the second two columns and it will explain multicollinearity and you can use it as a source if you get caught."
His language is oblique and imprecise. He conflates all kinds of issues. .... And just doesn't have the toolset to understand MARS.
And now we know ....
community for the discussion of real estate appraisal. Register a free account to be able to post and unlock additional forums and features.
