The Premier Online Community for Real Estate Appraisers!
-
Welcome to AppraisersForum.com, the premier online
community for the discussion of real estate appraisal. Register a free account to be able to post and unlock additional forums and features.
How Do I Make Adjustments In Square Footage
- Thread starter Patrick Eubanks
- Start date
- Status
Jungle Boy
Senior Member
- Joined
- Oct 30, 2002
- Professional Status
- Real Estate Agent or Broker
- State
- Florida
I agree with the first part of the quote.
The second part concerning the 10 adjustment & the 20 adj...
In practise, the two numbers are far apart. 20 is the double of 10.
Obviously there would be some that would fall below 10, some between 10 & 20, and some may be above 20.
Could it be said that the adjustment progresses from 10- to 20+ for homes that were built between so & so dates. Or is this too basic ?
tommcsherry
Senior Member
- Joined
- Mar 11, 2004
- Professional Status
- Certified Residential Appraiser
- State
- Nevada
All:
I just came across this...it's new and on sale at the AI site for $35 (for members)...45 for non members. Sounds like it might be interesting.
Practical Applications in Appraisal Valuation Modeling
M. Steven Kane, Mark R. Linné, MAI, CRE, CAE, ASA, and Jeffrey A. Johnson, MAI
Managing and analyzing the torrent of data available today can be a struggle, but technological tools and sophisticated analytical techniques have emerged to help appraisers meet this challenge. Practical Applications in Appraisal Valuation Modeling charts new territory and illustrates how the techniques of statistical analysis once used only in mass appraisal and in the classroom have real-world applications and may become an essential component of appraisal practice.
A resource for both valuation veterans and curious newcomers, this new book takes the reader through the analytical process step by step, from exploratory data analysis through linear regression modeling. The benefits and pitfalls of statistical modeling are examined and sample applications are demonstrated using the types of real estate situations and data appraisers commonly encounter.
Maybe we ought to get up a group purchase.
I just came across this...it's new and on sale at the AI site for $35 (for members)...45 for non members. Sounds like it might be interesting.
Practical Applications in Appraisal Valuation Modeling
M. Steven Kane, Mark R. Linné, MAI, CRE, CAE, ASA, and Jeffrey A. Johnson, MAI
Managing and analyzing the torrent of data available today can be a struggle, but technological tools and sophisticated analytical techniques have emerged to help appraisers meet this challenge. Practical Applications in Appraisal Valuation Modeling charts new territory and illustrates how the techniques of statistical analysis once used only in mass appraisal and in the classroom have real-world applications and may become an essential component of appraisal practice.
A resource for both valuation veterans and curious newcomers, this new book takes the reader through the analytical process step by step, from exploratory data analysis through linear regression modeling. The benefits and pitfalls of statistical modeling are examined and sample applications are demonstrated using the types of real estate situations and data appraisers commonly encounter.
Maybe we ought to get up a group purchase.
tommcsherry
Senior Member
- Joined
- Mar 11, 2004
- Professional Status
- Certified Residential Appraiser
- State
- Nevada
Also, if anone in interested...here's a link I found to an NAIFA article that discusses modeling, some of the available software, and how to use Excel2000 to build a regression template:
http://www.naifa.com/gram/2002oct/smith-oct02.html
http://www.naifa.com/gram/2002oct/smith-oct02.html
SHaugabrooks
Freshman Member
- Joined
- Aug 23, 2002
- Professional Status
- Certified Residential Appraiser
- State
- Florida
The institutute also teaches a class on this matter of regression. I am definetly taking this one. See attachment for details.
Austin
Elite Member
- Joined
- Jan 16, 2002
- Professional Status
- Certified General Appraiser
- State
- Virginia
I have already ordered my copy of the AI's new book and it is due Monday. The AI has another book that came out a couple of years ago that I purchased and is good too. We discussed the AI's regression class on this forum and one of the people that wrote the above referenced book is the instructor. I have heard mixed reviews from people that have had that course.
This seems to imply that you can use the entire dataset of all sales of SFR's without regard to segmentation, and as long as R-squared is above about 90, you've got a "good" model. Correct? Incorrect?
Because if that inference is correct, there's a tick in my longjohns. This "model" would allow you to derive adjustments for a 3500 sf two-story house from data produced by 900 sf starter homes. That, to me, is like fingernails on a blackboard--and a source of the famous Mark Twain quote "Lies, damned lies, and statistics." I would think that "comparable to begin with" is a necessity.
Tom,
With regard to the AI book on modeling, if a few of us buy it and work through the problems here on the forum, maybe we can teach each other the methods. Might be especially effective if we can get Steve, Austin, and a few of the others who've worked with this method to write in with criticisms and pointers from time to time.
Austin
Elite Member
- Joined
- Jan 16, 2002
- Professional Status
- Certified General Appraiser
- State
- Virginia
Jim:
You are basically correct but with a large enough database you can program in the segments with numbers. Using a large database to appraise a small segment of the range is not necessary except for mass appraisal possibly. I worked through all of that and a lot of other stuff that appraisers don’t even know exists until they use regression to see what is going on. If we could post graphs I would post the graph of what 2000 sales look like in this market across the spectrum. If I restrict the regression to modern one-story houses less than ten years old it is a pretty good correlation, but as you say when you have a wide size range in the data base you inject covariance into the equation because there is a high correlation between quality of construction and size in this market.
Generally speaking an R^2 above 72 is acceptable. Test statistics are not a good measure of how good the regression is for various reasons. The test is how well does the equation predict the prices of the sales with the least amount of variance. If the variance between actual and predicted prices is low, who cares what the test statistics say. That is a circular reference to you premise, if you use a less number but more comparable sales the test statistics are worse because the test statistics take into account the number of sales in the sample. If you have 15 sales with four independent variables you could have a perfect regression equation but terrible test statistics.
There was an article in the Appraisal Journal last year about this. A PHD was hired to evaluate a regression analysis based appraisal done on industrial property by some poor appraiser. The PHD's conclusion was that it was a bad appraisal because the test statistics were bad and he recommended holding up on using regression methods until new test statistics were devised. Actually the raw data in this appraisal was sufficient for most limited scope appraisals of that type property but the appraiser's major sin was making a couple of assumptions about access affecting value. I worked the problem and got a good result but I don’t remember the details. The real problem was that the appraiser did not know how to use the correct sequence of adjustment extraction procedure.
You are basically correct but with a large enough database you can program in the segments with numbers. Using a large database to appraise a small segment of the range is not necessary except for mass appraisal possibly. I worked through all of that and a lot of other stuff that appraisers don’t even know exists until they use regression to see what is going on. If we could post graphs I would post the graph of what 2000 sales look like in this market across the spectrum. If I restrict the regression to modern one-story houses less than ten years old it is a pretty good correlation, but as you say when you have a wide size range in the data base you inject covariance into the equation because there is a high correlation between quality of construction and size in this market.
Generally speaking an R^2 above 72 is acceptable. Test statistics are not a good measure of how good the regression is for various reasons. The test is how well does the equation predict the prices of the sales with the least amount of variance. If the variance between actual and predicted prices is low, who cares what the test statistics say. That is a circular reference to you premise, if you use a less number but more comparable sales the test statistics are worse because the test statistics take into account the number of sales in the sample. If you have 15 sales with four independent variables you could have a perfect regression equation but terrible test statistics.
There was an article in the Appraisal Journal last year about this. A PHD was hired to evaluate a regression analysis based appraisal done on industrial property by some poor appraiser. The PHD's conclusion was that it was a bad appraisal because the test statistics were bad and he recommended holding up on using regression methods until new test statistics were devised. Actually the raw data in this appraisal was sufficient for most limited scope appraisals of that type property but the appraiser's major sin was making a couple of assumptions about access affecting value. I worked the problem and got a good result but I don’t remember the details. The real problem was that the appraiser did not know how to use the correct sequence of adjustment extraction procedure.
tommcsherry
Senior Member
- Joined
- Mar 11, 2004
- Professional Status
- Certified Residential Appraiser
- State
- Nevada
Austin:
Let me know what you think of the book....especially if you think it's good enuf to be able to develop a working model dummies like me could easily use to get up and running.
Jim:
Good idea...pending Austin's comments on the book.
Happy Easter all.
Let me know what you think of the book....especially if you think it's good enuf to be able to develop a working model dummies like me could easily use to get up and running.
Jim:
Good idea...pending Austin's comments on the book.
Happy Easter all.
Posted by Jim,
It is hard for me to say where analysis leaves off and regression analysis begins. Some brilliant appraiser wrote an article showing how you could pair, graph or use the regression equation to get the same answer under certain conditions. However, there is something of a big difference between 1) working with a line of best fit (or trend line as Austin calls it) which you can even draw by hand on your workfile folder and 2) a fully-tested statistically-signficant correlation.
Regression was developed in the context of random samples and populations. When we start sorting sales, we de-randomize the sample. (This is where I start to disagree with my esteemed colleague Austin). Working with small and de-randomized samples, the whole idea of “correlation” gets distorted. If the sold properties are similar enough, then price per doorknob and price per light bulb will come in with Rsq’s of 90% because they are (as Austin likes to say) covariant with square footage. That is, more house means more doors and more lights.
The “test” statistics that Austin refers to include a calculation that indicates the chance that your “correlation” occurred by accident. Suppose a scientist wants to test the hypothesis that a flipped coin will always land on heads. He spends six months testing the coin. He flips it twice, it comes up heads twice and he publishes his findings in a paper. The problem is that this pattern of results occurs by random chance 25% of the time. So even though his “correlation” is 100%, the results are not “statistically significant.”
I know that example is off the wall, but it is to illustrate a point. We all knew the experiment had an absurd design – and yet – I would say 25% chance that it is coincidence sounds low compared to the magnitude of the absurdity. This, in my opinion, is why it is important to run the whole show if you are going to tell the client it is “regression” and you have high “correlation.” It is better in some ways to have 70% correlation that is almost surely accurate, than it is to have 95% correlation that could easily be an illusion.
We had one of these discussions a little over a year ago about a specific three-sale sample. The graph pattern had an important quirk. Two sales were close together, one above the other. The third sale was off to the right. The line of best fit had to pass through the third point and half way between the first two. There is just nowhere else for it to go. Now imagine a graph as big as the solar system and move only that third point out to about where Pluto is. The line of best fit still passes through the third point and halfway between the first two. The “value” of the first two points does not change and the “value” of their distance from the line does not change. The Rsq “correlation” does not change either.
This is a good illustration of how random chance offsets high correlation in a small sample. There are just too many ways to get this result by accident. With three sales that are visually simlar and on the same street, I think 90% price-size correlation would be “low.” So Jim, we have to be careful lest Mark Twain's ghost could rise and say - there are liars, damn liars and three-sale R-squared.
You have a knack. I hardly know where to begin, but I would like to point out that Mark Twain predates regression.
It is hard for me to say where analysis leaves off and regression analysis begins. Some brilliant appraiser wrote an article showing how you could pair, graph or use the regression equation to get the same answer under certain conditions. However, there is something of a big difference between 1) working with a line of best fit (or trend line as Austin calls it) which you can even draw by hand on your workfile folder and 2) a fully-tested statistically-signficant correlation.
Regression was developed in the context of random samples and populations. When we start sorting sales, we de-randomize the sample. (This is where I start to disagree with my esteemed colleague Austin). Working with small and de-randomized samples, the whole idea of “correlation” gets distorted. If the sold properties are similar enough, then price per doorknob and price per light bulb will come in with Rsq’s of 90% because they are (as Austin likes to say) covariant with square footage. That is, more house means more doors and more lights.
The “test” statistics that Austin refers to include a calculation that indicates the chance that your “correlation” occurred by accident. Suppose a scientist wants to test the hypothesis that a flipped coin will always land on heads. He spends six months testing the coin. He flips it twice, it comes up heads twice and he publishes his findings in a paper. The problem is that this pattern of results occurs by random chance 25% of the time. So even though his “correlation” is 100%, the results are not “statistically significant.”
I know that example is off the wall, but it is to illustrate a point. We all knew the experiment had an absurd design – and yet – I would say 25% chance that it is coincidence sounds low compared to the magnitude of the absurdity. This, in my opinion, is why it is important to run the whole show if you are going to tell the client it is “regression” and you have high “correlation.” It is better in some ways to have 70% correlation that is almost surely accurate, than it is to have 95% correlation that could easily be an illusion.
We had one of these discussions a little over a year ago about a specific three-sale sample. The graph pattern had an important quirk. Two sales were close together, one above the other. The third sale was off to the right. The line of best fit had to pass through the third point and half way between the first two. There is just nowhere else for it to go. Now imagine a graph as big as the solar system and move only that third point out to about where Pluto is. The line of best fit still passes through the third point and halfway between the first two. The “value” of the first two points does not change and the “value” of their distance from the line does not change. The Rsq “correlation” does not change either.
This is a good illustration of how random chance offsets high correlation in a small sample. There are just too many ways to get this result by accident. With three sales that are visually simlar and on the same street, I think 90% price-size correlation would be “low.” So Jim, we have to be careful lest Mark Twain's ghost could rise and say - there are liars, damn liars and three-sale R-squared.
- Status