Linear Regression Halp

Well, it seems I missed the class on transformations or they were never taught to us, so I need a bit of help.

1)  Is there any limit to how many different types of transformations can be used in one regression?  For instance, I was messing around and log-transformed some variables, left some alone, squared another variable and used the Fisher transformation on yet another.  My R squared was like .95, but it seemed like I was cheating a bit.

2)  Upon getting the results of the transformed regression, must you do anything to the coefficients or anything before interpreting them, or can you just leave them as is?


Edited by Hysteria - 30 April 2010 at 9:04pm

It sounds like you're in statistics.  In that case, my advise is to take a real math class.

I don't what the hell you're talking about, but this is how we do linear regressions in real math classes (Numerical Methods):

Using the model:



we find:



note ybar is the average value for y and xbar is the average value for x.

Edit:  Forgot the error calculation:




Edited by WGP guy2 - 30 April 2010 at 10:04pm

What the hell? My god I have trouble with college algebra.


I'll just stick with medicine, which ironically, is so much easier for me.

Wow.... Math stopped beaing math when the letters start flying......

Hysteria wrote: Not statistics, econometrics.  Yes, I know how to take a real linear regression like you describe.  However, if the data are not quite linear (or for many other reasons), you can use real regression technique called data transformation.

Ah, I see.  We touched on that but never went in depth.

Of course, you could just use Newton, Lagrange, or natrual spline interp and have essentially 0 error.  Damn economists.   

procarbinefreak wrote: statistics have their place in the world.  they're used every day where I work... and are rather important.

Oh I didn't say it wasn't important.  It's just not real math. 

^^^ In that case . . .



. . . I do want fries with that.