Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: RE:xtpcse vs. xtpcse, noconstant

From   "Nick Cox" <>
To   <>
Subject   st: RE: RE:xtpcse vs. xtpcse, noconstant
Date   Tue, 10 Aug 2004 12:07:32 +0100

Sounds risky to me. 

As said, the R^2 is just not a guide to model 
merit in this case. Otherwise I clearly can't 
comment on your results, but I'd advise seeking out 


John A. Nelder. 1998. 
The Selection of Terms in Response-Surface Models
-How Strong is the Weak-Heredity Principle?
The American Statistician 52(4): 15-318.

and following references from there. 


> Thank you, Nick.
>   Mine is a yield response model and since I am including 
> mostly inputs as
> explanatory vars and some categorical dummies, so perhaps a 
> model without
> const could work since no matter what I include, I get only 
> good results
> with noconst.
> Cris
> Date: Mon, 9 Aug 2004 09:14:27 +0100
> From: "Nick Cox" <>
> Subject: st: RE:
> On the R^2, your starting point is now a prediction of zero,
> not a prediction of the mean response.
> In a much simpler case, below,  dropping the constant
> gives a higher R-sq but a totally ludicrous model. Why then
> does the R-sq look so good? Because the predictions
> - -- which range from 11 to 30 mpg -- are much closer to
> the data than a prediction of 0 than the predictions of
> the first model to the mean of -mpg-. Your model is more
> complicated, and I can't see your data, but I guess that
> the same applies. If there is a really good reason,
> like a law of physics, to force predictions through
> the origin, then do it. (One can certainly improve
> on a linear regression of -mpg- on -weight-, a secondary
> point.)

*   For searches and help try:

© Copyright 1996–2015 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index