[R-SIG-Finance] R-squared or t-stat?

[Dirk Eddelbuettel]

On Mon, Sep 24, 2007 at 01:44:21PM +0200, Nicolas Mougeot wrote:
> Hi,
> 
> let us assume that you want to regress Y on either X and Z or W, ie you 
> hesitate between 1 bivariate model and a univariate one. As W is different 
> from X and Z, usual criterion such as AIC or BIC do not work. 
> if you had to quickly choose between the two models, would you rely more 
> on the R-squared or on the t-stat?
> my issue is that model with X and Z provide a higher adjusted R-squared 
> but low and non significant t-stat while W yields a significant t-stat but 
> a lower R-squared

IIRC the only proper way to do this comparison is to set up an
encompassing model, ie y ~ x + z + w.  There is a lot in the
literature, search for example for 'J test'. 

Hth, Dirk

-- 
Three out of two people have difficulties with fractions.