[R] how does R compute Std. Error's?

[Eik Vettorazzi]

Hi,
I think, the correct formula for beta should be
beta.hat=I(T(X)*X)*T(X)*y
and as far as I remember
beta.hat \sim N(beta,sigma^2*I(T(X)*X),
not I(T(X)*X)/sigma^2
 

Rnewb schrieb:
> i am trying to duplicate R's computation of standard errors but having some
> trouble.  i loaded some data into R and ran summary(lm(y~x1+x2+x3+0,
> data=data)), but i am not sure how the "Std. Error" values are computed.
>
> let y be the nx1 vector of dependent variables and X be the nx3 matrix of
> independent variables.  let T(.) denote the transpose of a matrix/vector,
> and let I(.) denote the inverse of a square matrix.  then i'm able to
> correctly compute the coefficients and residual standard error using the
> following formulas:
>
> beta = I(T(X)*X) * y
> resid err = sqrt(T(y)*y - 2*T(beta)*y + T(beta)*T(X)*X*beta) / sqrt(n - 3)
>
> i then try to compute the coefficient standard errors via:
>
> coeff err(i) = sqrt(I(T(X)*X)[i,i]) / (resid err)
>
> where .[i,i] means the ith entry on the diagonal of the given matrix. 
> however, doing this gives values that are off by a multiplicative factor. 
> the factor is the same for all coefficients, but it is not 1, and the value
> varies for different data sets.  what is this term?
>
> thanks,
> Rnewb
>   

-- 
Eik Vettorazzi
Institut für Medizinische Biometrie und Epidemiologie
Universitätsklinikum Hamburg-Eppendorf

Martinistr. 52
20246 Hamburg

T ++49/40/7410-58243
F ++49/40/7410-57790