[R] polr probit versus stata oprobit

[Thomas Lumley]

On Thu, 11 Nov 2004, Jean Eid wrote:

> Thank you Thomas for your answer. It was the weights that are giving me
> problems and I still have no idea why. i.e. when I try your example,
> everything work fine. However when I do not include the weights=Freq and
> [fw=Freq] in both softwares, I do get verry different results.
>

I still don't understand what example you are using to find the 
difference.  I tried two ways of not using weights

1)  Expand the data to have a record for each observation (so 1681 rows 
instead of 72).
     Fitting these expanded data  without weights gives the same answers as 
fitting the compressed data with weights, in both MASS::polr and Stata's 
oprobit.


2) Pretend that the housing data have only 72 observations and ignore the 
weights (though why you would do this...)
    The true coefficients are all zero in this situation. R gives numbers 
zero to about 6 digits and Stata gives zero to about 30 digits.  The 
intercepts are the same in both packages.


 	-thomas