Thursday, March 16, 2006
st: IMR is hightly correlated with X?
i must have made some terriable mistakes!
I am sorry to bother in this way but it seems that only you can help me out.
I am estimating an ordered probit model with the following form:
Size = beta*X+e
where Size can be 1, 2, 3 or 4
after the ordered probit i tried to calculate the so called truncated means, to be used to correct the bias of another Fee model.
suppose the cut off points of the previous ordered probit is u1, u2 and u3 and the linear prediction of the AHS is xb, ignoring the cut off points, I calculate the truncated means with the following foumular:
gen lambda1 = -normden(u1-xb)/norm(u1-xb) if Osize==1 gen lambda2 = (normden(u1-xb)-normden(u2-xb))/(norm(u2-xb)-norm(u1-xb)) if Osize==2 gen lambda3 = (normden(u2-xb)-normden(u3-xb))/(norm(u3-xb)-norm(u2-xb)) if Osize==3 gen lambda4 = normden(u3-xb)/(1-norm(u3-xb)) if Osize==4
where, normden() is the standard normal density function and norm() is it's CDF.
to my great surprise, i find for every Size group, Lambda is highly correlated with xb. the correlation can even be -1.00
after that, i also tried a binary choice model to calculate the IMR, i also find high correlation between xb and IMR.
as some of the variables in X will also enter my Fee model, the high correlation will make most of the variables insignificant.
I guess there is something wrong with my previous procedures, can you please help me out?
* * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/
Links to this post: