Many thanks for your response. In my case X1 and X2 do have more than one element. Thanks for clarifying though that OLS is appropriate only for prediction.
sk [EMAIL PROTECTED] (Herman Rubin) wrote in message news:<[EMAIL PROTECTED]>... > In article <[EMAIL PROTECTED]>, > sk <[EMAIL PROTECTED]> wrote: > >Hi everyone, > > >I have two structural equations of the following form: > > >Y1= a Y2+ b X1 > >Y2= c Y1+ d X2 > > >I first etimated these two equations using OLS (i.e. ignoring > >endogeneity). Next I estimated them using 2sls. Essentially what > >happened was the endogeneous variables (Y2,Y1) changed signs (went > >from positive significant to negative significant). I was wondering if > >this was cause for concern. Does it reflect some underlying problems > >with the data/ model? Or is this just an outcome of endogeneity? I > >should mention 3sls produced coefficients with the same signs as 2sls. > >Any insight would be much appreciated. > > If this is your model, there is no need to do anything > that difficult; just use X2 as the instrumental variable > for the first equation, and X1 for the second, or just > run the regression of Y1 and Y2 on X1 and X2 and transform > these regression equations to the desired form. This is > what 2sls and 3sls approximate when X1 and X2 have more > than one element. But in a just identified model, it > is not necessary to approximate. > > OLS is appropriate for prediction ONLY; that is, if you > will get the values of Y2 and X1, and what to predict Y1 > from those, use OLS. It can be far off when it comes to > the structural equations. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
