On Mon, 7 Mar 2011, Lemma Tenessa wrote: After estimating an ordered probit you can use the "fcast" command to generate a series of predictions for the ordinal dependent variable. I guess that in a two-stage procedure you would replace the actual ordinal values of your P and D variables with the first-stage predictions. However, I don't know if applying ordered probit again at a second step would give you any kind of consistent estimator. You may be better approaching this as a maximum likelihood problem (see gretl's "mle" command). What you're asking here is not really specific to gretl. You'll need to examine the literature on IV methods as applied to the sort of data you have. It looks as if these references might be useful: Andrew Chesher, "Instrumental Variable Models for Discrete Outcomes", http://www.econ.uiuc.edu/~roger/seminar/Chesher.pdf Skrondal, A. and Rabe-Hesketh, S. (2004). Generalized latent variable modeling: Multilevel, longitudinal and structural equation models (Chapman & Hall). Allin Cottrell