Sure. Basically the procedure works like this. 1. Optimize to get the MLEs and the Hessian. 2. Take the negative-inverse of the Hessian to get the variance co-variance matrix. 3. Simulate N draws from the multivariate normal centered at the MLE using the variance-covariance matrix as Sigma. 4. Set two sets of X values (we'll call them X1 and X2) as well as two sets of Z values (Z1 and Z2) 5. For each of the N draws: 5a. Convert from the parameters beta and gamma to the natural parameters pi and lambda by pushing everything through the systematic component. 5b. You should now have 4 numbers: pi_1, pi_2, lambda_1, lambda_2 5c. Use the formula for the expectation to calculate the first difference: E(Y|pi_2, lambda_2) - E(Y|pi_1, lambda_1) 6. Now you have N first differences. This is the sampling distribution. 7. Report the mean of your distribution as the point estimate, the standard deviation as the SE, and any quantile as a confidence interval. Brandon On Wed, Mar 30, 2011 at 3:26 PM, Yanfang Su <suyanfang(a)gmail.com> wrote: _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l