[Gretl-devel] ordered logit/probit

On Mon, 14 Sep 2009, Hamaad Shah wrote: You're right, there was a problem here (now fixed in CVS and the Windows snapshot). The $yhat accessor produces correct values for z-hat_i, the estimate of the latent variable at each observation (as described in chapter 24 of the User's Guide). And from these values one can compute predictions for the response variable that are in agreement with R. BUT two other things were wrong: (a) The count of "cases correctly predicted" (printed beneath the model) was not done in the standard manner, where predicted y is the response-value with the greatest estimated probability. (b) The values you'd get from the "fcast" command for ordered logit and probit were, unfortunately, rubbish. I'm afraid we never noticed that we were generating "forecasts" as if the model were a regular binary logit/probit. The results from gretl's "fcast" for an ordered model should now agree with what R says [if you use predict() after polr()]. Note that correct predictions based on gretl's $yhat can be constructed as in the following example (though now that fcast is working correctly you don't have to do this): <script> open greene22_2 genr rat = Z8 - 1 probit rat 0 Z1 Z2 series zhat = $yhat matrix cut = $coeff[3:] scalar n = $nobs scalar ncut = rows(cut) series pred = NA loop i=1..n --quiet # find the response with the greatest probability # for the probit case pmax = cdf(N, cut[1]-zhat[i]) cpmax = 1 loop j=2..ncut --quiet prob = cdf(N, cut[j]-zhat[i]) - cdf(N, cut[j-1]-zhat[i]) if (prob > pmax) pmax = prob cpmax = j endif endloop prob = 1 - cdf(N, cut[ncut]-zhat[i]) if (prob > pmax) cpmax = ncut+1 endif pred[i] = cpmax-1 endloop print rat pred --byobs </script> The series "pred" above is what "fcast" now produces. Allin Cottrell