tag:blogger.com,1999:blog-6288862798546085706.post6935037766375022888..comments2022-05-23T14:51:12.138-04:00Comments on Econometrics By Simulation: Simulating Multinomial logit in Stata - UpdatedFrancishttp://www.blogger.com/profile/16658586705916884436noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-6288862798546085706.post-90994522673645265372013-06-13T22:13:59.900-04:002013-06-13T22:13:59.900-04:00Dear Mateusz,
Interesting point.. however, when I ...Dear Mateusz,<br />Interesting point.. however, when I tried the Gumbel variate as you suggest, my conditional logit model does not recover the parameters well.. in contrast, when using the normal distribution as provided by Francis, it works well. I guess this is because this step..<br /><br />gen runif = runiform()<br />gen flavor = 3<br />replace flavor = 2 if runif < pvan+pchoc<br />replace flavor = 1 if runif < pvan<br /><br />.. induces a Gumbel error distribution? Or am I missing something?<br />Cheers<br />TomAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-6288862798546085706.post-18431485390155248672013-05-16T16:20:20.863-04:002013-05-16T16:20:20.863-04:00Thanks a lot Mateusz! This comment really helped m...Thanks a lot Mateusz! This comment really helped me correct an error in my thinking about these types of problems.Francishttps://www.blogger.com/profile/16658586705916884436noreply@blogger.comtag:blogger.com,1999:blog-6288862798546085706.post-4540924233034911532013-05-16T07:49:06.622-04:002013-05-16T07:49:06.622-04:00This simulation (by drawing numbers from a normal ...This simulation (by drawing numbers from a normal (0,1) is equivalent to simulating Gumbel errors with what variance?Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-6288862798546085706.post-79281517010990165852012-11-20T04:12:21.739-05:002012-11-20T04:12:21.739-05:00Hi,
drawing the unobserved component from the nor...Hi, <br />drawing the unobserved component from the normal distribution while generating the XB's isn't really ok. I mean, you wouldn't see much difference, as the parameter are estimated up to scale anyway (that's the reason why, as you said: "the estimator does not work as well as expected in terms of absolute magnitude"), but the logit model implies that the unobserved part of the utility be distributed type I extreme value. So instead of normal distribution, I believe you should use: -ln(-ln(runiform())) which is the inverse of the CDF in the EV distribution. <br />Regards,<br />Mateuszmmhttps://www.blogger.com/profile/01766687929582218745noreply@blogger.comtag:blogger.com,1999:blog-6288862798546085706.post-57346167371220850572012-07-13T19:42:52.768-04:002012-07-13T19:42:52.768-04:00You just saved my day. Thank you so much for this ...You just saved my day. Thank you so much for this amazing blog, it's such an inspirational way to learn Stata (and to teach too).<br /><br />Greetings from Chile!Camilohttp://pesadillaenvancouver.wordpress.com/noreply@blogger.com