From Charles.Raux at laet.ish-lyon.cnrs.fr Wed Oct 3 19:50:14 2018 From: Charles.Raux at laet.ish-lyon.cnrs.fr (RAUX Charles) Date: Wed, 3 Oct 2018 09:50:14 +0000 Subject: [Limdep Nlogit List] R2 Mc Fadden and other R2 in NLOGIT? In-Reply-To: References: <3BE1FEAD-3100-496F-B0D2-BB6F3015B109@gmail.com> Message-ID: <72D6B3FEC474FA40B4981EDD2D672981F21E6829@CNMB02WVP.core-res.rootcore.local> Hello, I am running nlogit (MNL nlogit, scaled smnlogit, latent class lclogit) models which give different R2 outputs and I am wondering about the signification of these. I went through the documentation and did not find any suitable answer. We know that McFadden's pseudo R2 = 1 - logL/logL0 where L0 stands for all slopes are zero, i.e. model with a constant only. When estimating an mnl: |-> nlogit ; choices=air, car, bus, train, none ; lhs=choicem ; model: U(air, car, bus, train) = + dd * duration + pc * price / U(none) = anone ; show model$ ... I get Discrete choice (multinomial logit) model Dependent variable Choice Log likelihood function -1734.80120 Estimation based on N = 1758, K = 6 Inf.Cr.AIC = 3481.6 AIC/N = 1.980 --------------------------------------- Log likelihood R-sqrd R2Adj Constants only -2159.8972 .1968 .1961 Note: R-sqrd = 1 - logL/Logl(constants) ... There is no output explicitly named "Mc Fadden pseudo R2". Is "Constants only" the same as "McFadden Pseudo R-squared"? When I run smnlogit |-> smnlogit; choices=air, car, bus, train, none ; lhs=choicem ; model: U(air, car, bus, train) = + * duration + pc * price / U(none) = anone ; pds=6; pts=100; halton ; showmodel$ I get Scaled Multinomial Logit Model Dependent variable CHOICEM Log likelihood function -1667.35565 Restricted log likelihood -2829.39185 Chi squared [ 10](P= .000) 2324.07240 Significance level .00000 McFadden Pseudo R-squared .4107018 Estimation based on N = 1758, K = 10 Inf.Cr.AIC = 3354.7 AIC/N = 1.908 --------------------------------------- Log likelihood R-sqrd R2Adj No coefficients -2829.3919 .4107 .4099 Constants only -2159.8972 .2280 .2269 At start values -1720.2080 .0307 .0293 Note: R-sqrd = 1 - logL/Logl(constants) ... I see the "McFadden Pseudo R-squared" which has the same value as "No coefficients R2" (.4107) but not the same as "Constants only R2". Could anybody clarify about these R2? Thanks Charles Raux From wgreene at stern.nyu.edu Thu Oct 4 02:25:47 2018 From: wgreene at stern.nyu.edu (William Greene) Date: Wed, 3 Oct 2018 12:25:47 -0400 Subject: [Limdep Nlogit List] R2 Mc Fadden and other R2 in NLOGIT? In-Reply-To: <72D6B3FEC474FA40B4981EDD2D672981F21E6829@CNMB02WVP.core-res.rootcore.local> References: <3BE1FEAD-3100-496F-B0D2-BB6F3015B109@gmail.com> <72D6B3FEC474FA40B4981EDD2D672981F21E6829@CNMB02WVP.core-res.rootcore.local> Message-ID: Charles: In the case of the MNL, if it is possible to compute the log likelihood for the constants only model by using the sample proportions, nlogit will return the pseudo-Rsquared. No, it doesn't list McFadden's name, but that is the statistic, as the legend implies. Note, if the choice set is not the same for everyone, or if there are restrictions on the constant terms, then it is not possible to compute the statistic based only on the sample shares, and the step will be skipped. But, in this case, you can compute the constants only model with the right command, then obtain the statistic with a hand calculator. For involved models, such as the scaled MNL, you have to be very careful, because the sample shares can't be used to obtain lnL for a scaled MNL with constants only. So, you might not want to trust an internally generated value. Again, use the model command to obtain the value for the model you want to be the base. Regards Bill Greene On Wed, Oct 3, 2018 at 5:50 AM RAUX Charles < Charles.Raux at laet.ish-lyon.cnrs.fr> wrote: > Hello, > I am running nlogit (MNL nlogit, scaled smnlogit, latent class lclogit) > models which give different R2 outputs and I am wondering about the > signification of these. I went through the documentation and did not find > any suitable answer. > We know that > McFadden's pseudo R2 = 1 - logL/logL0 > where L0 stands for all slopes are zero, i.e. model with a constant only. > > When estimating an mnl: > > |-> nlogit ; choices=air, car, bus, train, none > ; lhs=choicem > ; model: U(air, car, bus, train) = > + dd * duration + pc * price > / U(none) = anone > ; show model$ > ... > > I get > > Discrete choice (multinomial logit) model > Dependent variable Choice > Log likelihood function -1734.80120 > Estimation based on N = 1758, K = 6 > Inf.Cr.AIC = 3481.6 AIC/N = 1.980 > --------------------------------------- > Log likelihood R-sqrd R2Adj > Constants only -2159.8972 .1968 .1961 > Note: R-sqrd = 1 - logL/Logl(constants) > ... > > There is no output explicitly named "Mc Fadden pseudo R2". > Is "Constants only" the same as "McFadden Pseudo R-squared"? > > When I run smnlogit > > |-> smnlogit; choices=air, car, bus, train, none > ; lhs=choicem > ; model: U(air, car, bus, train) = > + * duration + pc * price > / U(none) = anone > ; pds=6; pts=100; halton > ; showmodel$ > > I get > > Scaled Multinomial Logit Model > Dependent variable CHOICEM > Log likelihood function -1667.35565 > Restricted log likelihood -2829.39185 > Chi squared [ 10](P= .000) 2324.07240 > Significance level .00000 > McFadden Pseudo R-squared .4107018 > Estimation based on N = 1758, K = 10 > Inf.Cr.AIC = 3354.7 AIC/N = 1.908 > --------------------------------------- > Log likelihood R-sqrd R2Adj > No coefficients -2829.3919 .4107 .4099 > Constants only -2159.8972 .2280 .2269 > At start values -1720.2080 .0307 .0293 > Note: R-sqrd = 1 - logL/Logl(constants) > ... > > I see the "McFadden Pseudo R-squared" which has the same value as "No > coefficients R2" (.4107) but not the same as "Constants only R2". > Could anybody clarify about these R2? > Thanks > Charles Raux > > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > > -- William Greene Department of Economics Stern School of Business, New York University 44 West 4 St., 7-90 New York, NY, 10012 URL: https://protect-au.mimecast.com/s/uxGMCq7BKYtDR9MqCZQu9k?domain=people.stern.nyu.edu Email: wgreene at stern.nyu.edu Ph. +1.212.998.0876 Editor in Chief: Journal of Productivity Analysis Editor in Chief: Foundations and Trends in Econometrics Associate Editor: Economics Letters Associate Editor: Journal of Business and Economic Statistics Associate Editor: Journal of Choice Modeling