From michael.burton at uwa.edu.au Mon Jan 17 15:08:31 2022
From: michael.burton at uwa.edu.au (Michael Burton)
Date: Mon, 17 Jan 2022 04:08:31 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
Message-ID:
I am wondering how the ?worst? choices are dealt with in estimation. Is the ?reverse? logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual ?stacked? approach taken, where marginal utilities of the ?worst? are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
From teagle at tceagle.com Tue Jan 18 04:55:08 2022
From: teagle at tceagle.com (Thomas Eagle)
Date: Mon, 17 Jan 2022 17:55:08 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To:
References:
Message-ID: <913d1861589c477aaeec50eb644082fb@tceagle.com>
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/EWVSCr81nyt8qMyQKi7LYUQ?domain=limdep.itls.usyd.edu.au
From teagle at tceagle.com Tue Jan 18 04:55:08 2022
From: teagle at tceagle.com (Thomas Eagle)
Date: Mon, 17 Jan 2022 17:55:08 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To:
References:
Message-ID: <913d1861589c477aaeec50eb644082fb@tceagle.com>
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/1DgNCwV1vMfGj65lquVM8FG?domain=limdep.itls.usyd.edu.au
From michael.burton at uwa.edu.au Tue Jan 18 12:28:52 2022
From: michael.burton at uwa.edu.au (Michael Burton)
Date: Tue, 18 Jan 2022 01:28:52 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To: <913d1861589c477aaeec50eb644082fb@tceagle.com>
References:
<913d1861589c477aaeec50eb644082fb@tceagle.com>
Message-ID:
Tom
That's not my problem.
If you estimate the 'stacked' model (Louviere et al p36-37) you are estimating a model that internally inconsistent, by construction. That's because the gumbel error term isn't symmetric. Where that's recognised (e.g Flynn et al, 2008, BMC Medical Research Methodology) its not seen as a big enough issue to prevent the approach being used.
Delle Site et al (2019) set out the theory (DOI: 10.1016/j.trb.2019.07.014) and a reading of that paper suggests the Nlogit6 can estimate it correctly (i.e. that the probabilities for the worst choices are defined by a 'reverse' logit). So the purpose of my question is to check if the best/worst command in Nlogit6 actually does that.
Michael
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Thomas Eagle
Sent: Tuesday, 18 January 2022 1:55 AM
To: Limdep and Nlogit Mailing List ; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/4L7PCP7LAXfKwxLgxU0ASsg?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/4L7PCP7LAXfKwxLgxU0ASsg?domain=aus01.safelinks.protection.outlook.com
From michael.burton at uwa.edu.au Tue Jan 18 12:28:52 2022
From: michael.burton at uwa.edu.au (Michael Burton)
Date: Tue, 18 Jan 2022 01:28:52 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To: <913d1861589c477aaeec50eb644082fb@tceagle.com>
References:
<913d1861589c477aaeec50eb644082fb@tceagle.com>
Message-ID:
Tom
That's not my problem.
If you estimate the 'stacked' model (Louviere et al p36-37) you are estimating a model that internally inconsistent, by construction. That's because the gumbel error term isn't symmetric. Where that's recognised (e.g Flynn et al, 2008, BMC Medical Research Methodology) its not seen as a big enough issue to prevent the approach being used.
Delle Site et al (2019) set out the theory (DOI: 10.1016/j.trb.2019.07.014) and a reading of that paper suggests the Nlogit6 can estimate it correctly (i.e. that the probabilities for the worst choices are defined by a 'reverse' logit). So the purpose of my question is to check if the best/worst command in Nlogit6 actually does that.
Michael
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Thomas Eagle
Sent: Tuesday, 18 January 2022 1:55 AM
To: Limdep and Nlogit Mailing List ; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/sGKQC2xMQzipYwD9Kc1T8OA?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/sGKQC2xMQzipYwD9Kc1T8OA?domain=aus01.safelinks.protection.outlook.com
From teagle at tceagle.com Tue Jan 18 15:39:08 2022
From: teagle at tceagle.com (Thomas Eagle)
Date: Tue, 18 Jan 2022 04:39:08 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To:
References:
<913d1861589c477aaeec50eb644082fb@tceagle.com>
Message-ID: <4f36ec434a004e66bf21c54009c501da@tceagle.com>
I think you'll find NLOGIT estimates it as I described as the best-worst option... but Bill Greene can explain it better than me. That is why I referred you to Louviere, Flynn, and Marley (2015) Best-Worst Scaling: Theory, Methods, and Applications. They lay out in more detail what Flynn, Louviere, et al wrote in 2008.
Just trying to help...
Tom
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Monday, January 17, 2022 5:29 PM
To: Limdep and Nlogit Mailing List ; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
Tom
That's not my problem.
If you estimate the 'stacked' model (Louviere et al p36-37) you are estimating a model that internally inconsistent, by construction. That's because the gumbel error term isn't symmetric. Where that's recognised (e.g Flynn et al, 2008, BMC Medical Research Methodology) its not seen as a big enough issue to prevent the approach being used.
Delle Site et al (2019) set out the theory (DOI: 10.1016/j.trb.2019.07.014) and a reading of that paper suggests the Nlogit6 can estimate it correctly (i.e. that the probabilities for the worst choices are defined by a 'reverse' logit). So the purpose of my question is to check if the best/worst command in Nlogit6 actually does that.
Michael
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au > On Behalf Of Thomas Eagle
Sent: Tuesday, 18 January 2022 1:55 AM
To: Limdep and Nlogit Mailing List >; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au > On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/RaKRCE8wmrt37rDEwtwcFp3?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/RaKRCE8wmrt37rDEwtwcFp3?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/QwiQCGv0oyC17xMvrSpSdwl?domain=limdep.itls.usyd.edu.au
From teagle at tceagle.com Tue Jan 18 15:39:08 2022
From: teagle at tceagle.com (Thomas Eagle)
Date: Tue, 18 Jan 2022 04:39:08 +0000
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
In-Reply-To:
References:
<913d1861589c477aaeec50eb644082fb@tceagle.com>
Message-ID: <4f36ec434a004e66bf21c54009c501da@tceagle.com>
I think you'll find NLOGIT estimates it as I described as the best-worst option... but Bill Greene can explain it better than me. That is why I referred you to Louviere, Flynn, and Marley (2015) Best-Worst Scaling: Theory, Methods, and Applications. They lay out in more detail what Flynn, Louviere, et al wrote in 2008.
Just trying to help...
Tom
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au On Behalf Of Michael Burton
Sent: Monday, January 17, 2022 5:29 PM
To: Limdep and Nlogit Mailing List ; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
Tom
That's not my problem.
If you estimate the 'stacked' model (Louviere et al p36-37) you are estimating a model that internally inconsistent, by construction. That's because the gumbel error term isn't symmetric. Where that's recognised (e.g Flynn et al, 2008, BMC Medical Research Methodology) its not seen as a big enough issue to prevent the approach being used.
Delle Site et al (2019) set out the theory (DOI: 10.1016/j.trb.2019.07.014) and a reading of that paper suggests the Nlogit6 can estimate it correctly (i.e. that the probabilities for the worst choices are defined by a 'reverse' logit). So the purpose of my question is to check if the best/worst command in Nlogit6 actually does that.
Michael
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au > On Behalf Of Thomas Eagle
Sent: Tuesday, 18 January 2022 1:55 AM
To: Limdep and Nlogit Mailing List >; limdep at limdep.itls.usyd.edu.au
Subject: Re: [Limdep Nlogit List] best/worst in NLOGIT6
There are several ways to do this. You should look at Louviere's book on Best-Worst modeling to see some examples. The MaxDiff approach is to generate tasks for the best choice and generate tasks for the worst choice. You use dummy or effects coding for the items in the tasks. The best tasks are all coded normally. The worst tasks have all their dummy or effects coded values multiplied by -1. If you wish to use the best worst coding then the best alternative is dropped from the generated worst tasks.
Tom Eagle
-----Original Message-----
From: limdep-bounces at mailman.sydney.edu.au > On Behalf Of Michael Burton
Sent: Sunday, January 16, 2022 8:09 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] best/worst in NLOGIT6
I am wondering how the 'worst' choices are dealt with in estimation. Is the 'reverse' logit applied to the worst choices, so the error term is multiplied by -1 and the best and worst choice specification is strictly consistent, or is the more usual 'stacked' approach taken, where marginal utilities of the 'worst' are flipped, but not the error term?
Thanks
Michael
Sent from Mail for Windows
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/uttaCMwGxOtqrKZN3Tke0e5?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/uttaCMwGxOtqrKZN3Tke0e5?domain=aus01.safelinks.protection.outlook.com
_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/AVWUCNLJyQU0WKB43Fjm8k1?domain=limdep.itls.usyd.edu.au
From annika.tienhaara at luke.fi Wed Jan 26 00:38:00 2022
From: annika.tienhaara at luke.fi (Tienhaara Annika (LUKE))
Date: Tue, 25 Jan 2022 13:38:00 +0000
Subject: [Limdep Nlogit List] Log normally distributed cost parameter
Message-ID:
Dear all,
We are estimating random parameters (mixed) logit models with environmental attributes following a normal distribution and the cost parameter lognormal distribution. How can we calculate the means and standard errors of the marginal WTPs for the attributes using Nlogit, based on the unconditional estimates? We have looked into simulating the WTP values for the population, but have not found instructions or examples on how to do this when the cost is lognormal. Estimation in WTP space is not possible, as there we have problems with model convergence.
Here is our model and results:
|-> NLOGIT;Lhs=CHOICE
;Choices=BAU,ALT1,ALT2
;Halton
;rpl
;pts=2000
;pds=6
;Fcn=a2(n),a3(n),b1(n),b2(n),b3(n),b4(n),b5(n),b6(n),b7(n),b8(n),b9(n),b10(n),b11(l)
;parameters
;Model:
U(BAU)=b1*p30+b2*p60+b3*e2+b4*e3+b5*k4+b6*k5+b7*i10+b8*i30+b9*v70+b10*v80+b11*mpc/
U(ALT1)=a2+b1*p30+b2*p60+b3*e2+b4*e3+b5*k4+b6*k5+b7*i10+b8*i30+b9*v70+b10*v80+b11*mpc/
U(ALT2)=a3+b1*p30+b2*p60+b3*e2+b4*e3+b5*k4+b6*k5+b7*i10+b8*i30+b9*v70+b10*v80+b11*mpc$
---------------------------------------------------------------------------
Random Parameters Multinom. Logit Model
Dependent variable CHOICE
Log likelihood function -2462.16944
Restricted log likelihood -3401.30365
Chi squared [ 26](P= .000) 1878.26841
Significance level .00000
McFadden Pseudo R-squared .2761101
Estimation based on N = 3096, K = 26
Inf.Cr.AIC = 4976.3 AIC/N = 1.607
---------------------------------------
Log likelihood R-sqrd R2Adj
No coefficients -3401.3036 .2761 .2731
Constants only -3228.5985 .2374 .2342
At start values -3003.1598 .1801 .1767
Note: R-sqrd = 1 - logL/Logl(constants)
---------------------------------------
Response data are given as ind. choices
Replications for simulated probs. =2000
Used Halton sequences in simulations.
RPL model with panel has 516 groups
Fixed number of obsrvs./group= 6
BHHH estimator used for asymp. variance
Number of obs.= 3096, skipped 0 obs
--------+------------------------------------------------------------------
| Standard Prob. 95% Confidence
CHOICE| Coefficient Error z |z|>Z* Interval
--------+------------------------------------------------------------------
|Random parameters in utility functions..............................
A2| 2.13805*** .38031 5.62 .0000 1.39265 2.88344
A3| 1.34720*** .36737 3.67 .0002 .62718 2.06723
B1| .70033*** .16475 4.25 .0000 .37743 1.02323
B2| .73475*** .16741 4.39 .0000 .40663 1.06288
B3| .20620 .15004 1.37 .1694 -.08788 .50028
B4| .73427*** .16651 4.41 .0000 .40792 1.06062
B5| -.19350 .14221 -1.36 .1736 -.47223 .08523
B6| .08965 .13512 .66 .5070 -.17519 .35448
B7| .69951*** .17191 4.07 .0000 .36256 1.03645
B8| .64813*** .17595 3.68 .0002 .30329 .99298
B9| .31783* .16785 1.89 .0583 -.01116 .64681
B10| .68545*** .17767 3.86 .0001 .33723 1.03367
B11| -4.10386*** .14879 -27.58 .0000 -4.39548 -3.81224
|Distns. of RPs. Std.Devs or limits of triangular....................
NsA2| .80219*** .22657 3.54 .0004 .35811 1.24627
NsA3| 1.29128*** .18269 7.07 .0000 .93322 1.64934
NsB1| 1.18243*** .19424 6.09 .0000 .80172 1.56314
NsB2| 1.16108*** .25288 4.59 .0000 .66544 1.65672
NsB3| .93324*** .23092 4.04 .0001 .48064 1.38584
NsB4| .95523*** .23226 4.11 .0000 .50001 1.41046
NsB5| .51003* .30579 1.67 .0953 -.08931 1.10936
NsB6| .88595*** .20339 4.36 .0000 .48731 1.28458
NsB7| .72713*** .24920 2.92 .0035 .23871 1.21555
NsB8| 1.33430*** .19572 6.82 .0000 .95070 1.71790
NsB9| .93112*** .24078 3.87 .0001 .45919 1.40305
NsB10| 1.01326*** .21938 4.62 .0000 .58328 1.44324
LsB11| 2.10463*** .16179 13.01 .0000 1.78753 2.42173
--------+------------------------------------------------------------------
Thank you in advance!
Best regards
Annika