From djourdain at ait.asia Thu Feb 15 21:26:52 2018 From: djourdain at ait.asia (Damien Jourdain) Date: Thu, 15 Feb 2018 12:26:52 +0200 Subject: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? Message-ID: <009301d3a647$8264f6a0$872ee3e0$@cirad.fr> Dear All, I developing a RPL model using choice experiment data The model is as followed: Calc; Ran(1234567)$ RPLOGIT ; Choices = 1,2,3 ; Lhs = CHOICE, CSET, ALT ; Rhs = L_IMP, L_RAU, L_GAP, L_PGS, FRESH, O_SUPE, O_SPEC, NPRICE ; Fcn = L_IMP(n), L_RAU(n), L_GAP(n), L_PGS(n), FRESH(n), O_SUPE(n), O_SPEC(n), NPRICE(l) ; Halton ; Pds = csi ; Pts = 20 ; Parameters ; Maxit = 150$ I have changed the price attribute to negative values, so I can use a lognormal distribution of for the price attribute. I am getting the following results --------+------------------------------------------------------------------- - | Standard Prob. 95% Confidence CHOICE| Coefficient Error z |z|>Z* Interval --------+------------------------------------------------------------------- - |Random parameters in utility functions.............................. L_IMP| -1.11608*** .29152 -3.83 .0001 -1.68745 -.54470 L_RAU| 1.49941*** .09880 15.18 .0000 1.30577 1.69304 L_GAP| 1.82794*** .10487 17.43 .0000 1.62239 2.03349 L_PGS| .63730** .25734 2.48 .0133 .13291 1.14168 FRESH| -.61318*** .05496 -11.16 .0000 -.72089 -.50546 O_SUPE| .43891*** .07567 5.80 .0000 .29060 .58721 O_SPEC| -.76256*** .17329 -4.40 .0000 -1.10221 -.42291 NPRICE| -1.61991*** .33228 -4.88 .0000 -2.27117 -.96865 |Distns. of RPs. Std.Devs or limits of triangular.................... NsL_IMP| 1.52346*** .31666 4.81 .0000 .90281 2.14410 NsL_RAU| .69380*** .13439 5.16 .0000 .43040 .95721 NsL_GAP| .01744 .24287 .07 .9427 -.45858 .49346 NsL_PGS| .95598*** .21017 4.55 .0000 .54405 1.36790 NsFRESH| .48681*** .05657 8.60 .0000 .37593 .59770 NsO_SUPE| 1.65455*** .11307 14.63 .0000 1.43293 1.87616 NsO_SPEC| 1.08890*** .12068 9.02 .0000 .85237 1.32544 LsNPRICE| .99479*** .18655 5.33 .0000 .62917 1.36041 If I am not wrong, I can calculate the population mean of the price E(beta) = exp(beta + sigma^2/2) |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ [CALC] = .3246179 Then, I am using the procedure described in section N29.8.2 of the Nlogit manual to examine the distribution of the parameters. Matrix; bn = beta_i; sn =sdbeta_i $ CREATE; BIMP=0; BRAU=0; BGAP=0; BPGS=0; BFRE =0; BSUP=0; BSPE=0; BNPR=0 $ CREATE ;SIMP=0; SRAU=0; SGAP=0; SPGS=0; SFRE =0; SSUP=0; SSPE=0; SNPR=0 $ NAMELIST; betan = BIMP,BRAU, BGAP, BPGS, BFRE, BSUP, BSPE, BNPR$ NAMELIST; sbetan = SIMP,SRAU, SGAP, SPGS, SFRE, SSUP, SSPE, SNPR$ CREATE ; betan =bn$ CREATE ; sbetan = sn$ CALC; List; XBR(BNPR)$ ? calculate the average of the beta for nprice |-> CALC; List; XBR(BNPR)$ [CALC] = .0257560 My understanding is that these two figures should be close to one another. Is there anything that could explain such difference between these two ways to estimate the results? Any help is welcomed? Best, Damien From djourdain at ait.asia Thu Feb 15 22:41:16 2018 From: djourdain at ait.asia (Damien Jourdain) Date: Thu, 15 Feb 2018 13:41:16 +0200 Subject: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? In-Reply-To: <60299233-429c-049c-738d-e89661957fb7@uw.edu.pl> References: <009301d3a647$8264f6a0$872ee3e0$@cirad.fr> <60299233-429c-049c-738d-e89661957fb7@uw.edu.pl> Message-ID: <00ac01d3a651$e6e119b0$b4a34d10$@cirad.fr> Dear Mik, Thank you for the suggestion. I tried that but there is still an important difference between the two. |-> create; expBNPR = exp(BNPR)$ |-> calc; list; xbr(expBNPR)$ [CALC] = 1.0329355 As the direct estimation from the coefficients is giving : > |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > [CALC] = .3246179 By the way, the more 'realistic' calculation is the direct estimation from the coefficients (at least it is of the same magnitude than the MNL coefficient for price) Damien -----Message d'origine----- De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, February 15, 2018 12:34 PM ? : Damien Jourdain Objet : Re: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? Dear Damien, Shoulnd't you have something like: create; expBNPR = exp(BNPR)$ first? Then calc; list; xbr(expBNPR)$ Cheers, Mik On 2018-02-15 11:26, Damien Jourdain wrote: > Dear All, > > > > I developing a RPL model using choice experiment data > > > > The model is as followed: > > > > Calc; Ran(1234567)$ > > RPLOGIT > > ; Choices = 1,2,3 > > ; Lhs = CHOICE, CSET, ALT > > ; Rhs = L_IMP, L_RAU, L_GAP, L_PGS, > > FRESH, O_SUPE, O_SPEC, NPRICE > > ; Fcn = L_IMP(n), L_RAU(n), L_GAP(n), > > L_PGS(n), FRESH(n), O_SUPE(n), O_SPEC(n), NPRICE(l) > > ; Halton > > ; Pds = csi > > ; Pts = 20 > > ; Parameters > > ; Maxit = 150$ > > > > I have changed the price attribute to negative values, so I can use a > lognormal distribution of for the price attribute. > > I am getting the following results > > > > --------+------------------------------------------------------------- > --------+------ > - > > | Standard Prob. 95% Confidence > > CHOICE| Coefficient Error z |z|>Z* Interval > > --------+------------------------------------------------------------- > --------+------ > - > > |Random parameters in utility > functions.............................. > > L_IMP| -1.11608*** .29152 -3.83 .0001 -1.68745 -.54470 > > L_RAU| 1.49941*** .09880 15.18 .0000 1.30577 1.69304 > > L_GAP| 1.82794*** .10487 17.43 .0000 1.62239 2.03349 > > L_PGS| .63730** .25734 2.48 .0133 .13291 1.14168 > > FRESH| -.61318*** .05496 -11.16 .0000 -.72089 -.50546 > > O_SUPE| .43891*** .07567 5.80 .0000 .29060 .58721 > > O_SPEC| -.76256*** .17329 -4.40 .0000 -1.10221 -.42291 > > NPRICE| -1.61991*** .33228 -4.88 .0000 -2.27117 -.96865 > > |Distns. of RPs. Std.Devs or limits of > triangular.................... > > NsL_IMP| 1.52346*** .31666 4.81 .0000 .90281 2.14410 > > NsL_RAU| .69380*** .13439 5.16 .0000 .43040 .95721 > > NsL_GAP| .01744 .24287 .07 .9427 -.45858 .49346 > > NsL_PGS| .95598*** .21017 4.55 .0000 .54405 1.36790 > > NsFRESH| .48681*** .05657 8.60 .0000 .37593 .59770 > > NsO_SUPE| 1.65455*** .11307 14.63 .0000 1.43293 1.87616 > > NsO_SPEC| 1.08890*** .12068 9.02 .0000 .85237 1.32544 > > LsNPRICE| .99479*** .18655 5.33 .0000 .62917 1.36041 > > > > If I am not wrong, I can calculate the population mean of the price > E(beta) = exp(beta + sigma^2/2) > > |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > > [CALC] = .3246179 > > > > Then, I am using the procedure described in section N29.8.2 of the > Nlogit manual to examine the distribution of the parameters. > > Matrix; bn = beta_i; sn =sdbeta_i $ > > CREATE; BIMP=0; BRAU=0; BGAP=0; BPGS=0; BFRE =0; BSUP=0; BSPE=0; > BNPR=0 $ > > CREATE ;SIMP=0; SRAU=0; SGAP=0; SPGS=0; SFRE =0; SSUP=0; SSPE=0; > SNPR=0 $ > > NAMELIST; betan = BIMP,BRAU, BGAP, BPGS, BFRE, BSUP, BSPE, BNPR$ > > NAMELIST; sbetan = SIMP,SRAU, SGAP, SPGS, SFRE, SSUP, SSPE, SNPR$ > > CREATE ; betan =bn$ > > CREATE ; sbetan = sn$ > > > > CALC; List; XBR(BNPR)$ ? calculate the average of the beta for nprice > > > > |-> CALC; List; XBR(BNPR)$ > > [CALC] = .0257560 > > > > My understanding is that these two figures should be close to one another. > Is there anything that could explain such difference between these two > ways to estimate the results? > > > > Any help is welcomed? > > > > Best, > > > > Damien > > > > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > From mrentrena at uco.es Fri Feb 16 00:32:26 2018 From: mrentrena at uco.es (=?UTF-8?Q?Macario_RODR=c3=8dGUEZ-ENTRENA?=) Date: Thu, 15 Feb 2018 14:32:26 +0100 Subject: [Limdep Nlogit List] WTP space estimation issue Message-ID: Dear Professor Bill and Nlogit users: I am experiencing some issues regarding a WTP space modelling. In the first stage, I estimated the model using a standard RPL model with one interaction (in the mean) and everything is fine since the results are quite informative for policy making. Nonetheless, to avoid known issues related to assuring finite moments for WTP estimations I am trying to estimate de model with such interaction in WTP space. When I do that the results are very inconsistent and changeable. Any suggestion would be very welcome since I am disconcerted. Here I show the inconsistencies (from my point of view): Thank you so much in advance Any suggestion will be highly appreciated Macario -------- RPL Output RPLOGIT ; Lhs = ELE ; cHOICES =A,B,SQ ; RPL=DENS ; FCN=EMI500(N),EMI700(N),ERO14(N),ERO28(N),BIO15(N),BIO20(N),EURC[N] ; MODEL: U(A,B,SQ) = <0,0,ASCSQ>+EMI500*EMI500+EMI700*EMI700+ERO14*ERO14+ ERO28*ERO28+BIO15*BIO15+BIO20*BIO20+EURC*EURC ; pwt; Halton ; pts=1000 ; ECM=(A,B) ; pds=PDS ; PAR; margin$ Iterative procedure has converged Normal exit:64 iterations. Status=0, F=.2066905D+04 --------+-------------------------------------------------------------------- |StandardProb.95% Confidence ELE|CoefficientErrorz|z|>Z*Interval --------+-------------------------------------------------------------------- |Random parameters in utility functions.......................... EMI500|2.06088***.243398.47.00001.583842.53793 EMI700|3.26465***.2594912.58.00002.756063.77324 ERO14|1.64629***.246196.69.00001.163762.12882 ERO28|2.90635***.2586211.24.00002.399463.41324 BIO15|.72678***.243842.98.0029.248861.20470 BIO20|.69049***.249902.76.0057.200691.18030 EURC|-.26158***.01361-19.22.0000-.28825-.23492 |Nonrandom parameters in utility functions....................... ASCSQ|1.75895***.355814.94.00001.061582.45632 |Heterogeneity in mean, Parameter:Variable....................... EMI5:DEN|-.71829***.26300-2.73.0063-1.23376-.20282 EMI7:DEN|-.85872***.23874-3.60.0003-1.32664-.39079 ERO1:DEN|-.11006.25682-.43.6683-.61341.39330 ERO2:DEN|-.43107*.22683-1.90.0574-.87565.01351 BIO1:DEN|.17070.22053.77.4389-.26152.60292 BIO2:DEN|.38344.242321.58.1136-.09150.85838 EURC:DEN|0.0.....(Fixed Parameter)..... |Distns. of RPs. Std.Devs or limits of triangular................ NsEMI500|.06975.56176.12.9012-1.031291.17079 NsEMI700|.64275**.296852.17.0304.060941.22457 NsERO14|.57823**.274212.11.0350.040801.11567 NsERO28|.97815***.204944.77.0000.576471.37983 NsBIO15|.50831*.269811.88.0596-.020501.03713 NsBIO20|.88335***.215324.10.0000.461331.30536 NsEURC|.16707***.0141911.77.0000.13926.19488 |Standard deviations of latent random effects.................... SigmaE01|-4.98259***.39175-12.72.0000-5.75041-4.21477 --------+-------------------------------------------------------------------- GMX Output GMXLOGIT ; userp ; Lhs = ELE ; cHOICES = A,B,SQ ; GMX = DENS ; FCN=EMI500(N),EMI700(N),ERO14(N),ERO28(N),BIO15(N),BIO20(N),EURC[*N] ; MODEL: U(A,B,SQ) = <0,0,ASCSQ>+EMI500*EMI500+EMI700*EMI700+ERO14*ERO14+ ERO28*ERO28+BIO15*BIO15+BIO20*BIO20+EURC*EURC ; pwt; Halton ; pts=1000 ; ECM=(A,B) ; pds=PDS ; PAR; margin$ Iterative procedure has converged Normal exit:89 iterations. Status=0, F=.2518070D+04 --------+-------------------------------------------------------------------- |StandardProb.95% Confidence ELE|CoefficientErrorz|z|>Z*Interval --------+-------------------------------------------------------------------- |Random parameters in utility functions.......................... EMI500|.09175.27163.34.7355-.44064.62414 EMI700|.02970.18100.16.8697-.32506.38446 ERO14|.11045.37497.29.7683-.62448.84538 ERO28|.09214.19681.47.6396-.29359.47788 BIO15|-.00831.29531-.03.9776-.58710.57049 BIO20|.09172.27327.34.7371-.44388.62733 EURC|1.0.....(Fixed Parameter)..... |Nonrandom parameters in utility functions....................... ASCSQ|-2.40941***.09551-25.23.0000-2.59661-2.22221 |Heterogeneity in mean, Parameter:Variable....................... EMI5:DEN|.21937.43417.51.6134-.631601.07034 EMI7:DEN|.21808.31480.69.4885-.39892.83508 ERO1:DEN|.30811.53738.57.5664-.745141.36135 ERO2:DEN|.10758.27105.40.6914-.42368.63883 BIO1:DEN|.50247.51724.97.3313-.511291.51623 BIO2:DEN|.26820.44615.60.5477-.606231.14263 EURC:DEN|-.00690.36556-.02.9849-.72338.70958 |Distns. of RPs. Std.Devs or limits of triangular................ NsEMI500|.0003512.69377.00 1.0000-24.8789824.87968 NsEMI700| .002145.19543.00.9997-10.1807210.18500 NsERO14|.005429.71863.00.9996-19.0427419.05358 NsERO28|.001816.84284.00.9998-13.4099113.41354 NsBIO15|.002237.20629.00.9998-14.1218414.12631 NsBIO20|.005363.29627.00.9987-6.455226.46594 CsEURC|0.0.....(Fixed Parameter)..... |Variance parameter tau in GMX scale parameter................... TauScale|3.36068***.3161810.63.00002.740983.98038 |Weighting parameter gamma in GMX model.......................... GammaMXL|0.0.....(Fixed Parameter)..... |Coefficient on EURCin preference space form................ Beta0WTP|-177.918200.3959-.89.3746-570.687214.851 S_b0_WTP|16.3943366.1541.04.9643-701.2545734.0430 |Sample MeanSample Std.Dev................................. Sigma(i)|.351002.53852.14.8900-4.624415.32641 --------+-------------------------------------------------------------------- -- From djourdain at ait.asia Fri Feb 16 01:33:20 2018 From: djourdain at ait.asia (Damien Jourdain) Date: Thu, 15 Feb 2018 16:33:20 +0200 Subject: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? In-Reply-To: <0a52bfb4-457f-3907-8e40-d9eac46008e2@uw.edu.pl> References: <009301d3a647$8264f6a0$872ee3e0$@cirad.fr> <60299233-429c-049c-738d-e89661957fb7@uw.edu.pl> <00ac01d3a651$e6e119b0$b4a34d10$@cirad.fr> <0a52bfb4-457f-3907-8e40-d9eac46008e2@uw.edu.pl> Message-ID: <00f001d3a669$f0620950$d1261bf0$@cirad.fr> Dear Mik, Thank you. I've looked again, and I think found the mistake I made. When creating the variables from the matrix, I forgot to add the "Sample; 1-1400". By failing to do so, I suppose the calculation of the average include all the zeros for the variable BNPR (from 1401 to 13500 ... since I have 13500 rows). This result in an average that is much smaller than the reality! I am now adding the following line Sample ; 1-1400$ before getting the parameters from the matrix When I tried again with this statement, I find the calculated from direct estimation being quite close to the average of posterior individual-specific estimates. If that is correct, there is no need to use the exponential of the coefficients. |-> CALC; List; XBR(BNPR)$ [CALC] = .3199974 |-> create; expBNPR = exp(BNPR)$ |-> calc; list; xbr(expBNPR)$ [CALC] = 1.6429415 |-> CALC; LIST; exp(-1.95961 + (1.31682^2)/2)$ [CALC] = .3353426 Again, thank you for your help and interest. Damien -----Message d'origine----- De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, February 15, 2018 3:21 PM ? : Damien Jourdain Objet : Re: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? Dear Damien, This is what I would expect - direct estimation using coefficients is likely always better than the one based on posterior individual-specific estimates (even though asymptotically they should be equivalent). Best regards, Mik On 2018-02-15 12:41, Damien Jourdain wrote: > Dear Mik, > > Thank you for the suggestion. > > I tried that but there is still an important difference between the two. > > |-> create; expBNPR = exp(BNPR)$ > |-> calc; list; xbr(expBNPR)$ > [CALC] = 1.0329355 > > As the direct estimation from the coefficients is giving : >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ >> [CALC] = .3246179 > By the way, the more 'realistic' calculation is the direct estimation > from the coefficients (at least it is of the same magnitude than the > MNL coefficient for price) > > > Damien > > > -----Message d'origine----- > > De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, > February 15, 2018 12:34 PM ? : Damien Jourdain Objet : Re: [Limdep > Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? > > > Dear Damien, > > Shoulnd't you have something like: > > create; expBNPR = exp(BNPR)$ > > first? > > Then > calc; list; xbr(expBNPR)$ > > Cheers, > Mik > > > > > On 2018-02-15 11:26, Damien Jourdain wrote: >> Dear All, >> >> >> >> I developing a RPL model using choice experiment data >> >> >> >> The model is as followed: >> >> >> >> Calc; Ran(1234567)$ >> >> RPLOGIT >> >> ; Choices = 1,2,3 >> >> ; Lhs = CHOICE, CSET, ALT >> >> ; Rhs = L_IMP, L_RAU, L_GAP, L_PGS, >> >> FRESH, O_SUPE, O_SPEC, NPRICE >> >> ; Fcn = L_IMP(n), L_RAU(n), L_GAP(n), >> >> L_PGS(n), FRESH(n), O_SUPE(n), O_SPEC(n), NPRICE(l) >> >> ; Halton >> >> ; Pds = csi >> >> ; Pts = 20 >> >> ; Parameters >> >> ; Maxit = 150$ >> >> >> >> I have changed the price attribute to negative values, so I can use a >> lognormal distribution of for the price attribute. >> >> I am getting the following results >> >> >> >> --------+------------------------------------------------------------ >> --------+- >> --------+------ >> - >> >> | Standard Prob. 95% Confidence >> >> CHOICE| Coefficient Error z |z|>Z* Interval >> >> --------+------------------------------------------------------------ >> --------+- >> --------+------ >> - >> >> |Random parameters in utility >> functions.............................. >> >> L_IMP| -1.11608*** .29152 -3.83 .0001 -1.68745 -.54470 >> >> L_RAU| 1.49941*** .09880 15.18 .0000 1.30577 1.69304 >> >> L_GAP| 1.82794*** .10487 17.43 .0000 1.62239 2.03349 >> >> L_PGS| .63730** .25734 2.48 .0133 .13291 1.14168 >> >> FRESH| -.61318*** .05496 -11.16 .0000 -.72089 -.50546 >> >> O_SUPE| .43891*** .07567 5.80 .0000 .29060 .58721 >> >> O_SPEC| -.76256*** .17329 -4.40 .0000 -1.10221 -.42291 >> >> NPRICE| -1.61991*** .33228 -4.88 .0000 -2.27117 -.96865 >> >> |Distns. of RPs. Std.Devs or limits of >> triangular.................... >> >> NsL_IMP| 1.52346*** .31666 4.81 .0000 .90281 2.14410 >> >> NsL_RAU| .69380*** .13439 5.16 .0000 .43040 .95721 >> >> NsL_GAP| .01744 .24287 .07 .9427 -.45858 .49346 >> >> NsL_PGS| .95598*** .21017 4.55 .0000 .54405 1.36790 >> >> NsFRESH| .48681*** .05657 8.60 .0000 .37593 .59770 >> >> NsO_SUPE| 1.65455*** .11307 14.63 .0000 1.43293 1.87616 >> >> NsO_SPEC| 1.08890*** .12068 9.02 .0000 .85237 1.32544 >> >> LsNPRICE| .99479*** .18655 5.33 .0000 .62917 1.36041 >> >> >> >> If I am not wrong, I can calculate the population mean of the price >> E(beta) = exp(beta + sigma^2/2) >> >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ >> >> [CALC] = .3246179 >> >> >> >> Then, I am using the procedure described in section N29.8.2 of the >> Nlogit manual to examine the distribution of the parameters. >> >> Matrix; bn = beta_i; sn =sdbeta_i $ >> >> CREATE; BIMP=0; BRAU=0; BGAP=0; BPGS=0; BFRE =0; BSUP=0; BSPE=0; >> BNPR=0 $ >> >> CREATE ;SIMP=0; SRAU=0; SGAP=0; SPGS=0; SFRE =0; SSUP=0; SSPE=0; >> SNPR=0 $ >> >> NAMELIST; betan = BIMP,BRAU, BGAP, BPGS, BFRE, BSUP, BSPE, BNPR$ >> >> NAMELIST; sbetan = SIMP,SRAU, SGAP, SPGS, SFRE, SSUP, SSPE, SNPR$ >> >> CREATE ; betan =bn$ >> >> CREATE ; sbetan = sn$ >> >> >> >> CALC; List; XBR(BNPR)$ ? calculate the average of the beta for nprice >> >> >> >> |-> CALC; List; XBR(BNPR)$ >> >> [CALC] = .0257560 >> >> >> >> My understanding is that these two figures should be close to one another. >> Is there anything that could explain such difference between these >> two ways to estimate the results? >> >> >> >> Any help is welcomed? >> >> >> >> Best, >> >> >> >> Damien >> >> >> >> _______________________________________________ >> Limdep site list >> Limdep at mailman.sydney.edu.au >> http://limdep.itls.usyd.edu.au >> From wgreene at stern.nyu.edu Sat Feb 17 01:29:14 2018 From: wgreene at stern.nyu.edu (William Greene) Date: Fri, 16 Feb 2018 09:29:14 -0500 Subject: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? In-Reply-To: <00f001d3a669$f0620950$d1261bf0$@cirad.fr> References: <009301d3a647$8264f6a0$872ee3e0$@cirad.fr> <60299233-429c-049c-738d-e89661957fb7@uw.edu.pl> <00ac01d3a651$e6e119b0$b4a34d10$@cirad.fr> <0a52bfb4-457f-3907-8e40-d9eac46008e2@uw.edu.pl> <00f001d3a669$f0620950$d1261bf0$@cirad.fr> Message-ID: Damien. That looks even better. Note, you can use DSTAT with matrices directly - the statistics are computed for the columns of the matrices. Note the example in my previous note. Cheers Bill Greene On Thu, Feb 15, 2018 at 9:33 AM, Damien Jourdain wrote: > Dear Mik, > > Thank you. > I've looked again, and I think found the mistake I made. > When creating the variables from the matrix, I forgot to add the "Sample; > 1-1400". By failing to do so, I suppose the calculation of the average > include all the zeros for the variable BNPR (from 1401 to 13500 ... since I > have 13500 rows). This result in an average that is much smaller than the > reality! > > I am now adding the following line Sample ; 1-1400$ before getting the > parameters from the matrix > > When I tried again with this statement, I find the calculated from direct > estimation being quite close to the average of posterior > individual-specific estimates. If that is correct, there is no need to use > the exponential of the coefficients. > > |-> CALC; List; XBR(BNPR)$ > [CALC] = .3199974 > > |-> create; expBNPR = exp(BNPR)$ > |-> calc; list; xbr(expBNPR)$ > [CALC] = 1.6429415 > > |-> CALC; LIST; exp(-1.95961 + (1.31682^2)/2)$ > [CALC] = .3353426 > > Again, thank you for your help and interest. > > Damien > > > -----Message d'origine----- > De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] > Envoy? : Thursday, February 15, 2018 3:21 PM > ? : Damien Jourdain > Objet : Re: [Limdep Nlogit List] Interpretation of RPL coefficients when > using lognormal distribution: why results from direct estimation and from > estimated parameters are not giving the same results? > > > Dear Damien, > > This is what I would expect - direct estimation using coefficients is > likely always better than the one based on posterior individual-specific > estimates (even though asymptotically they should be equivalent). > > Best regards, > Mik > > > On 2018-02-15 12:41, Damien Jourdain wrote: > > Dear Mik, > > > > Thank you for the suggestion. > > > > I tried that but there is still an important difference between the two. > > > > |-> create; expBNPR = exp(BNPR)$ > > |-> calc; list; xbr(expBNPR)$ > > [CALC] = 1.0329355 > > > > As the direct estimation from the coefficients is giving : > >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > >> [CALC] = .3246179 > > By the way, the more 'realistic' calculation is the direct estimation > > from the coefficients (at least it is of the same magnitude than the > > MNL coefficient for price) > > > > > > Damien > > > > > > -----Message d'origine----- > > > > De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, > > February 15, 2018 12:34 PM ? : Damien Jourdain Objet : Re: [Limdep > > Nlogit List] Interpretation of RPL coefficients when using lognormal > distribution: why results from direct estimation and from estimated > parameters are not giving the same results? > > > > > > Dear Damien, > > > > Shoulnd't you have something like: > > > > create; expBNPR = exp(BNPR)$ > > > > first? > > > > Then > > calc; list; xbr(expBNPR)$ > > > > Cheers, > > Mik > > > > > > > > > > On 2018-02-15 11:26, Damien Jourdain wrote: > >> Dear All, > >> > >> > >> > >> I developing a RPL model using choice experiment data > >> > >> > >> > >> The model is as followed: > >> > >> > >> > >> Calc; Ran(1234567)$ > >> > >> RPLOGIT > >> > >> ; Choices = 1,2,3 > >> > >> ; Lhs = CHOICE, CSET, ALT > >> > >> ; Rhs = L_IMP, L_RAU, L_GAP, L_PGS, > >> > >> FRESH, O_SUPE, O_SPEC, NPRICE > >> > >> ; Fcn = L_IMP(n), L_RAU(n), L_GAP(n), > >> > >> L_PGS(n), FRESH(n), O_SUPE(n), O_SPEC(n), NPRICE(l) > >> > >> ; Halton > >> > >> ; Pds = csi > >> > >> ; Pts = 20 > >> > >> ; Parameters > >> > >> ; Maxit = 150$ > >> > >> > >> > >> I have changed the price attribute to negative values, so I can use a > >> lognormal distribution of for the price attribute. > >> > >> I am getting the following results > >> > >> > >> > >> --------+------------------------------------------------------------ > >> --------+- > >> --------+------ > >> - > >> > >> | Standard Prob. 95% > Confidence > >> > >> CHOICE| Coefficient Error z |z|>Z* Interval > >> > >> --------+------------------------------------------------------------ > >> --------+- > >> --------+------ > >> - > >> > >> |Random parameters in utility > >> functions.............................. > >> > >> L_IMP| -1.11608*** .29152 -3.83 .0001 -1.68745 > -.54470 > >> > >> L_RAU| 1.49941*** .09880 15.18 .0000 1.30577 > 1.69304 > >> > >> L_GAP| 1.82794*** .10487 17.43 .0000 1.62239 > 2.03349 > >> > >> L_PGS| .63730** .25734 2.48 .0133 .13291 > 1.14168 > >> > >> FRESH| -.61318*** .05496 -11.16 .0000 -.72089 > -.50546 > >> > >> O_SUPE| .43891*** .07567 5.80 .0000 .29060 > .58721 > >> > >> O_SPEC| -.76256*** .17329 -4.40 .0000 -1.10221 > -.42291 > >> > >> NPRICE| -1.61991*** .33228 -4.88 .0000 -2.27117 > -.96865 > >> > >> |Distns. of RPs. Std.Devs or limits of > >> triangular.................... > >> > >> NsL_IMP| 1.52346*** .31666 4.81 .0000 .90281 2.14410 > >> > >> NsL_RAU| .69380*** .13439 5.16 .0000 .43040 .95721 > >> > >> NsL_GAP| .01744 .24287 .07 .9427 -.45858 .49346 > >> > >> NsL_PGS| .95598*** .21017 4.55 .0000 .54405 1.36790 > >> > >> NsFRESH| .48681*** .05657 8.60 .0000 .37593 .59770 > >> > >> NsO_SUPE| 1.65455*** .11307 14.63 .0000 1.43293 > 1.87616 > >> > >> NsO_SPEC| 1.08890*** .12068 9.02 .0000 .85237 > 1.32544 > >> > >> LsNPRICE| .99479*** .18655 5.33 .0000 .62917 > 1.36041 > >> > >> > >> > >> If I am not wrong, I can calculate the population mean of the price > >> E(beta) = exp(beta + sigma^2/2) > >> > >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > >> > >> [CALC] = .3246179 > >> > >> > >> > >> Then, I am using the procedure described in section N29.8.2 of the > >> Nlogit manual to examine the distribution of the parameters. > >> > >> Matrix; bn = beta_i; sn =sdbeta_i $ > >> > >> CREATE; BIMP=0; BRAU=0; BGAP=0; BPGS=0; BFRE =0; BSUP=0; BSPE=0; > >> BNPR=0 $ > >> > >> CREATE ;SIMP=0; SRAU=0; SGAP=0; SPGS=0; SFRE =0; SSUP=0; SSPE=0; > >> SNPR=0 $ > >> > >> NAMELIST; betan = BIMP,BRAU, BGAP, BPGS, BFRE, BSUP, BSPE, BNPR$ > >> > >> NAMELIST; sbetan = SIMP,SRAU, SGAP, SPGS, SFRE, SSUP, SSPE, SNPR$ > >> > >> CREATE ; betan =bn$ > >> > >> CREATE ; sbetan = sn$ > >> > >> > >> > >> CALC; List; XBR(BNPR)$ ? calculate the average of the beta for nprice > >> > >> > >> > >> |-> CALC; List; XBR(BNPR)$ > >> > >> [CALC] = .0257560 > >> > >> > >> > >> My understanding is that these two figures should be close to one > another. > >> Is there anything that could explain such difference between these > >> two ways to estimate the results? > >> > >> > >> > >> Any help is welcomed? > >> > >> > >> > >> Best, > >> > >> > >> > >> Damien > >> > >> > >> > >> _______________________________________________ > >> Limdep site list > >> Limdep at mailman.sydney.edu.au > >> http://limdep.itls.usyd.edu.au > >> > > _______________________________________________ > 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/lziwCL7rK8tm1GrMIBihA2?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 From richard.turner at imarketresearch.com Sat Feb 17 07:20:21 2018 From: richard.turner at imarketresearch.com (Richard Turner) Date: Fri, 16 Feb 2018 15:20:21 -0500 Subject: [Limdep Nlogit List] How to deal with large numbers of attributes? Message-ID: Greetings, What is the best way to handle large numbers of attributes in discrete choice experiments? Is it better to do a *partial profile design* or to do some* two-step approach* such as conducting an "initial study" using a partial profile design, then conduct a final study using the most important attributes, which were derived from the initial study (implicit in the second method would be to synthesize the learnings from both studies to get some ranking of all the attributes)? I've done some searching, but haven't found any "defining" papers on the subject. Any advice and/or direction is greatly appreciated! Regards, Richard From miq at wne.uw.edu.pl Sat Feb 17 07:33:23 2018 From: miq at wne.uw.edu.pl (=?UTF-8?Q?Miko=c5=82aj_Czajkowski?=) Date: Fri, 16 Feb 2018 21:33:23 +0100 Subject: [Limdep Nlogit List] How to deal with large numbers of attributes? In-Reply-To: References: Message-ID: <026cc501-7aed-ac1d-9d14-a71965da8973@wne.uw.edu.pl> Dear Richard, It seems to me like the answer to this question would depend on the goal of the modeller - whether he wants to learn a lot about the most important attributes only, or have some idea about all the attributes. I am not sure is a lot of concrete advice can be given in these kinds of situations. Cheers, Mik On 2018-02-16 21:20, Richard Turner wrote: > Greetings, > > What is the best way to handle large numbers of attributes in discrete > choice experiments? > > Is it better to do a *partial profile design* or to do some* two-step > approach* such as conducting an "initial study" using a partial profile > design, then conduct a final study using the most important attributes, > which were derived from the initial study (implicit in the second method > would be to synthesize the learnings from both studies to get some ranking > of all the attributes)? > > I've done some searching, but haven't found any "defining" papers on the > subject. > > Any advice and/or direction is greatly appreciated! > > Regards, > > Richard > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > From david.hensher at sydney.edu.au Sat Feb 17 07:56:16 2018 From: david.hensher at sydney.edu.au (David Hensher) Date: Fri, 16 Feb 2018 20:56:16 +0000 Subject: [Limdep Nlogit List] How to deal with large numbers of attributes? In-Reply-To: <026cc501-7aed-ac1d-9d14-a71965da8973@wne.uw.edu.pl> References: , <026cc501-7aed-ac1d-9d14-a71965da8973@wne.uw.edu.pl> Message-ID: <59E77ACA-1789-4E2A-8AEE-DF90B5679F44@sydney.edu.au> This relates to the literature on attribute non attendance where different attributes are relevant to different people and selecting a limited set initially without strong evidence of the universal relevant set is behaviourally concerning. Depending on how many attributes, up to 20 or so is fine and one can ask questions on which attributes are attended to. Lots of papers on this by people such as Hensher, Louviere, Scarpa, and the special issue a couple of years ago in J of choice modelling on process heuristics and especially the design of designs (DoD) approach initially developed by Hensher Sent from my iPhone 0418 433 057 David A Hensher Note: hgroup at optusnet.com.au has been cancelled so instead use hgroup at hensher.com.au David.hensher at bigpond.com David.hensher at sydney.edu.au These emails are linked so use one only On 17 Feb 2018, at 7:33 am, Miko?aj Czajkowski > wrote: Dear Richard, It seems to me like the answer to this question would depend on the goal of the modeller - whether he wants to learn a lot about the most important attributes only, or have some idea about all the attributes. I am not sure is a lot of concrete advice can be given in these kinds of situations. Cheers, Mik On 2018-02-16 21:20, Richard Turner wrote: Greetings, What is the best way to handle large numbers of attributes in discrete choice experiments? Is it better to do a *partial profile design* or to do some* two-step approach* such as conducting an "initial study" using a partial profile design, then conduct a final study using the most important attributes, which were derived from the initial study (implicit in the second method would be to synthesize the learnings from both studies to get some ranking of all the attributes)? I've done some searching, but haven't found any "defining" papers on the subject. Any advice and/or direction is greatly appreciated! Regards, Richard _______________________________________________ Limdep site list Limdep at mailman.sydney.edu.au http://limdep.itls.usyd.edu.au _______________________________________________ Limdep site list Limdep at mailman.sydney.edu.au http://limdep.itls.usyd.edu.au From miq at wne.uw.edu.pl Sat Feb 17 08:17:02 2018 From: miq at wne.uw.edu.pl (=?UTF-8?Q?Miko=c5=82aj_Czajkowski?=) Date: Fri, 16 Feb 2018 22:17:02 +0100 Subject: [Limdep Nlogit List] How to deal with large numbers of attributes? In-Reply-To: <59E77ACA-1789-4E2A-8AEE-DF90B5679F44@sydney.edu.au> References: <026cc501-7aed-ac1d-9d14-a71965da8973@wne.uw.edu.pl> <59E77ACA-1789-4E2A-8AEE-DF90B5679F44@sydney.edu.au> Message-ID: Dear David, As far as I understood Richard's question, option (1) *partial profile design* is having many versions of the study using different attributes vs. option (2) would be an initial study like (1) + final study aimed at learning more about the most prominent attributes. Attribute non-attendance would be a thing to econometrically control in each case, (1) and (2), but does it help determine if option (1) or (2) is preferable? Best regards, Mik On 2018-02-16 21:56, David Hensher via Limdep wrote: > This relates to the literature on attribute non attendance where different attributes are relevant to different people and selecting a limited set initially without strong evidence of the universal relevant set is behaviourally concerning. > > Depending on how many attributes, up to 20 or so is fine and one can ask questions on which attributes are attended to. Lots of papers on this by people such as Hensher, Louviere, Scarpa, and the special issue a couple of years ago in J of choice modelling on process heuristics and especially the design of designs (DoD) approach initially developed by Hensher > > Sent from my iPhone > 0418 433 057 > David A Hensher > > Note: hgroup at optusnet.com.au has been cancelled so instead use > hgroup at hensher.com.au David.hensher at bigpond.com > David.hensher at sydney.edu.au > These emails are linked so use one only > > > On 17 Feb 2018, at 7:33 am, Miko?aj Czajkowski > wrote: > > > Dear Richard, > > It seems to me like the answer to this question would depend on the goal of the modeller - whether he wants to learn a lot about the most important attributes only, or have some idea about all the attributes. I am not sure is a lot of concrete advice can be given in these kinds of situations. > > Cheers, > Mik > > > On 2018-02-16 21:20, Richard Turner wrote: > Greetings, > > What is the best way to handle large numbers of attributes in discrete > choice experiments? > > Is it better to do a *partial profile design* or to do some* two-step > approach* such as conducting an "initial study" using a partial profile > design, then conduct a final study using the most important attributes, > which were derived from the initial study (implicit in the second method > would be to synthesize the learnings from both studies to get some ranking > of all the attributes)? > > I've done some searching, but haven't found any "defining" papers on the > subject. > > Any advice and/or direction is greatly appreciated! > > Regards, > > Richard > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > > > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > From david.hensher at sydney.edu.au Sat Feb 17 08:36:34 2018 From: david.hensher at sydney.edu.au (David Hensher) Date: Fri, 16 Feb 2018 21:36:34 +0000 Subject: [Limdep Nlogit List] How to deal with large numbers of attributes? In-Reply-To: References: <026cc501-7aed-ac1d-9d14-a71965da8973@wne.uw.edu.pl> <59E77ACA-1789-4E2A-8AEE-DF90B5679F44@sydney.edu.au> Message-ID: <5A874EE2.70602@sydney.edu.au> Dear Mik It does sound a bit like the DoD approach and I think it would be very interesting to design a number of designs (each with subsets of attributes but some common attributes) and test for differences - bit like hierarchical information integration that some of us did many years ago (Louviere, Hensher, Timmermans). Then settle on one design if they want that or indeed jointly estimate across all designs. David On 17/02/2018 8:17 AM, Miko?aj Czajkowski wrote: > > Dear David, > > As far as I understood Richard's question, option (1) *partial profile > design* is having many versions of the study using different > attributes vs. option (2) would be an initial study like (1) + final > study aimed at learning more about the most prominent attributes. > Attribute non-attendance would be a thing to econometrically control > in each case, (1) and (2), but does it help determine if option (1) or > (2) is preferable? > > Best regards, > Mik > > > On 2018-02-16 21:56, David Hensher via Limdep wrote: >> This relates to the literature on attribute non attendance where >> different attributes are relevant to different people and selecting a >> limited set initially without strong evidence of the universal >> relevant set is behaviourally concerning. >> >> Depending on how many attributes, up to 20 or so is fine and one can >> ask questions on which attributes are attended to. Lots of papers on >> this by people such as Hensher, Louviere, Scarpa, and the special >> issue a couple of years ago in J of choice modelling on process >> heuristics and especially the design of designs (DoD) approach >> initially developed by Hensher >> >> Sent from my iPhone >> 0418 433 057 >> David A Hensher >> >> Note: hgroup at optusnet.com.au has been >> cancelled so instead use >> hgroup at hensher.com.au >> David.hensher at bigpond.com >> David.hensher at sydney.edu.au >> These emails are linked so use one only >> >> >> On 17 Feb 2018, at 7:33 am, Miko?aj Czajkowski >> > wrote: >> >> >> Dear Richard, >> >> It seems to me like the answer to this question would depend on the >> goal of the modeller - whether he wants to learn a lot about the most >> important attributes only, or have some idea about all the >> attributes. I am not sure is a lot of concrete advice can be given in >> these kinds of situations. >> >> Cheers, >> Mik >> >> >> On 2018-02-16 21:20, Richard Turner wrote: >> Greetings, >> >> What is the best way to handle large numbers of attributes in discrete >> choice experiments? >> >> Is it better to do a *partial profile design* or to do some* two-step >> approach* such as conducting an "initial study" using a partial profile >> design, then conduct a final study using the most important attributes, >> which were derived from the initial study (implicit in the second method >> would be to synthesize the learnings from both studies to get some >> ranking >> of all the attributes)? >> >> I've done some searching, but haven't found any "defining" papers on the >> subject. >> >> Any advice and/or direction is greatly appreciated! >> >> Regards, >> >> Richard >> _______________________________________________ >> Limdep site list >> Limdep at mailman.sydney.edu.au >> http://limdep.itls.usyd.edu.au >> >> >> _______________________________________________ >> Limdep site list >> Limdep at mailman.sydney.edu.au >> http://limdep.itls.usyd.edu.au >> >> _______________________________________________ >> Limdep site list >> Limdep at mailman.sydney.edu.au >> http://limdep.itls.usyd.edu.au >> > > _______________________________________________ > Limdep site list > Limdep at mailman.sydney.edu.au > http://limdep.itls.usyd.edu.au > > -- DAVID HENSHER FASSA, PhD | Professor and Founding Director Institute of Transport and Logistics Studies | The University of Sydney Business School THE UNIVERSITY OF SYDNEY Rm 201, Building H73 | The University of Sydney | NSW | 2006 Street Address: 378 Abercrombie St, Darlington NSW 2008 T +61 2 9114 1871 | F +61 2 9114 1863 | M +61 418 433 057 E David.Hensher at sydney.edu.au | W sydney.edu.au/business/itls Celebrating 25 years of ITLS: 1991-2016 https://protect-au.mimecast.com/s/X0DNCvl0PoC2mMKJHQ51YG?domain=youtu.be ERA Rank 5 (Transportation and Freight Services) Co-Founder of the International Conference Series on Competition and Ownership of Land Passenger Transport (The 'Thredbo' Series) https://protect-au.mimecast.com/s/le6HCwVLQmiAMY95UqD2yg?domain=thredbo-conference-series.org Join the ITLS group on LinkedIn Second edition of Applied Choice Analysis now available at www.cambridge.org/9781107465923 CRICOS 00026A This email plus any attachments to it are confidential. Any unauthorised use is strictly prohibited. If you receive this email in error, please delete it and any attachments. Please think of our environment and only print this e-mail if necessary. From Lixian.Qian at xjtlu.edu.cn Sun Feb 18 13:36:31 2018 From: Lixian.Qian at xjtlu.edu.cn (Lixian Qian) Date: Sun, 18 Feb 2018 02:36:31 +0000 Subject: [Limdep Nlogit List] Hierarchical Baysian Approach in NLogit Message-ID: Dear all, I am wondering whether NLogit can support the hierarchical baysian (HB) approach when estimating the DCM. Thanks. Best, Lixian From djourdain at ait.asia Tue Feb 20 18:45:31 2018 From: djourdain at ait.asia (Damien Jourdain) Date: Tue, 20 Feb 2018 09:45:31 +0200 Subject: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? In-Reply-To: References: <009301d3a647$8264f6a0$872ee3e0$@cirad.fr> <60299233-429c-049c-738d-e89661957fb7@uw.edu.pl> <00ac01d3a651$e6e119b0$b4a34d10$@cirad.fr> <0a52bfb4-457f-3907-8e40-d9eac46008e2@uw.edu.pl> <00f001d3a669$f0620950$d1261bf0$@cirad.fr> Message-ID: <008401d3aa1e$cba5d520$62f17f60$@cirad.fr> Dear Pr. Greene, Thank you for this tip. It is easier to operate indeed! Damien -----Message d'origine----- De : Limdep [mailto:limdep-bounces at mailman.sydney.edu.au] De la part de William Greene Envoy? : Friday, February 16, 2018 4:29 PM ? : Limdep and Nlogit Mailing List Objet : Re: [Limdep Nlogit List] Interpretation of RPL coefficients when using lognormal distribution: why results from direct estimation and from estimated parameters are not giving the same results? Damien. That looks even better. Note, you can use DSTAT with matrices directly - the statistics are computed for the columns of the matrices. Note the example in my previous note. Cheers Bill Greene On Thu, Feb 15, 2018 at 9:33 AM, Damien Jourdain wrote: > Dear Mik, > > Thank you. > I've looked again, and I think found the mistake I made. > When creating the variables from the matrix, I forgot to add the > "Sample; 1-1400". By failing to do so, I suppose the calculation of > the average include all the zeros for the variable BNPR (from 1401 to > 13500 ... since I have 13500 rows). This result in an average that is > much smaller than the reality! > > I am now adding the following line Sample ; 1-1400$ before getting the > parameters from the matrix > > When I tried again with this statement, I find the calculated from > direct estimation being quite close to the average of posterior > individual-specific estimates. If that is correct, there is no need to > use the exponential of the coefficients. > > |-> CALC; List; XBR(BNPR)$ > [CALC] = .3199974 > > |-> create; expBNPR = exp(BNPR)$ > |-> calc; list; xbr(expBNPR)$ > [CALC] = 1.6429415 > > |-> CALC; LIST; exp(-1.95961 + (1.31682^2)/2)$ > [CALC] = .3353426 > > Again, thank you for your help and interest. > > Damien > > > -----Message d'origine----- > De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, > February 15, 2018 3:21 PM ? : Damien Jourdain Objet : Re: [Limdep > Nlogit List] Interpretation of RPL coefficients when using lognormal > distribution: why results from direct estimation and from estimated > parameters are not giving the same results? > > > Dear Damien, > > This is what I would expect - direct estimation using coefficients is > likely always better than the one based on posterior > individual-specific estimates (even though asymptotically they should be equivalent). > > Best regards, > Mik > > > On 2018-02-15 12:41, Damien Jourdain wrote: > > Dear Mik, > > > > Thank you for the suggestion. > > > > I tried that but there is still an important difference between the two. > > > > |-> create; expBNPR = exp(BNPR)$ > > |-> calc; list; xbr(expBNPR)$ > > [CALC] = 1.0329355 > > > > As the direct estimation from the coefficients is giving : > >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > >> [CALC] = .3246179 > > By the way, the more 'realistic' calculation is the direct > > estimation from the coefficients (at least it is of the same > > magnitude than the MNL coefficient for price) > > > > > > Damien > > > > > > -----Message d'origine----- > > > > De : Miko?aj Czajkowski [mailto:mc at uw.edu.pl] Envoy? : Thursday, > > February 15, 2018 12:34 PM ? : Damien Jourdain Objet : Re: [Limdep > > Nlogit List] Interpretation of RPL coefficients when using lognormal > distribution: why results from direct estimation and from estimated > parameters are not giving the same results? > > > > > > Dear Damien, > > > > Shoulnd't you have something like: > > > > create; expBNPR = exp(BNPR)$ > > > > first? > > > > Then > > calc; list; xbr(expBNPR)$ > > > > Cheers, > > Mik > > > > > > > > > > On 2018-02-15 11:26, Damien Jourdain wrote: > >> Dear All, > >> > >> > >> > >> I developing a RPL model using choice experiment data > >> > >> > >> > >> The model is as followed: > >> > >> > >> > >> Calc; Ran(1234567)$ > >> > >> RPLOGIT > >> > >> ; Choices = 1,2,3 > >> > >> ; Lhs = CHOICE, CSET, ALT > >> > >> ; Rhs = L_IMP, L_RAU, L_GAP, L_PGS, > >> > >> FRESH, O_SUPE, O_SPEC, NPRICE > >> > >> ; Fcn = L_IMP(n), L_RAU(n), L_GAP(n), > >> > >> L_PGS(n), FRESH(n), O_SUPE(n), O_SPEC(n), NPRICE(l) > >> > >> ; Halton > >> > >> ; Pds = csi > >> > >> ; Pts = 20 > >> > >> ; Parameters > >> > >> ; Maxit = 150$ > >> > >> > >> > >> I have changed the price attribute to negative values, so I can use > >> a lognormal distribution of for the price attribute. > >> > >> I am getting the following results > >> > >> > >> > >> --------+---------------------------------------------------------- > >> --------+-- > >> --------+- > >> --------+------ > >> - > >> > >> | Standard Prob. 95% > Confidence > >> > >> CHOICE| Coefficient Error z |z|>Z* Interval > >> > >> --------+---------------------------------------------------------- > >> --------+-- > >> --------+- > >> --------+------ > >> - > >> > >> |Random parameters in utility > >> functions.............................. > >> > >> L_IMP| -1.11608*** .29152 -3.83 .0001 -1.68745 > -.54470 > >> > >> L_RAU| 1.49941*** .09880 15.18 .0000 1.30577 > 1.69304 > >> > >> L_GAP| 1.82794*** .10487 17.43 .0000 1.62239 > 2.03349 > >> > >> L_PGS| .63730** .25734 2.48 .0133 .13291 > 1.14168 > >> > >> FRESH| -.61318*** .05496 -11.16 .0000 -.72089 > -.50546 > >> > >> O_SUPE| .43891*** .07567 5.80 .0000 .29060 > .58721 > >> > >> O_SPEC| -.76256*** .17329 -4.40 .0000 -1.10221 > -.42291 > >> > >> NPRICE| -1.61991*** .33228 -4.88 .0000 -2.27117 > -.96865 > >> > >> |Distns. of RPs. Std.Devs or limits of > >> triangular.................... > >> > >> NsL_IMP| 1.52346*** .31666 4.81 .0000 .90281 2.14410 > >> > >> NsL_RAU| .69380*** .13439 5.16 .0000 .43040 .95721 > >> > >> NsL_GAP| .01744 .24287 .07 .9427 -.45858 .49346 > >> > >> NsL_PGS| .95598*** .21017 4.55 .0000 .54405 1.36790 > >> > >> NsFRESH| .48681*** .05657 8.60 .0000 .37593 .59770 > >> > >> NsO_SUPE| 1.65455*** .11307 14.63 .0000 1.43293 > 1.87616 > >> > >> NsO_SPEC| 1.08890*** .12068 9.02 .0000 .85237 > 1.32544 > >> > >> LsNPRICE| .99479*** .18655 5.33 .0000 .62917 > 1.36041 > >> > >> > >> > >> If I am not wrong, I can calculate the population mean of the price > >> E(beta) = exp(beta + sigma^2/2) > >> > >> |-> CALC; LIST; EXP(-1.61991 + (0.99479^2)/2)$ > >> > >> [CALC] = .3246179 > >> > >> > >> > >> Then, I am using the procedure described in section N29.8.2 of the > >> Nlogit manual to examine the distribution of the parameters. > >> > >> Matrix; bn = beta_i; sn =sdbeta_i $ > >> > >> CREATE; BIMP=0; BRAU=0; BGAP=0; BPGS=0; BFRE =0; BSUP=0; BSPE=0; > >> BNPR=0 $ > >> > >> CREATE ;SIMP=0; SRAU=0; SGAP=0; SPGS=0; SFRE =0; SSUP=0; SSPE=0; > >> SNPR=0 $ > >> > >> NAMELIST; betan = BIMP,BRAU, BGAP, BPGS, BFRE, BSUP, BSPE, BNPR$ > >> > >> NAMELIST; sbetan = SIMP,SRAU, SGAP, SPGS, SFRE, SSUP, SSPE, SNPR$ > >> > >> CREATE ; betan =bn$ > >> > >> CREATE ; sbetan = sn$ > >> > >> > >> > >> CALC; List; XBR(BNPR)$ ? calculate the average of the beta for > >> nprice > >> > >> > >> > >> |-> CALC; List; XBR(BNPR)$ > >> > >> [CALC] = .0257560 > >> > >> > >> > >> My understanding is that these two figures should be close to one > another. > >> Is there anything that could explain such difference between these > >> two ways to estimate the results? > >> > >> > >> > >> Any help is welcomed? > >> > >> > >> > >> Best, > >> > >> > >> > >> Damien > >> > >> > >> > >> _______________________________________________ > >> Limdep site list > >> Limdep at mailman.sydney.edu.au > >> http://limdep.itls.usyd.edu.au > >> > > _______________________________________________ > 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/vOgiCwVLQmiPLk3LhVsScn?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 _______________________________________________ Limdep site list Limdep at mailman.sydney.edu.au http://limdep.itls.usyd.edu.au From siibawei2013 at gmail.com Wed Feb 21 16:55:34 2018 From: siibawei2013 at gmail.com (Alhassan Siiba) Date: Wed, 21 Feb 2018 13:55:34 +0800 Subject: [Limdep Nlogit List] Problems with running MNL in NLOGIT v.5 Message-ID: Dear All, I am trying to run MNL model with coefficients in NLOGIT v. 5. However, I kept on getting different error messages in that regard. Some of the error messages are: "Error 101: LOGIT - one of the cells (outcomes) has no observations". "Error 221: The data on the LHS variable appear not to be coded 0,1,2.." I am new to NLOGIT. I am using SP data with the dependent variable being RANK for different travel modes and the Independent variables being the attributes of the travel modes. Please, can someone help me out? Thank you. Kind regards, Siiba. From ssingh10 at unl.edu Thu Feb 22 05:36:31 2018 From: ssingh10 at unl.edu (Sunil Kumar Singh) Date: Wed, 21 Feb 2018 18:36:31 +0000 Subject: [Limdep Nlogit List] Mediation Question Message-ID: Limdep Community, I am trying to run a probit model as specified below. I also want to check if the impact of X1, X2, and X1*X2 is mediated through X4. However, I am not sure how to specify the model or run some ancillary test for this. Would appreciate any help. skip$ Calc ; ran (12345)$ probit ; lhs = Y ; rhs = one,X1,X2,X3, X1*X2,X4 ; Panel ; pds = COUNT ; rpm ; fcn = X4(n) ; pts = 100 ; Halton$ Thanks, Sunil