[Limdep Nlogit List] ASCs calibration
Thao Thai
Thao.T.Thai at monash.edu
Fri Sep 11 10:44:19 AEST 2020
Hi Nlogit users,
Train's textbook (2000) page 75 pointed out that when using the results
from SP for simulations, ASCs should be calibrated to the actual shares of
alternatives.
(1) I followed Train's suggestion to adjust ASCs to new values and run the
syntax below, the coefficients are different from the coefficients
obtained from the original SP. I guess it is not the correct way to do ASCs
calibration?
|-> sample;all$
|-> Nlogit
;lhs = cho, cset, alti
;choices = H, C, P, I, G, N
;Alg=BFGS
;checkdata
;model:
U(H) = rl_h1 * RL_H1 + rl_h2 * RL_H2
+ fl_ * FL_H
+ cr_h1 * CR_H1 + cr_2 * CR_H2
+ lo_h1 * LO_H1
+ sa_ * SA_H
/
U(C) = com[0.960]
+ rl_c1 * RL_C1 + rl_c2 * RL_C2
+ fl_ * FL_C
+ cr_c1 * CR_C1 + cr_2 * CR_C2
+ lo_c1 * LO_C1 + lo_c2 * LO_C2
+ sa_ * SA_C/
U(P) = pri[-1.225]
+ rl_p1 * RL_P1
+ fl_ * FL_P
+ cr_p1 * CR_P1 + cr_2 * CR_P2
+ lo_p1 * LO_P1 + lo_p2 * LO_P2
+ sa_ * SA_P/
U(I) = ind[-2.440]
+ rl_i1 * RL_I1 + rl_i2 * RL_I2
+ fl_ * FL_I
+ cr_i1 * CR_I1
+ lo_i1 * LO_I1
+ sa_ * SA_I/
U(G) = gov[-0.548]
+ rl_g1 * RL_G1
+ fl_ * FL_G
+ cr_g1 * CR_G1
+ lo_g1 * LO_G1
+ sa_ * SA_G/
U(N) = non[-0.844]
+ rl_n1 * RL_N1
+ fl_ * FL_N
+ cr_n1 * CR_N1
+ lo_n1 * LO_N1 + lo_n2 * LO_N2
+ sa_ * SA_N
$
(2) I then used the endogenous weighting to the SP model, the coefficient
results are different from the original SP model. I guess this way is not
correct either.
|-> sample;all$
|-> Nlogit
;lhs = cho, cset, alti
;choices = H, C, P, I, G, N/0.242,0.5134,0.0388,0.0468,0.0922,0.0668
;Alg=BFGS
;checkdata
;model:
U(H) = rl_h1 * RL_H1 + rl_h2 * RL_H2
+ fl_ * FL_H
+ cr_h1 * CR_H1 + cr_2 * CR_H2
+ lo_h1 * LO_H1
+ sa_ * SA_H
/
U(C) = com
+ rl_c1 * RL_C1 + rl_c2 * RL_C2
+ fl_ * FL_C
+ cr_c1 * CR_C1 + cr_2 * CR_C2
+ lo_c1 * LO_C1 + lo_c2 * LO_C2
+ sa_ * SA_C/
U(P) = pri
+ rl_p1 * RL_P1
+ fl_ * FL_P
+ cr_p1 * CR_P1 + cr_2 * CR_P2
+ lo_p1 * LO_P1 + lo_p2 * LO_P2
+ sa_ * SA_P/
U(I) = ind
+ rl_i1 * RL_I1 + rl_i2 * RL_I2
+ fl_ * FL_I
+ cr_i1 * CR_I1
+ lo_i1 * LO_I1
+ sa_ * SA_I/
U(G) = gov
+ rl_g1 * RL_G1
+ fl_ * FL_G
+ cr_g1 * CR_G1
+ lo_g1 * LO_G1
+ sa_ * SA_G/
U(N) = non
+ rl_n1 * RL_N1
+ fl_ * FL_N
+ cr_n1 * CR_N1
+ lo_n1 * LO_N1 + lo_n2 * LO_N2
+ sa_ * SA_N
$
(3) I then run a model with fixed coefficient estimates but ASCs with
endogenous weighting. The ASCs estimates are similar to the adjusted ASCs
in step (1) as per suggestion of Train (2000). However, the base shares are
not the same as the shares that I specified in the model
|-> sample;all$
|-> Nlogit
;lhs = cho, cset, alti
;choices = H, C, P, I, G, N/0.241,0.5153,0.0386,0.0466,0.0919,0.0666
;checkdata
;show
'Alg=BFGS
;calibrate
;model:
U(H) = rl_h1[0.0700481] * RL_H1 + rl_h2[0.231038] * RL_H2
+ fl_[0.178665] * FL_H
+ cr_h1[0.271042] * CR_H1 + cr_2[0.0529584] * CR_H2
+ lo_h1[-0.414659] * LO_H1
+ sa_[0.0123069] * SA_H
/
U(C) = com
+ rl_c1[0.387869] * RL_C1 + rl_c2[0.263637] * RL_C2
+ fl_[0.178665] * FL_C
+ cr_c1[0.19661] * CR_C1 + cr_2[0.0529584] * CR_C2
+ lo_c1[-0.222456] * LO_C1 + lo_c2[-0.820609] * LO_C2
+ sa_[0.0123069] * SA_C
/
U(P) = pri
+ rl_p1[0.0167174] * RL_P1
+ fl_[0.178665] * FL_P
+ cr_p1[0.333571] * CR_P1 + cr_2[0.0529584] * CR_P2
+ lo_p1[-0.951557] * LO_P1 + lo_p2[-1.03276] * LO_P2
+ sa_[0.0123069] * SA_P
/
U(I) = ind
+ rl_i1[0.643119] * RL_I1 + rl_i2[0.790318] * RL_I2
+ fl_[0.178665] * FL_I
+ cr_i1[0.610239] * CR_I1
+ lo_i1[-0.64615] * LO_I1
+ sa_[0.0123069] * SA_I
/
U(G) = gov
+ rl_g1[-0.331743] * RL_G1
+ fl_[0.178665] * FL_G
+ cr_g1[0.516844] * CR_G1
+ lo_g1[-0.57783] * LO_G1
+ sa_[0.0123069] * SA_G
/
U(N) = non
+ rl_n1[-0.0814267] * RL_N1
+ fl_[0.178665] * FL_N
+ cr_n1[0.36292] * CR_N1
+ lo_n1[-0.585064] * LO_N1 + lo_n2[-0.466387] * LO_N2
+ sa_[0.0123069] * SA_N$
The base share results below
+----------+--------------+--------------+------------------+
|Choice | Base | Scenario | Scenario - Base |
| |%Share Number |%Share Number |ChgShare ChgNumber|
+----------+--------------+--------------+------------------+
|H | 24.823 604 | 22.801 555 | -2.023% -49 |
|C | 48.452 1179 | 52.560 1279 | 4.108% 100 |
|P | 4.596 112 | 4.208 102 | -.388% -10 |
|I | 5.192 126 | 4.804 117 | -.388% -9 |
|G | 9.971 243 | 9.150 223 | -.820% -20 |
|N | 6.965 170 | 6.476 158 | -.489% -12 |
|Total |100.000 2434 |100.000 2434 | .000% 0 |
+----------+--------------+--------------+------------------+
Question: (1) Could you kindly please let me know how I should calibrate
ASCs in simulation in Nlogit?
(2) I'd like to do the simulation using a Mixed logit with
all ASCs being random parameters. Please let me know if there are any
special considerations when calibrating ASCs.
Thank you so much. I truly look forward to your help.
Best regards,
Thao
--
*Thao Thai*| MHEcon(Adv), BPharm
PhD candidate
Centre for Health Economics <https://protect-au.mimecast.com/s/Y8vrCoV1kpfmjXKqF10EXH?domain=monash.edu>
Monash Business School | Monash University
Tel (03) 99029847| Thao.T.Thai at monash.edu <Thao.T.Thai at monash.edu>
More information about the Limdep
mailing list