Mixture of logit

Choice model with latent variable. No measurement equation for the indicators.

author:

Michel Bierlaire, EPFL

date:

Thu Apr 13 16:58:21 2023

import biogeme.biogeme_logging as blog
import biogeme.biogeme as bio
from biogeme import models
import biogeme.distributions as dist
from biogeme.expressions import (
    Beta,
    RandomVariable,
    Integrate,
    exp,
    log,
)

from read_or_estimate import read_or_estimate

from biogeme.data.optima import (
    read_data,
    age_65_more,
    ScaledIncome,
    moreThanOneCar,
    moreThanOneBike,
    individualHouse,
    male,
    haveChildren,
    haveGA,
    highEducation,
    WaitingTimePT,
    Choice,
    TimePT_scaled,
    TimeCar_scaled,
    MarginalCostPT_scaled,
    CostCarCHF_scaled,
    distance_km_scaled,
    PurpHWH,
    PurpOther,
)


logger = blog.get_screen_logger(level=blog.INFO)
logger.info('Example b03choice_only.py')
Example b03choice_only.py

Parameters to be estimated

coef_intercept = Beta('coef_intercept', 0.0, None, None, 1)
coef_age_65_more = Beta('coef_age_65_more', 0.0, None, None, 0)
coef_haveGA = Beta('coef_haveGA', 0.0, None, None, 0)
coef_moreThanOneCar = Beta('coef_moreThanOneCar', 0.0, None, None, 0)
coef_moreThanOneBike = Beta('coef_moreThanOneBike', 0.0, None, None, 0)
coef_individualHouse = Beta('coef_individualHouse', 0.0, None, None, 0)
coef_male = Beta('coef_male', 0.0, None, None, 0)
coef_haveChildren = Beta('coef_haveChildren', 0.0, None, None, 0)
coef_highEducation = Beta('coef_highEducation', 0.0, None, None, 0)

Latent variable: structural equation.

Define a random parameter, normally distributed, designed to be used for numerical integration.

omega = RandomVariable('omega')
density = dist.normalpdf(omega)
sigma_s = Beta('sigma_s', 1, -1, 1, 0)

thresholds = [None, 4, 6, 8, 10, None]
formula_income = models.piecewise_formula(variable=ScaledIncome, thresholds=thresholds)

CARLOVERS = (
    coef_intercept
    + coef_age_65_more * age_65_more
    + formula_income
    + coef_moreThanOneCar * moreThanOneCar
    + coef_moreThanOneBike * moreThanOneBike
    + coef_individualHouse * individualHouse
    + coef_male * male
    + coef_haveChildren * haveChildren
    + coef_haveGA * haveGA
    + coef_highEducation * highEducation
    + sigma_s * omega
)

Choice model: parameters.

ASC_CAR = Beta('ASC_CAR', 0.0, None, None, 0)
ASC_PT = Beta('ASC_PT', 0.0, None, None, 1)
ASC_SM = Beta('ASC_SM', 0.0, None, None, 0)
BETA_COST_HWH = Beta('BETA_COST_HWH', 0.0, None, None, 0)
BETA_COST_OTHER = Beta('BETA_COST_OTHER', 0.0, None, None, 0)
BETA_DIST = Beta('BETA_DIST', 0.0, None, None, 0)
BETA_TIME_CAR_REF = Beta('BETA_TIME_CAR_REF', -0.0001, None, 0, 0)
BETA_TIME_CAR_CL = Beta('BETA_TIME_CAR_CL', -1.0, -3, 3, 0)
BETA_TIME_PT_REF = Beta('BETA_TIME_PT_REF', -0.0001, None, 0, 0)
BETA_TIME_PT_CL = Beta('BETA_TIME_PT_CL', -1.0, -3, 3, 0)
BETA_WAITING_TIME = Beta('BETA_WAITING_TIME', 0.0, None, None, 0)

Definition of utility functions.

BETA_TIME_PT = BETA_TIME_PT_REF * exp(BETA_TIME_PT_CL * CARLOVERS)

V0 = (
    ASC_PT
    + BETA_TIME_PT * TimePT_scaled
    + BETA_WAITING_TIME * WaitingTimePT
    + BETA_COST_HWH * MarginalCostPT_scaled * PurpHWH
    + BETA_COST_OTHER * MarginalCostPT_scaled * PurpOther
)

BETA_TIME_CAR = BETA_TIME_CAR_REF * exp(BETA_TIME_CAR_CL * CARLOVERS)

V1 = (
    ASC_CAR
    + BETA_TIME_CAR * TimeCar_scaled
    + BETA_COST_HWH * CostCarCHF_scaled * PurpHWH
    + BETA_COST_OTHER * CostCarCHF_scaled * PurpOther
)

V2 = ASC_SM + BETA_DIST * distance_km_scaled

Associate utility functions with the numbering of alternatives.

V = {0: V0, 1: V1, 2: V2}

Conditional on omega, we have a logit model (called the kernel).

condprob = models.logit(V, None, Choice)

We integrate over omega using numerical integration.

loglike = log(Integrate(condprob * density, 'omega'))

Read the data

database = read_data()

Create the Biogeme object.

the_biogeme = bio.BIOGEME(database, loglike)
the_biogeme.modelName = 'b03choice_only'
Biogeme parameters read from biogeme.toml.

If estimation results are saved on file, we read them to speed up the process. If not, we estimate the parameters.

results = read_or_estimate(the_biogeme=the_biogeme, directory='saved_results')
print(f'Estimated betas: {len(results.data.betaValues)}')
print(f'Final log likelihood: {results.data.logLike:.3f}')
print(f'Output file: {results.data.htmlFileName}')
Estimated betas: 24
Final log likelihood: -1068.884
Output file: b03choice_only.html
results.get_estimated_parameters()
Value Active bound Rob. Std err Rob. t-test Rob. p-value
ASC_CAR 0.931176 0.0 0.149727 6.219169 4.997935e-10
ASC_SM 2.014524 0.0 0.294670 6.836552 8.112178e-12
BETA_COST_HWH -1.766405 0.0 0.198007 -8.920916 0.000000e+00
BETA_COST_OTHER -0.828576 0.0 0.132385 -6.258836 3.878613e-10
BETA_DIST -6.174887 0.0 0.861775 -7.165311 7.760459e-13
BETA_TIME_CAR_CL -1.480684 0.0 2.973269 -0.497999 6.184849e-01
BETA_TIME_CAR_REF -15.285957 0.0 5.001630 -3.056195 2.241655e-03
BETA_TIME_PT_CL -1.069380 0.0 2.160096 -0.495061 6.205570e-01
BETA_TIME_PT_REF -5.426451 0.0 1.538572 -3.526941 4.203910e-04
BETA_WAITING_TIME -0.039497 0.0 0.009740 -4.054926 5.015009e-05
beta_ScaledIncome_10_inf 0.033239 0.0 0.109399 0.303836 7.612531e-01
beta_ScaledIncome_4_6 0.087575 0.0 0.281559 0.311036 7.557735e-01
beta_ScaledIncome_6_8 -0.178503 0.0 0.436064 -0.409350 6.822829e-01
beta_ScaledIncome_8_10 -0.094509 0.0 0.413156 -0.228749 8.190639e-01
beta_ScaledIncome_minus_inf_4 0.058319 0.0 0.191729 0.304173 7.609964e-01
coef_age_65_more -0.037623 0.0 0.177962 -0.211412 8.325657e-01
coef_haveChildren -0.021627 0.0 0.122668 -0.176307 8.600530e-01
coef_haveGA -0.824816 0.0 1.679053 -0.491239 6.232574e-01
coef_highEducation -0.017555 0.0 0.124911 -0.140543 8.882307e-01
coef_individualHouse -0.123526 0.0 0.275458 -0.448441 6.538351e-01
coef_male 0.155699 0.0 0.328803 0.473532 6.358334e-01
coef_moreThanOneBike -0.402377 0.0 0.820967 -0.490126 6.240447e-01
coef_moreThanOneCar 1.120108 0.0 2.255210 0.496676 6.194179e-01
sigma_s 1.000000 1.0 2.007941 0.498023 6.184681e-01


Total running time of the script: (0 minutes 0.275 seconds)

Gallery generated by Sphinx-Gallery