Resultados piloto encuesta 1

tesis
transporte
Resultados de piloto de encuesta sobre elección entre micro bus y tren entre Santiago y Rancagua
Published

May 11, 2026

Caracterización de la muestra

El primer piloto contó con 224 respuestas desde Netquest, de las cuales 96 respuestas fueron consideradas válidas (gente que efectivamente usaba el corredor).

Es un poco complejo el índice de incidencia (estamos bajo el 50% de opbacioón objetivo vs población encuestada). Ojalá se pueda hacer algo al respecto desde Netquest.

A continuación, se muestran algunos gráficos de caracterización:

Caract. Demográfica

Caract. viajes

Resultados Items Likert

Modelos preliminares

Se plantearon 3 modelos de Logit Mixto. Las especificaciones se muestran a continuación

MXL - Efecto panel + variables instrumentales (medibles)

MXL - Efecto panel + variables instrumentales + variables hedónicas

MXL - Indicadores de satisfacción por modo agregados

Resultados

A continuación, se muestran los resultados de las estimaciones de los modelos anteriores:

Table 1: Comparison of Mixed Logit Specifications (Hess Error-Component Spec, Panel Data)
Parameter Category Model 0 Model 1 Model 2
\(ASC_{train}\) (Fixed) – 0.0000 (–) 0.0000 (–) 0.0000 (–)
\(ASC_{bus}\) – -1.4232 (-4.64)*** -1.6964 (-5.02)*** -0.6692 (-0.38)
\(ASC_{micro}\) – -1.2302 (-2.57)** -1.2080 (-2.49)** -0.1689 (-0.09)
\(\beta_{tt}\) (Travel Time) Instrumental -0.01295 (-2.78)*** -0.01253 (-2.62)*** -0.01212 (-2.55)**
\(\beta_{cost}\) (Cost) Instrumental -0.000173 (-2.86)*** -0.000150 (-2.45)** -0.000151 (-2.48)**
\(\beta_{hw}\) (Headway) Instrumental -0.00452 (-1.26) -0.00334 (-0.92) -0.00332 (-0.90)
\(\beta_{var}\) (Variability) Instrumental -0.00107 (-0.18) -\(3.4 \times 10^{-5}\) (-0.01) -\(8.2 \times 10^{-5}\) (-0.01)
\(\beta_{lmt}\) (Last-Mile Time) Instrumental -0.00220 (-0.17) -0.00783 (-0.58) -0.00612 (-0.45)
\(\beta_{crowd}\) (Crowding) Hedonic – -0.1569 (-3.58)*** -0.1570 (-3.57)***
\(\beta_{aircon}\) (Air Conditioning) Hedonic – -0.1223 (-0.92) -0.1200 (-0.90)
\(\beta_{sellers}\) (Vendors) Hedonic – -0.1033 (-0.73) -0.0985 (-0.70)
Satisfaction Indices (sum of 9 Likert items per mode, range 9–45)
\(\beta_{satisf,train}\) Latent proxy – – -0.1628 (-3.54)***
\(\beta_{satisf,bus}\) Latent proxy – – -0.1040 (-1.96)*
\(\beta_{satisf,micro}\) Latent proxy – – -0.1551 (-3.58)***
Error Components (Hess et al. 2008 spec, single shared \(\sigma\))
\(\sigma_{panel}\) – 1.6260 (8.88)*** 1.6592 (8.90)*** 1.4820 (9.11)***
Fit Statistics
Observations / Individuals – 1225 / 96 1225 / 96 1225 / 96
Estimated parameters – 8 11 14
Final Log-Likelihood – -695.26 -688.38 -677.46
AIC – 1406.52 1398.77 1382.91
BIC – 1447.41 1454.98 1454.46
Adj. \(\rho^2\) (vs. Equal Shares) – 0.3450 0.3486 0.3560

Robust t-ratios entre parentesís. Significancias: *** p < 0.01, ** p < 0.05, * p < 0.10.

Código

MXL - Efecto panel

# ==============================================================================
# LOGIT MIXTO PANEL - ESPECIFICACION HESS et al. (2008)
# ==============================================================================
# Componentes de error iid (entre alternativas, no entre observaciones)
# en las TRES alternativas, con UNA sola sigma compartida.
# Identificacion garantizada por la independencia de los draws entre
# alternativas y la estructura panel.
# Referencia: Apollo Manual, seccion 6.2 / Figura 6.7.
# ==============================================================================

# -- Configuracion SLURM (procesado por submit_to_nlhpc.py) -------------------
# @slurm partition   = "debug"
# @slurm time        = "00:30:00"
# @slurm cores       = 44
# @slurm mem_per_cpu = 345
# @db C:/Users/Pablo/Desktop/tesis-magister/SP/netquest-soft_start-v2/data/master_db_netquest_soft_start_v2.csv


# 1. ENTORNO -------------------------------------------------------------------
if (!nzchar(Sys.getenv("SLURM_JOB_ID")) &&
  requireNamespace("rstudioapi", quietly = TRUE) &&
  rstudioapi::isAvailable()) {
  setwd(dirname(rstudioapi::getSourceEditorContext()$path))
}

# install.packages("apollo")

library(apollo)
library(dplyr)

rm(list = ls())
apollo_initialise()

# Cores: NLHPC lo dicta via SLURM; local default = 8
nCores_use <- as.integer(Sys.getenv("SLURM_CPUS_PER_TASK", unset = "8"))
cat("[INFO] Usando", nCores_use, "cores\n")

# Path BBDD
db_dir <- Sys.getenv(
  "NLHPC_DB_DIR",
  unset = "C:/Users/Pablo/Desktop/tesis-magister/SP/netquest-soft_start-v2/data/"
)
database <- read.csv(file.path(db_dir, "master_db_netquest_soft_start_v2.csv"))
names(database) <- trimws(names(database))
database[is.na(database)] <- 0

# FILTRO: Eliminar observaciones donde se eligio el opt-out (alternativa 4)
database <- subset(database, choice != 4)

# ==============================================================================
# 2. CONFIGURACION DE APOLLO
# ==============================================================================
apollo_control <- list(
  modelName       = "modelo3-Panel_Hess_SinOptOut - Netquest v2",
  modelDescr      = "Logit Mixto Panel - Hess et al. 2008 (EC iid en las 3 alternativas) - Sin Opt-Out - Netquest v2",
  indivID         = "id_response",
  mixing          = TRUE,
  nCores          = 8,
  outputDirectory = "output"
)

# ==============================================================================
# 3. DEFINICION DE PARAMETROS A ESTIMAR
# ==============================================================================
apollo_beta <- c(
  asc_train = 0,
  asc_bus   = 0,
  asc_micro = 0,

  # Nivel de Servicio
  b_tt      = 0,
  b_cost    = 0,
  b_lmt     = 0,
  b_hw      = 0,

  # Confort y Confiabilidad
  b_var     = 0,
  b_crowd   = 0,
  b_aircon  = 0,
  b_sellers = 0,

  # ----------------------------------------------------------------------------
  # CAMBIO CLAVE 1: una sola sigma compartida en lugar de sigma_bus y sigma_micro
  # Valor inicial 0.5 (no 0) para evitar partida en una region plana de la LL.
  # ----------------------------------------------------------------------------
  sigma_panel = 0.5
)

# Constante del tren como referencia (esto no cambia respecto al modelo anterior)
apollo_fixed <- "asc_train"

# ==============================================================================
# 4. CONFIGURACION DE DRAWS
# ------------------------------------------------------------------------------
# CAMBIO CLAVE 2: TRES draws normales independientes, uno por alternativa
# (antes habian dos: draw_bus y draw_micro). Ahora se incluye tambien draw_train.
# La independencia entre estos tres draws es lo que garantiza identificacion
# en datos panel pese a tener EC en TODAS las alternativas.
# ==============================================================================
apollo_draws = list(
  interDrawsType = "mlhs",
  interNDraws    = 2000,
  interUnifDraws = c(),
  interNormDraws = c("draw_train", "draw_bus", "draw_micro"),
  intraDrawsType = "halton",
  intraNDraws    = 0,
  intraUnifDraws = c(),
  intraNormDraws = c()
)

# ==============================================================================
# 5. COEFICIENTES ALEATORIOS
# ------------------------------------------------------------------------------
# Tres EC, todos con la MISMA sigma_panel (homocedasticidad). Cada EC usa
# un draw distinto, lo que los hace independientes entre alternativas pero
# constantes a lo largo de las T situaciones de cada persona (efecto panel).
# ==============================================================================
apollo_randCoeff = function(apollo_beta, apollo_inputs){
  randcoeff = list()
  randcoeff[["ec_train"]] = sigma_panel * draw_train
  randcoeff[["ec_bus"]]   = sigma_panel * draw_bus
  randcoeff[["ec_micro"]] = sigma_panel * draw_micro
  return(randcoeff)
}

apollo_inputs <- apollo_validateInputs()

# ==============================================================================
# 6. FUNCION DE PROBABILIDADES
# ------------------------------------------------------------------------------
# CAMBIO CLAVE 3: el EC se suma a las TRES utilidades (incluida la del tren),
# no solo a las J-1 alternativas como en el modelo anterior.
# ==============================================================================
apollo_probabilities <- function(apollo_beta, apollo_inputs, functionality = "estimate"){

  apollo_attach(apollo_beta, apollo_inputs)
  on.exit(apollo_detach(apollo_beta, apollo_inputs))

  P <- list()
  V <- list()

  V[["train"]] <- asc_train + ec_train +
    b_tt      * train_tt +
    b_cost    * train_cost +
    b_lmt     * train_lmt +
    b_hw      * train_headway +
    b_var     * train_arrival_time_variability +
    b_crowd   * train_crowding_ordinal +
    b_aircon  * train_aircon_ordinal +
    b_sellers * train_sellers_ordinal

  V[["bus"]]   <- asc_bus + ec_bus +
    b_tt      * bus_tt +
    b_cost    * bus_cost +
    b_lmt     * bus_lmt +
    b_hw      * bus_headway +
    b_var     * bus_arrival_time_variability +
    b_crowd   * bus_crowding_ordinal +
    b_aircon  * bus_aircon_ordinal +
    b_sellers * bus_sellers_ordinal

  V[["micro"]] <- asc_micro + ec_micro +
    b_tt      * micro_tt +
    b_cost    * micro_cost +
    b_lmt     * micro_lmt +
    b_hw      * micro_headway +
    b_var     * micro_arrival_time_variability +
    b_crowd   * micro_crowding_ordinal +
    b_aircon  * micro_aircon_ordinal +
    b_sellers * micro_sellers_ordinal

  mnl_settings <- list(
    alternatives = c(train = 1, bus = 2, micro = 3),
    avail        = list(train = av_train, bus = av_bus, micro = av_micro),
    choiceVar    = choice,
    utilities    = V
  )

  P[["model"]] <- apollo_mnl(mnl_settings, functionality)
  P <- apollo_panelProd(P, apollo_inputs, functionality)
  P <- apollo_avgInterDraws(P, apollo_inputs, functionality)
  P <- apollo_prepareProb(P, apollo_inputs, functionality)

  return(P)
}

# ==============================================================================
# 7. ESTIMACION Y RESULTADOS
# ==============================================================================
model <- apollo_estimate(apollo_beta, apollo_fixed, apollo_probabilities, apollo_inputs)

apollo_modelOutput(model)
apollo_saveOutput(model)