gretl-devel@gretlml.univpm.it
by Edward Ugarte
Here are the Hansl scripts I’ve been working on — reproducible, licensed, and purposeful.
🔗 github.com/Edward-Ugarte?tab=repositories
3 weeks
- 1
- 0
Panel time subsampling glitch
by Sven Schreiber
Hi,
reproducible example of the error I'm getting when I try to specify the
month elements with single digits in the panel context (only):
<hansl>
# plain monthly time series for comparison
open hamilton
smpl 1975:1 1980:6 # works with single-digit month numbers
# artificial monthly panel
T = 24
N = 10
TN = T*N
pd = 12
nulldata TN --preserve
setobs T 1:1 --stacked-time
setobs pd 2000:1 --panel-time
smpl 2000:02 2001:01 --time # works with double-digit month numbers
smpl full
smpl 2000:3 2001:5 --time # fails with single-digit month numbers
</hansl>
This is gretl 2025b.
thanks
sven
3 weeks, 3 days
- 4
- 3
gretl-devel@gretlml.univpm.it
by Edward Ugarte
# === Preparación de datos ===
clear # Limpia cualquier conjunto de datos previo
nulldata 1000 # Crea 1000 observaciones en Gretl
# === Generación de ingresos simulados ===
scalar media_ingreso = 1750 # Media esperada de ingresos
scalar desvio_ingreso = 500 # Variabilidad esperada
series ingresos_sim = normal(media_ingreso, desvio_ingreso) # Simulación de ingresos
# === Cálculo de tamaño muestral con error proporcional ===
scalar Z = 1.96 # Nivel de confianza del 95%
scalar media_sim = mean(ingresos_sim)
scalar E = 0.05 * media_sim # Margen de error dinámico (5% de la media)
scalar sigma_sim = sd(ingresos_sim) # Desviación estándar real
# Tamaño muestral sin corrección
scalar n_simple = (Z^2 * sigma_sim^2) / (E^2)
# Tamaño muestral con corrección por población finita
scalar N_real = 1000 # Población ajustada
scalar n_ajustado = n_simple / (1 + (n_simple / N_real))
# === Cálculo de intervalos de confianza ===
scalar n = floor(n_ajustado) # Tamaño muestral final ajustado
scalar x = mean(ingresos_sim) # Media de la muestra generada
scalar s = sd(ingresos_sim) # Desviación estándar de la muestra
# --- Intervalo SIN corrección por población finita ---
scalar E_sin = Z * (s / sqrt(n))
scalar inf_sin = x - E_sin
scalar sup_sin = x + E_sin
# --- Intervalo CON corrección por población finita ---
scalar f = n / N_real
scalar E_con = E_sin * sqrt(1 - f)
scalar inf_con = x - E_con
scalar sup_con = x + E_con
# === Cálculo de límites poblacionales ===
scalar inf_pobl_sin = x - E_sin
scalar sup_pobl_sin = x + E_sin
scalar inf_pobl_con = x - E_con
scalar sup_pobl_con = x + E_con
# === Prueba de hipótesis para la media ===
scalar mu_0 = 1700 # Valor poblacional de referencia
scalar Z_obs = (x - mu_0) / (s / sqrt(n))
scalar Z_critico = Z
# === Mostrar resultados ===
printf "\n📊 Tamaño muestral y error proporcional\n"
printf "Margen de error dinámico (5%% de la media): %.2f\n", E
printf "Tamaño muestral sin corrección: %.2f\n", n_simple
printf "Tamaño muestral con corrección finita: %.2f\n", n
printf "\n➡ Intervalos de confianza:\n"
printf "Sin corrección: [%.2f, %.2f]\n", inf_sin, sup_sin
printf "Con corrección: [%.2f, %.2f]\n", inf_con, sup_con
printf "\n✅ Límites poblacionales estimados:\n"
printf "Sin corrección: [%.2f, %.2f]\n", inf_pobl_sin, sup_pobl_sin
printf "Con corrección: [%.2f, %.2f]\n", inf_pobl_con, sup_pobl_con
printf "\n📌 Prueba de hipótesis para la media poblacional\n"
printf "H0: Media poblacional = %.2f\n", mu_0
printf "Media muestral: %.2f\n", x
printf "Estadístico de prueba Z: %.2f\n", Z_obs
printf "Valor crítico: ±%.2f\n", Z_critico
if abs(Z_obs) > Z_critico
printf "\n🚀 Se **rechaza** la hipótesis nula (H0). La media es significativamente diferente.\n"
else
printf "\n✅ No hay suficiente evidencia para rechazar H0. La media podría ser %.2f.\n", mu_0
endif
series X1 = normal(0, 1) # Ejemplo de regresora
series X2 = uniform(0, 10)
series X3 = normal(5, 2)
# Transformar ingresos en categorías discretas para Poisson
series ingresos_cat = ceil(ingresos_sim / 500) # Agrupamos en intervalos de 500
# Aplicar regresión de Poisson
poisson ingresos_cat const X1 X2 X3
# Configuración sin borrar datos
set verbose off
# Aplicar regresión de Poisson sin transformación adicional
poisson ingresos_cat const X1 X2 X3
nulldata 1000
set verbose off
# === Paso 2: Generación de ingresos simulados ===
series ingresos_sim = normal(1750, 500)
series ingresos_cat = ceil(ingresos_sim / 500) # Agrupar ingresos en categorías discretas
# Definir estadísticas clave
scalar media_ingreso = mean(ingresos_sim)
scalar desvio_ingreso = sd(ingresos_sim)
scalar sigma2 = desvio_ingreso^2
# Definir regresoras
series X1 = normal(0, 1)
series X2 = uniform(0, 10)
series X3 = normal(5, 2)
# === Paso 3: Muestreo Bayesiano con Gibbs Sampling ===
scalar beta0 = media_ingreso + mean(normal(0, sqrt(sigma2)))
scalar beta1 = mean(normal(0, 1))
scalar beta2 = mean(normal(0, 1))
scalar beta3 = mean(normal(0, 1))
matrix beta0_iter = zeros(1000, 1)
matrix beta1_iter = zeros(1000, 1)
matrix beta2_iter = zeros(1000, 1)
matrix beta3_iter = zeros(1000, 1)
loop i=1..1000
scalar epsilon = mean(normal(0, sqrt(sigma2)))
scalar ingresos_pred = beta0 + beta1 * mean(X1) + beta2 * mean(X2) + beta3 * mean(X3) + epsilon
beta0 = mean(normal(mean(ingresos_pred), sqrt(sigma2)))
beta1 = mean(normal(mean(X1), sqrt(sigma2)))
beta2 = mean(normal(mean(X2), sqrt(sigma2)))
beta3 = mean(normal(mean(X3), sqrt(sigma2)))
beta0_iter[i, 1] = beta0
beta1_iter[i, 1] = beta1
beta2_iter[i, 1] = beta2
beta3_iter[i, 1] = beta3
endloop
# === Paso 4: Transformación a series para visualización ===
series beta0_series = beta0_iter[,1]
series beta1_series = beta1_iter[,1]
series beta2_series = beta2_iter[,1]
series beta3_series = beta3_iter[,1]
# === Paso 5: Construcción del cuadro de resultados Bayesianos ===
printf "\n📊 **Estimaciones finales Bayesianas:**\n"
printf "Variable dependiente: ingresos_cat\n"
printf "Desviaciones típicas basadas en muestreo Gibbs\n"
printf "-----------------------------------------------------------\n"
printf " Coeficiente Desv. típica\n"
printf "-----------------------------------------------------------\n"
printf " Const %.5f %.5f\n", mean(beta0_series), sd(beta0_series)
printf " X1 %.5f %.5f\n", mean(beta1_series), sd(beta1_series)
printf " X2 %.5f %.5f\n", mean(beta2_series), sd(beta2_series)
printf " X3 %.5f %.5f\n", mean(beta3_series), sd(beta3_series)
printf "-----------------------------------------------------------\n\n"
# Mostrar estadísticas adicionales
printf "📌 **Estadísticas del ajuste Bayesiano:**\n"
printf "Media de la vble. dep. %.5f\n", mean(ingresos_cat)
printf "D.T. de la vble. dep. %.5f\n", sd(ingresos_cat)
printf "Criterio de Akaike %.5f\n", 2 * (mean(beta0_series) + mean(beta1_series) + mean(beta2_series) + mean(beta3_series))
Author: Edward Ugarte
Economist specializing in statistical modeling, Bayesian inference, and econometric scripting in open-source environments.
Package Description
This package integrates Poisson regression and Bayesian Gibbs Sampling, providing robust statistical inference for discrete count models. Designed for use within Gretl, it enables sample size estimation, interval analysis, and parameter uncertainty quantification in econometric studies.
Why Poisson and Bayesian Methods Were Used
Poisson regression is commonly used when the dependent variable consists of count data, such as income categories. It assumes that the mean and variance of the dependent variable are equal, which is characteristic of the Poisson distribution.
Bayesian Gibbs Sampling complements Poisson regression by incorporating parameter uncertainty into the model, allowing for a probabilistic approach rather than relying on a single point estimate. By iterating over posterior distributions, Gibbs Sampling refines parameter estimates, making it useful when sample sizes are small or when prior information is valuable.
Key Features
- Poisson regression modeling for discrete dependent variables.
- Bayesian Gibbs Sampling framework for parameter uncertainty quantification.
- Optimized sample size calculation, considering finite population corrections.
- Interval estimation and hypothesis testing integrated within Gretl.
- Modular structure for adaptability and usability across different econometric analyses.
Warnings When Using Poisson Regression
- Overdispersion: Poisson regression assumes that the mean and variance of the dependent variable are equal. If the variance is significantly larger, a negative binomial model may be more suitable.
- High prevalence of zeros: If the dependent variable contains excessive zeros, a zero-inflated Poisson model may be required.
- Collinearity among predictors: Strong correlations between independent variables can lead to unreliable coefficient estimates.
Warnings When Using Bayesian Methods
- Choice of priors: Poorly chosen priors can bias results or slow convergence in Gibbs Sampling. Prior distributions should reflect reasonable assumptions based on domain knowledge.
- Slow convergence in Gibbs Sampling: If chains are highly dependent, convergence may be slow, requiring diagnostics such as trace plots to verify stability.
- Sensitivity in small samples: Bayesian methods, especially Gibbs Sampling, rely on iterative updates. When sample sizes are small, the results may depend strongly on the choice of priors.
Compatibility
- Tested on Gretl version 2025b with stable results across Monte Carlo simulations.
- Designed for integration within Gretl’s econometric environment, ensuring seamless execution of both Poisson and Bayesian models.
Author's Background
Edward Ugarte specializes in Monte Carlo simulations, finite population corrections, and econometric modeling. His work focuses on developing accessible, open-source econometric tools that enhance the efficiency of statistical estimations in research environments.
Installation
- Import the package into Gretl.
- Run the scripts according to the usage instructions.
Let me know if you want any refinements or additional details. Always ready to enhance the package for clarity and efficiency.
3 weeks, 3 days
- 3
- 5
Re: [Gretl-users] gretl 2025b released
by Sven Schreiber
[moved to devel]
Am 05.06.2025 um 18:40 schrieb Allin Cottrell:
>
> Windows users: Note that there's now an ARM64 package available:
> https://gretl.sourceforge.net/win32/
Thanks, Allin. I will close the corresponding ticket on the tracker in a
minute.
Some more or less cosmetic remarks on the Windows download page:
- Aren't there some limitations with the Arm builds, and should they
perhaps be mentioned? The MPI thing is maybe only for experts, but I
thought that X13 and so on also are not for Arm? Or is that then simply
covered by emulation? (Only on Win11?)
- The top of the page says that Windows XP is required and SSE2. OK, so
the XP thing refers to the 32-bit build. But SSE2 is x64-only, if I
understand correctly, so now with the Arm build this is no longer
exactly true, or is it?
- The standard 64bit build is labeled with "Intel" processor, but in the
Windows world the AMD chips obviously have been very important (as
opposed to the Mac). What's meant is x64 here, I guess.
As I said, pretty much cosmetic.
cheers
sven
3 weeks, 6 days
- 1
- 0
- ← Newer
- 1
- Older →