Análisis estadístico y exploración de modelos para mejorar la representatividad nacional en datos de conveniencia en línea.
Loading...
Date
2024
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Las investigaciones mediante encuestas muestran un incremento en las últimas décadas, utilizándose en diversas áreas de investigación para obtener de manera científica las opiniones de las personas. Con el avance de las tecnologías, especialmente el uso de internet, se realizan numerosas encuestas en línea debido a su rapidez y menor costo comparado con encuestas presenciales. Sin embargo, estas encuestas generalmente producen muestras no probabilísticas, lo que implica que pueden tener sesgos y esto limita la capacidad de realizar inferencias poblacionales con los datos. Aunque en muchos casos esto no representa un problema, si se desean hacer inferencias poblacionales, no se puede utilizar una muestra de este tipo.
Este estudio evalúa varias metodologías aplicables a una muestra no probabilística obtenida de internet, con el objetivo de mejorar la representatividad de la población en dicha muestra y permitir la realización de inferencias poblacionales. Se iguala la distribución de hombres y mujeres en tres grupos etarios utilizando tres metodologías diferentes: factores de expansión, método de sobremuestreo y un enfoque híbrido de sobremuestreo y submuestreo. Cada una de estas metodologías presenta variaciones en su funcionamiento, permitiendo evaluar si alguna de ellas iguala de mejor manera la distribución de la muestra con la de la población.
La muestra utilizada se obtiene de la página Mo2, que contiene datos sobre el uso de tiempo de la población chilena. Tras aplicar estas metodologías, se observa que los tres métodos entregan valores promedios similares de uso de tiempo, aunque en ciertas actividades algunos métodos muestran ligeras diferencias. Si bien todos los métodos logran su objetivo de igualar la distribución de la muestra no probabilística con la de la población, el método SmoteTomek (método híbrido) muestra un rendimiento deficiente al eliminar datos. Por ello, se recomienda sustituir este método por otro que funcione de manera diferente al ya aplicado.
Este estudio destaca la viabilidad de utilizar diversas metodologías para mejorar la representatividad de muestras no probabilísticas obtenidas en línea. Cualquiera de los métodos discutidos puede ser útil para este propósito, proporcionando información valiosa para futuras investigaciones y proyectos de encuestas.
Survey-based research has shown an increase in recent decades, being utilized across various research areas to scientifically gather people's opinions. With the advancement of technologies, especially the use of the internet, numerous online surveys are conducted due to their speed and lower cost compared to in-person surveys. However, these surveys typically yield non-probabilistic samples, which may introduce biases and thereby limit the ability to make population inferences with the data. While this may not be problematic in many cases, it precludes the use of non-probabilistic samples when population inferences are desired. This study evaluates several methodologies applicable to a non-probabilistic sample obtained from the internet, aiming to enhance the sample's representativeness of the population and enable population inferences. The distribution of males and females across three age groups are equalized using three different methodologies: weighting factors, oversampling method, and a hybrid approach of oversampling and undersampling. Each of these methodologies varies in its approach, allowing an assessment of which one best matches the sample distribution with that of the population. The sample used is obtained from the Mo2 website, which contains data on time use among the Chilean population. After applying these methodologies, it is observed that all three methods provide similar average values for time use, although slight differences are noted in certain activities. While all methods achieve their goal of equalizing the distribution of the non-probability sample with that of the population, the SmoteTomek method (hybrid method) performs poorly in data elimination. Therefore, it is recommended to replace this method with another that operates differently from the one already applied. This study underscores the feasibility of employing diverse methodologies to enhance the representativeness of non-probabilistic samples obtained online. Any of the discussed methods can serve this purpose effectively, providing valuable insights for future survey research and projects.
Survey-based research has shown an increase in recent decades, being utilized across various research areas to scientifically gather people's opinions. With the advancement of technologies, especially the use of the internet, numerous online surveys are conducted due to their speed and lower cost compared to in-person surveys. However, these surveys typically yield non-probabilistic samples, which may introduce biases and thereby limit the ability to make population inferences with the data. While this may not be problematic in many cases, it precludes the use of non-probabilistic samples when population inferences are desired. This study evaluates several methodologies applicable to a non-probabilistic sample obtained from the internet, aiming to enhance the sample's representativeness of the population and enable population inferences. The distribution of males and females across three age groups are equalized using three different methodologies: weighting factors, oversampling method, and a hybrid approach of oversampling and undersampling. Each of these methodologies varies in its approach, allowing an assessment of which one best matches the sample distribution with that of the population. The sample used is obtained from the Mo2 website, which contains data on time use among the Chilean population. After applying these methodologies, it is observed that all three methods provide similar average values for time use, although slight differences are noted in certain activities. While all methods achieve their goal of equalizing the distribution of the non-probability sample with that of the population, the SmoteTomek method (hybrid method) performs poorly in data elimination. Therefore, it is recommended to replace this method with another that operates differently from the one already applied. This study underscores the feasibility of employing diverse methodologies to enhance the representativeness of non-probabilistic samples obtained online. Any of the discussed methods can serve this purpose effectively, providing valuable insights for future survey research and projects.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Industrial.
Keywords
Población Estadísticas, Encuestas Chile, Encuestas