Comparación de métodos de inferencia causal para mitigar el sesgo de no participación en campañas publicitarias
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La medición del impacto de campañas publicitarias es un desafío cuando existen sesgos debido al incumplimiento en el grupo de tratamiento. Este estudio compara distintos métodos para estimar el efecto causal de una campaña publicitaria digital, evaluando en particular la regresión lineal, el propensity score y el efecto causal promedio local (LATE) utilizando variables instrumentales.
Se desarrolla una simulación que contempla cuatro escenarios con distintos grados de influencia de la variable instrumental sobre la exposición al anuncio. Los resultados indican que la regresión simple no estima correctamente el impacto del anuncio debido al sesgo. El propensity score mejora la estimación, pero su precisión depende de la correcta especificación de los confusores. El modelo LATE ofrece estimaciones precisas cuando se cumple el supuesto de validez del instrumento, aunque su efectividad se reduce considerablemente si la variable instrumental tiene poca influencia sobre la exposición.
El análisis de sensibilidad confirma la estabilidad de los resultados. Además, se realiza una comparación de los métodos en términos de su interpretabilidad. La regresión lineal destaca por su facilidad de explicación. El propensity score, aunque más complejo de explicar, permite evitar el uso de todas las variables de control, facilitando la comprensión en contextos donde las variables son numerosas. El modelo LATE, por su parte, presenta un desafío mayor en términos de comprensión debido a la necesidad de entender el concepto de variable instrumental y los supuestos subyacentes.
Finalmente, se discuten posibles líneas de investigación futuras para mejorar la corrección del sesgo en la medición del impacto publicitario.
Measuring the impact of advertising campaigns is challenging when there are biases due to non-compliance in the treatment group. This study compares different methods to estimate the causal effect of a digital advertising campaign, specifically evaluating linear regression, propensity score, and local average treatment effect (LATE) using instrumental variables. A simulation is developed, considering four scenarios with varying degrees of influence of the instrumental variable on exposure to the ad. The results show that simple regression does not correctly estimate the impact of the ad due to bias. The propensity score improves the estimation, but its accuracy depends on the correct specification of confounders. The LATE model provides accurate estimates when the instrument validity assumption is met, although its effectiveness decreases significantly if the instrumental variable has little influence on exposure. Sensitivity analysis confirms the stability of the results. Additionally, a comparison of methods is made in terms of interpretability. Linear regression stands out for its ease of explanation. The propensity score, although more complex to explain, allows for avoiding the use of all control variables, facilitating understanding in contexts with numerous variables. The LATE model, on the other hand, presents a greater challenge in terms of comprehension due to the need to understand the concept of instrumental variables and the underlying assumptions. Finally, future research directions are discussed to improve bias correction in measuring advertising impact.
Measuring the impact of advertising campaigns is challenging when there are biases due to non-compliance in the treatment group. This study compares different methods to estimate the causal effect of a digital advertising campaign, specifically evaluating linear regression, propensity score, and local average treatment effect (LATE) using instrumental variables. A simulation is developed, considering four scenarios with varying degrees of influence of the instrumental variable on exposure to the ad. The results show that simple regression does not correctly estimate the impact of the ad due to bias. The propensity score improves the estimation, but its accuracy depends on the correct specification of confounders. The LATE model provides accurate estimates when the instrument validity assumption is met, although its effectiveness decreases significantly if the instrumental variable has little influence on exposure. Sensitivity analysis confirms the stability of the results. Additionally, a comparison of methods is made in terms of interpretability. Linear regression stands out for its ease of explanation. The propensity score, although more complex to explain, allows for avoiding the use of all control variables, facilitating understanding in contexts with numerous variables. The LATE model, on the other hand, presents a greater challenge in terms of comprehension due to the need to understand the concept of instrumental variables and the underlying assumptions. Finally, future research directions are discussed to improve bias correction in measuring advertising impact.
Description
Keywords
INFERENCIA CAUSAL, SESGO DE NO PARTICIPACIóN