CMStatistics 2021: Start Registration
View Submission - CMStatistics
B0984
Title: Handling endogeneity in linear quantile regression models for longitudinal data Authors:  Marco Alfo - University La Sapienza, Rome (Italy) [presenting]
Francesca Martella - La Sapienza University of Rome (Italy)
Abstract: Individual specific effects are often included in regression models for longitudinal data. These are used to account for the effect of time-constant unobserved covariates, often referred to as unobserved heterogeneity, which may be considered as random variables (random effects). In this case, the individual-specific effects may also account for inter-individual dependence. However, the assumptions on the dependence between the random effects and the observed covariates, that is between unobserved and observed covariates, are a crucial point when integrating out the random effects to derive the model likelihood. If dependence is not properly taken into account, the resulting estimator may be inconsistent. There exist solutions in the linear regression context and extensions to the context of generalized linear models, the correlated random effect estimator. Its properties are due to the geometric properties of the (generalized) least-squares method. When we move to linear quantile regression, the same properties are not valid any longer and this estimator may not be optimal. We review methods to deal with endogeneity in the context of mixed linear quantile regression for longitudinal data and propose a general solution by exploiting a finite mixture specification.