One common approach to handle covariate measurement error in Generalized Linear Models (GLM) is classical error modeling. In the past 20 years, classical error modeling has been brought to (Non-Parametric) Maximum Likelihood (NPML) es- timation, by means of finite mixture modeling: the supposedly continuous true score is modeled as a discrete (multinomial) static latent variable, and is handled as a part of the model. Nonetheless, the true score is not allowed to vary over time: if the true score has own underlying dynamics, these are either unaccounted for or mistaken for measurement error, or possibly both. The aim of the present paper is to formulate a joint model for the outcome variable, the covariate observed with error (measure- ment model), and the true score model that accounts for the underlying dynamics in the true score. The true score and its dynamics are modeled non-parametrically as a first-order latent (hidden) Markov chain. Estimation is done extending the NPML approach, in a full maximum likelihood environment with a well-know modification of the EM algorithm (forward-backward algorithm). From an applied researcher per- spective, our methodology can safely handle both the case where the latent underly- ing characteristic is stable over time, as well as providing a suitable framework even when changes across measurement occasions are substantial. Within a GLM frame- work, it is demonstrated, by means of extensive simulation studies, that this is cru- cial to get correct estimates of the regression coefficients, as well as good coverages. In the real-data application, the effect of heart rate on the occurrence of cardiovas- cular diseases in a sample of Chinese elderly patients is measured. Modeling the true (unobserved) heart rate and its dynamics - which, in elderly patients, are likely to be non negligible - will be showed to allow a correct assessment of risk factors of cardio- vascular diseases occurrence.

Covariate measurement error in generalized linear models for longitudinal data: a latent Markov approach

Roberto Di Mari;Antonio Punzo;
2018-01-01

Abstract

One common approach to handle covariate measurement error in Generalized Linear Models (GLM) is classical error modeling. In the past 20 years, classical error modeling has been brought to (Non-Parametric) Maximum Likelihood (NPML) es- timation, by means of finite mixture modeling: the supposedly continuous true score is modeled as a discrete (multinomial) static latent variable, and is handled as a part of the model. Nonetheless, the true score is not allowed to vary over time: if the true score has own underlying dynamics, these are either unaccounted for or mistaken for measurement error, or possibly both. The aim of the present paper is to formulate a joint model for the outcome variable, the covariate observed with error (measure- ment model), and the true score model that accounts for the underlying dynamics in the true score. The true score and its dynamics are modeled non-parametrically as a first-order latent (hidden) Markov chain. Estimation is done extending the NPML approach, in a full maximum likelihood environment with a well-know modification of the EM algorithm (forward-backward algorithm). From an applied researcher per- spective, our methodology can safely handle both the case where the latent underly- ing characteristic is stable over time, as well as providing a suitable framework even when changes across measurement occasions are substantial. Within a GLM frame- work, it is demonstrated, by means of extensive simulation studies, that this is cru- cial to get correct estimates of the regression coefficients, as well as good coverages. In the real-data application, the effect of heart rate on the occurrence of cardiovas- cular diseases in a sample of Chinese elderly patients is measured. Modeling the true (unobserved) heart rate and its dynamics - which, in elderly patients, are likely to be non negligible - will be showed to allow a correct assessment of risk factors of cardio- vascular diseases occurrence.
2018
9788891910233
covariate measurement error
errors–in–variables
longitudinal data
Generalized linear models
latent Markov models
File in questo prodotto:
File Dimensione Formato  
paper_SIS2018.pdf

solo gestori archivio

Dimensione 3.58 MB
Formato Adobe PDF
3.58 MB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.11769/363410
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact