Generalized linear models with coarsened covariates: a practical Bayesian approach.
Psychol Methods
; 19(2): 281-99, 2014 Jun.
Article
in En
| MEDLINE
| ID: mdl-24364382
ABSTRACT
Coarsened covariates are a common and sometimes unavoidable phenomenon encountered in statistical modeling. Covariates are coarsened when their values or categories have been grouped. This may be done to protect privacy or to simplify data collection or analysis when researchers are not aware of their drawbacks. Analyses with coarsened covariates based on ad hoc methods can compromise the validity of inferences. One valid method for accounting for a coarsened covariate is to use a marginal likelihood derived by summing or integrating over the unknown realizations of the covariate. However, algorithms for estimation based on this approach can be tedious to program and can be computationally expensive. These are significant obstacles to their use in practice. To overcome these limitations, we show that when expressed as a Bayesian probability model, a generalized linear model with a coarsened covariate can be posed as a tractable missing data problem where the missing data are due to censoring. We also show that this model is amenable to widely available general-purpose software for simulation-based inference for Bayesian probability models, providing researchers a very practical approach for dealing with coarsened covariates.
Full text:
1
Collection:
01-internacional
Database:
MEDLINE
Main subject:
Linear Models
/
Bayes Theorem
Type of study:
Prognostic_studies
Limits:
Humans
Language:
En
Journal:
Psychol Methods
Journal subject:
PSICOLOGIA
Year:
2014
Document type:
Article
Country of publication:
EEUU
/
ESTADOS UNIDOS
/
ESTADOS UNIDOS DA AMERICA
/
EUA
/
UNITED STATES
/
UNITED STATES OF AMERICA
/
US
/
USA