Intuitive joint priors for high-dimensional Bayesian linear models
Lade...
Datum
Autor:innen
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Verlag
Sonstige Titel
Zusammenfassung
In high-dimensional Bayesian regression, prior distributions can decisively affect posterior inference.
These models often involve multilevel terms, grouped predictors, interactions, or dependent
covariates, yet standard shrinkage priors are typically formulated for exchangeable coefficients
and do not directly encode this structure. Moreover, coefficient-level hyperparameters can
obscure how prior choices affect predictive behavior, making it difficult to incorporate domain
knowledge. To address these problems, we develop a predictive-first framework in which we
specify priors on the coefficient of determination R2 and use variance decompositions to translate
these specifications into joint shrinkage priors on model parameters. Our construction induces
regularization across and within structured components of the linear predictor. We extend this
framework to multilevel models, grouped predictors, and dependent covariates. We show that the
resulting prior families have desirable shrinkage properties and introduce interpretable quantities
that help users understand how prior choices affect regularization. We evaluate the proposed
priors in simulation studies and real-data applications, examining when and to what extent
incorporating structure improves parameter recovery and out-of-sample prediction.
Beschreibung
Inhaltsverzeichnis
Schlagwörter
Bayesian inference, Bayesian workflow, prior specification, shrinkage priors, high-dimensional regression
Schlagwörter nach RSWK
Bayes-Verfahren, Regressionsmodell, Prädiktor, A-priori-Verteilung
