2. Notation and Preliminary Results

In this section, we first introduce a general specification for the linear mixed model on the log-scale along with a basic result conditional on the variance components. Then, we shortly describe the GIG distribution that will be considered in further analyses.

2.1. The Log-Normal Mixed Model

Let us consider a n-dimensional vector of strictly positive responses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {y}$$\end{document} ; once defined \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {w}=\log \mathbf {y}$$\end{document} , a linear mixed model is assumed:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbf {w}= \mathbf {X}\varvec{\beta }+\mathbf {Zu}+\varvec{\varepsilon }, \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\beta }\in \mathbb {R}^p$$\end{document} is a vector of fixed effects, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {u}\in \mathbb {R}^m$$\end{document} is a vector of random effects and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\varepsilon }\in \mathbb {R}^{n}$$\end{document} is the vector of residuals. The design matrices are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}\in \mathbb {R}^{n\times p}$$\end{document} , that is assumed to be full rank, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}\in \mathbb {R}^{n\times m}$$\end{document} . The following Bayesian hierarchical model will be studied:

(1) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned}&\mathbf {w}|\mathbf {u}, \varvec{\beta }, \sigma ^2\sim \mathcal {N}_n\left( \mathbf {X}\varvec{\beta }+\mathbf {Zu}, \mathbf {I}_n\sigma ^2 \right) ;\\&\quad \mathbf {u}|\tau ^2_1,...,\tau ^2_q\sim \mathcal {N}_m\left( \mathbf {0}, \mathbf {D}\right) ,\ \mathbf {D}=\oplus ^q_{s=1}\mathbf {I}_{m_s}\tau _s^2. \end{aligned} \end{aligned}$$\end{document}

Note that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q\ge 1$$\end{document} random factors are allowed, so that q different variances related to the random components \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\tau }^2=(\tau ^2_1,...,\tau ^2_q)$$\end{document} are included in the model. Therefore, it is possible to split the vector of random effects in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {u}=[\mathbf {u}_1^T,...,\mathbf {u}_s^T,...,\mathbf {u}_q^T]^T$$\end{document} , where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {u}_s\in \mathbb {R}^{m_s}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum _{s=1}^q m_s=m$$\end{document} . The design matrix of the random effects can be partitioned too: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}=[\mathbf {Z}_1\cdots \mathbf {Z}_s\cdots \mathbf {Z}_q]$$\end{document} . We note that the design matrix of the random effects is not necessarily non-singular. For an introduction to the use of these models in the behavioral sciences framework, see, e.g., Jackman (Reference Jackman2009, Chapter 7).

The introduced model is fairly general. All standard one and multi-ways ANOVA models as well as mixed models suitable for the analysis of repeated measures with both nested and crossed effects (Baayen et al., Reference Baayen, Davidson and Bates2008) can be obtained as special cases. ANCOVA models, accounting for the effect of possible covariates, are also encompassed by (1), including models that allow for possible nonlinear effects of these covariates whose shape cannot be anticipated: in fact, spline regression can be represented by means of mixed models (see Crainiceanu et al., Reference Crainiceanu, Ruppert and Wand2005). Equation (1) covers situations in which the assumption of independence between random effects fails, provided no additional parameter is involved: more specifically, if known positive matrices replace \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {I}_{m_s}$$\end{document} , (1) can be reparameterized to allow for correlated random effects (Hobert and Casella, Reference Hobert and Casella1996). On the contrary, models involving additional parameters describing the correlation between random effects are beyond the scope of (1) and thereby of our analysis. Nonetheless, a discussion of models in which correlated random effects are specified within grouping factors can be found in Sect. 7.

We now restate a known result on the posterior distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\beta }$$\end{document} in order to set notations and define quantities that will be used later on.

Proposition 1

Considering the model (1) with a flat improper prior on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\beta }$$\end{document} then:

(2) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{\beta }|\sigma ^2, \varvec{\tau }^2, \mathbf {w}\sim \mathcal {N}_p\left( \bar{\varvec{\beta }},\mathbf {V}_\beta \right) , \end{aligned}$$\end{document}

where:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} \mathbf {V}_\beta&=\left( \frac{\left( \mathbf {X}^T\mathbf {X}\right) }{\sigma ^2}+ \mathbf {X}^T\mathbf {M}\mathbf {X}\right) ^{-1} ,\quad \bar{\varvec{\beta }}=\mathbf {V}_\beta \left( \frac{\mathbf {X}^T\mathbf {X}}{\sigma ^2}\hat{\varvec{\beta }}+\mathbf {X}^T\mathbf {M}\mathbf {X}\tilde{\varvec{\beta }}\right) ,\\ \mathbf {M}&=\left( \frac{\mathbf {V}^{-1}_Z}{\sigma ^2}-\frac{\mathbf {P}_Z}{\sigma ^2} \right) ,\quad \hat{\varvec{\beta }}= \left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\mathbf {X}^T\mathbf {y},\quad \tilde{\varvec{\beta }}= \left( \mathbf {X}^T\mathbf {M}\mathbf {X}\right) ^{-1}\mathbf {X}^T\mathbf {M}\mathbf {y},\\ \mathbf {P}_Z&=\mathbf {Z}\left( \mathbf {Z}^T\mathbf {Z}\right) ^{-}\mathbf {Z}^T,\quad \mathbf {V}_Z^{-1}=\mathbf {Z}\left( \mathbf {Z}^T\mathbf {Z}\right) ^{-} \left( \left( \mathbf {Z}^T\mathbf {Z}\right) ^{-}+\frac{\mathbf {D}}{\sigma ^2}\right) ^{-1}\left( \mathbf {Z}^T\mathbf {Z}\right) ^{-}\mathbf {Z}^T, \end{aligned} \end{aligned}$$\end{document}

and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( \mathbf {Z}^T\mathbf {Z}\right) ^{-}$$\end{document} is the Moore–Penrose inverse of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}^T\mathbf {Z}$$\end{document} .

As anticipated in the introduction, in this paper we focus on the estimation of the expectation of y and on predictive distributions. Let the vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {z}}$$\end{document} represent a point in the covariates space conditionally on which we can be interested in estimating the expectation of y. More specifically, let us first consider:

(3) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbb {E}\left[ \tilde{y}|\varvec{\beta },\sigma ^2,\varvec{\tau }^2\right] =\theta _m(\tilde{\mathbf {x}})=\exp \left\{ \tilde{\mathbf {x}}^T\varvec{\beta }+\frac{1}{2}\left( \sigma ^2+\sum _{s=1}^q \tau _s^2\right) \right\} , \end{aligned}$$\end{document}

where the random effects are integrated out. We use the notation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{y}$$\end{document} instead of y to emphasize we are working conditionally on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {z}}$$\end{document} . The expectation of y conditional on the random effects is another quantity that can be relevant in prediction problems:

(4) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbb {E}\left[ \tilde{y}|\mathbf {u},\varvec{\beta },\sigma ^2 \right] =\theta _c(\tilde{\mathbf {x}},\tilde{\mathbf {z}})=\exp \left\{ \tilde{\mathbf {x}}^T\varvec{\beta }+\tilde{\mathbf {z}}^T\mathbf {u}+\frac{\sigma ^2}{2} \right\} . \end{aligned}$$\end{document}

Finally, the posterior predictive distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\tilde{y}|\mathbf {y})$$\end{document} and its posterior moments are further quantities to investigate. Note that:

(5) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p\left( \tilde{y}|\mathbf {y}\right) \propto \int _{\varvec{\Theta }} p\left( \tilde{y}|\varvec{\theta }\right) p\left( \varvec{\theta }|\mathbf {y}\right) \mathrm {d}\varvec{\theta }, \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\theta }=(\varvec{\beta }, \mathbf {u}, \sigma ^2, \varvec{\tau }^2)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Theta $$\end{document} is the parameter space. In practice, the posterior expectation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {E}\left[ \tilde{y}|\mathbf {y}\right] $$\end{document} might be used to predict unobserved values like missing values or unsampled units.

2.2. The Generalized Inverse Gaussian Distribution

In this paper, we assume a GIG prior for the variance components. In general, a random variable V is GIG distributed, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \sim GIG (\lambda ,\delta ,\gamma )$$\end{document} , if its density can be written as follows:

(6) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(v)=\left( \frac{\gamma }{\delta } \right) ^\lambda \frac{1}{2K_\lambda (\delta \gamma )}v^{\lambda -1}\exp \left\{ -\frac{1}{2}\left( \delta ^2 v^{-1}+\gamma ^2 v\right) \right\} \mathbf {1}_{\mathbb {R}^+}. \end{aligned}$$\end{document}

If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta > 0$$\end{document} , the permissible values for the other parameters are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \ge 0$$\end{document} when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda <0$$\end{document} , and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma > 0$$\end{document} if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda =0$$\end{document} . If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \ge 0$$\end{document} , then \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} should be strictly positive. The first reason to consider the GIG is that many important distributions may be obtained as special cases. For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda >0$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma > 0$$\end{document} , the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Gamma (\lambda ,\gamma ^2/2)$$\end{document} distribution emerges as the limit when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \rightarrow 0$$\end{document} . An inverse-gamma is obtained when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda < 0$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta >0$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \rightarrow 0$$\end{document} ; an inverse Gaussian distribution is obtained when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda =-\frac{1}{2}$$\end{document} . A uniform distribution over the range (0, A) for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sqrt{V}$$\end{document} implies that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(v) \propto v^{-1/2}\mathbf {1}_{(0,A)}$$\end{document} , which may approximated by the density of a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$GIG(0.5,\delta ,(2A^2)^{-1})$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \rightarrow 0$$\end{document} and truncated at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A^2$$\end{document} . This special case is relevant to discuss the uniform prior on the standard deviation advocated by Gelman (Reference Gelman2006). For more details on the GIG distribution see Bibby and Sørensen (Reference Bibby and Sørensen2003).

3. Theoretical Results

In this section, we study the existence of moments for the posterior distributions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(\tilde{\mathbf {x}})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(\tilde{\mathbf {x}},\tilde{\mathbf {z}})$$\end{document} , defined in (3) and (4), and for the posterior predictive distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\tilde{y}|\mathbf {y})$$\end{document} (5). As anticipated in the introduction, we assume GIG distributions for the hyper-parameters:

(7) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sigma ^2\sim & {} GIG\left( \lambda _\sigma ,\delta _\sigma ,\gamma _\sigma \right) , \end{aligned}$$\end{document}

(8) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tau _s^2\sim & {} GIG\left( \lambda _{\tau ,s},\delta _{\tau ,s},\gamma _{\tau ,s}\right) ,\ \forall s. \end{aligned}$$\end{document}

Before stating the main result of this section, let us define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_s\in \mathbb {R}^{p\times p}$$\end{document} as a matrix whose entries are all 0s with the exception of the first \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l \times l$$\end{document} square block \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_{s;1,1}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l=p-rank\{ \mathbf {X}^T\left( \mathbf {I}-\mathbf {P_Z} \right) \mathbf {X}\}$$\end{document} is the rank deficiency of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}^T\left( \mathbf {I}-\mathbf {P_Z} \right) \mathbf {X}$$\end{document} and it coincides with the number of columns of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}$$\end{document} that are included in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} too. To simplify the statement of our result, it is useful to work with a modified design matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}_o$$\end{document} obtained by placing the columns included in both \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} as the first l columns, without loss of generality. Consequently, we note that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_{s;1,1}$$\end{document} coincides with the inverse of upper left \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l \times l$$\end{document} block on the diagonal of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}_o^T\left( \mathbf {Z}(\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {C}_s (\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {Z}^T\right) \mathbf {X}_o$$\end{document} , where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {C}_s$$\end{document} is the null matrix with the exception of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {I}_{m_s}$$\end{document} as block on the diagonal in correspondence to the s-th variance component of the random effect. Eventually, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}_{o}$$\end{document} is the covariate pattern of the new observation ordered consistently with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}_o$$\end{document} .

Proof

See appendix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\square $$\end{document}

Few comments on Theorem 1 are in order. We first note that the conditions on the existence of posterior moments depend only on constraints on the tail parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} . Moreover, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(\tilde{\mathbf {x}})$$\end{document} requires a condition on the parameters of all variance components prior, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(\tilde{\mathbf {x}},\tilde{\mathbf {z}})$$\end{document} and the posterior predictive distribution need only a condition on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\sigma ^2)$$\end{document} , to ensure the finiteness of the posterior moments.

Statement (i) parallels the result by Fabrizi and Trivisano (Reference Fabrizi and Trivisano2016) for the log-normal linear model: the square of the moment order r is multiplied by the leverage associated with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} , i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}^T\left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\tilde{\mathbf {x}}$$\end{document} . The same condition on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\sigma }$$\end{document} appears also for the moments of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(\tilde{\mathbf {x}})$$\end{document} .

As far as the posterior predictive distribution, it concerns, i.e., case (iii), the existence of its posterior moments is related only to the term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2$$\end{document} . It must be noted that, unlike case (i), the quantity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r^2$$\end{document} enters the condition as a separate term, making the value on the right side of the constraint rapidly increasing with the moment order. The result is in line with the higher variability that characterizes the posterior predictive distribution with respect to the posteriors of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(\tilde{\mathbf {x}},\tilde{\mathbf {z}})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(\tilde{\mathbf {x}})$$\end{document} .

From Theorem 1 and its proof, it is apparent that, generally speaking, a prior containing an exponential term in the form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\exp \{-c\omega ^2 \}$$\end{document} must be given as prior for the generic variance component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega ^2$$\end{document} , where c is set in order to have finite moments up to a pre-specified order. This helps us to understand which special cases within the GIG family and which distributions outside this group can be considered. Popular choices for priors on the variance components such as Jeffrey’s priors, uniform (both on the variance and on the standard deviation), half-t (including half-Cauchy) do not contain the exponential term in question. Other priors such as the inverse gamma (that is a special case of the GIG distribution when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma \rightarrow 0$$\end{document} ) or the log-normal, even if they contain an exponential term, cannot be used as this term does not go to 0 when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega ^2\rightarrow +\infty $$\end{document} .

Other distributions, outside the GIG family, can be considered as prior for the variance components, as for instance the half-normal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$HN(\zeta )$$\end{document} , mentioned as reasonable prior for the standard deviation by Gelman (Reference Gelman2006), provided that a small hyper-parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta $$\end{document} is chosen. In view of Theorem 1, it can be shown that, for example, the prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \sigma \sim HN(\zeta _\sigma )$$\end{document} should be specified in compliance with the following constraint:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \zeta _\sigma <\sqrt{\frac{1}{r+r^2\tilde{\mathbf {x}}^T\left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\tilde{\mathbf {x}}}}. \end{aligned}$$\end{document}

Nonetheless, we note that to satisfy this constraint the tail decay of such a prior might be too rapid and an excessive amount of prior information might be included in the model, whereas the GIG distribution provides useful tools to control it and to specify a more suitable prior distribution.

3.1. The Random Intercepts Model

The constraints on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma ^2_{\tau ,s}$$\end{document} that appear in condition (ii) of Theorem 1 look rather complicated as we assumed a general structure for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} . To better understand the meaning of the result, we can show the results obtained when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} is simpler. Let us consider the following simple random intercept model, that can be applied in the analysis of repeated measurement data where a random effect is introduced to account for within individual correlation:

(9) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} w_{ij}&=\log \left( y_{ij}\right) = \mathbf {x}_{ij}^T\varvec{\beta } +v_j +\varepsilon _{ij};\ j=1,...,m;\ i=1,...,n_j;\\&\varepsilon _{ij}|\sigma ^2 {\mathop {\sim }\limits ^{ind}} N\left( 0,\sigma ^2\right) ,\ v_j|\tau ^2 {\mathop {\sim }\limits ^{ind}} N\left( 0,\tau ^2\right) . \end{aligned} \end{aligned}$$\end{document}

In the random intercepts model, the number m of the columns of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} coincides with the number of clusters observed in the data and each row contains a single 1, denoting that the correspondent unit belongs to the cluster (typically the subject in longitudinal data), and 0s otherwise. Moreover, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}_o$$\end{document} is the simple design matrix, since the first column is the usual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{1}_n$$\end{document} vector corresponding to the general intercept and the first element of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {x}_{o,i}$$\end{document} is 1. Moreover, it is easy to verify that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l=p-rank\{\mathbf {X}^T\left( \mathbf {I}-\mathbf {P_Z} \right) \mathbf {X}\}=1$$\end{document} and therefore the unique non-null entry of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_s$$\end{document} is the first element of the first column. Eventually, exploiting the particular structure of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} , after some algebra, it is possible to verify that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_{s;1,1}=m^{-1}$$\end{document} (i.e., the inverse of the number of groups determined by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} ). Provided that priors (7) and (8) are adopted, the condition on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma ^2_\sigma $$\end{document} does not change, whereas the eventual condition on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma ^2_{\tau }$$\end{document} simplifies to:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \gamma ^2_{\tau } > r + \frac{r^2}{m}. \end{aligned}$$\end{document}

4. Practical Implementation Issues

In this section, we consider two issues related to practical implementation. In Sect. 4.1, we consider how to set GIG priors’ hyper-parameters. Theorem 1 provides lower bounds for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} parameters; we complement this information offering some guidance on how to remove the dependence on specific \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} in the choice of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} and on how to choose values for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta $$\end{document} parameters. The setting of these parameters can be relevant in the analysis of small samples. Specifically, we devise a weakly informative strategy based on the uniform shrinkage principle that will lead us to the specification of Gamma priors on the variance components.

In Section 4.2, we provide some details on how to generate samples from the posterior of model parameters (and the random effects). We only need a direct Gibbs sampler where elementary samplers can be used for each of the full conditionals: a nice feature that depends on the conjugacy relationship between the normal and the GIG distributions. To encourage the use of the method by practitioners and automatically set the advised priors, functions included in the BayesLN package can be used (Gardini et al., Reference Gardini, Fabrizi and Trivisano2020).

4.1. Hyper-Parameters Choice

The lower bounds in Theorem 1 depend on r, the order of posterior moments for which we need to impose the existence. In principle, a priori we would set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} s to the lower bound allowing the existence of moments up to the order r we are interested in, with the aim of avoiding priors with exceedingly light tails. In practice, it is advisable to set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} parameters somewhat larger than the existence lower bound to avoid numerical instability caused by dealing with integrals that although finite are very large. We can achieve this, for instance, by choosing values of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} s allowing the existence of moments up to the order \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r+c$$\end{document} , with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c>0$$\end{document} . A discussion on the selection of c can be found in Section S1 in the supplementary material. In short, choices of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c \ge 0.5$$\end{document} are advisable. Throughout the simulations and applications of this paper, we will use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=1$$\end{document} .

The existence conditions stated in Theorem 1 also depend on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \tilde{\mathbf {x}}^T\left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\tilde{\mathbf {x}}$$\end{document} . Since we want moments of order r to exist for all the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} included in the analysis, the dependence on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\mathbf {x}}$$\end{document} can be removed by setting:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \gamma _\sigma =\sqrt{(r+c)+(r+c)^2h_m}, \end{aligned}$$\end{document}

with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_m=\max _{i \in s_p} \tilde{\mathbf {x}}_i^T\left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\tilde{\mathbf {x}}_i$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_p$$\end{document} is the set of points in the covariates’s space for which we are interested in making predictions. If the moments of the posterior predictive distribution are required, then \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(r+c)^2$$\end{document} must be included in the previous condition. In the same line, we propose to set:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \gamma _{\tau ,s}=\sqrt{(r+c)+(r+c)^2l_m}, \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l_m=\max _{i \in s_p} \tilde{\mathbf {x}}_{o,i}^T\mathbf {L}_s\tilde{\mathbf {x}}_{o_i}$$\end{document} .

In general, the advice is to fix the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} equal to the most restrictive condition (i.e., the greatest one) with respect to the quantities that are of interest in the analysis.

As expected, constraints on the existence of posterior moments lead to priors with light tails for the variance components. In order to avoid excessively informative priors, we propose a weakly informative strategy for the selection of remaining parameters. To illustrate our heuristic, let us work on the notable special case where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q=1$$\end{document} . Consequently, for simplicity, we denote with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2$$\end{document} the variance component associated with the unique random effect. Some remarks on the generalization to the case \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q>1$$\end{document} are reported later. Let the intraclass correlation coefficient be defined as:

(10) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho =\frac{\tau ^2}{\sigma ^2+\tau ^2}. \end{aligned}$$\end{document}

This quantity is of interest in the analysis of hierarchical model, both from a statistical viewpoint and from the applied perspective. Chaloner (Reference Chaloner1987) proposes to specify \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho \sim \mathcal {U}(0,1)$$\end{document} to obtain good frequentist properties for the parameters estimates. The uniform prior distribution for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} has been extensively studied and used (Daniels, Reference Daniels1999). If both variance components \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2$$\end{document} are GIG distributed, Favaro et al. (Reference Favaro, Lijoi and Pruenster2012) show that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} follows a normalized generalized inverse Gaussian distribution, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho \sim N-GIG(\lambda _\tau ,\delta _\tau ,\gamma _\tau ,\lambda _\sigma ,\delta _\sigma , \gamma _\sigma )$$\end{document} . If we assume, for the time being, to set the same hyper-parameters for both priors, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \sigma ^2 \sim GIG(\lambda ,\delta ,\gamma ) \, and \, \tau ^2 \sim GIG(\lambda ,\delta ,\gamma )$$\end{document} , then the normalized GIG density for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} simplifies to:

(11) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(\rho )=\frac{K_{2\lambda }\left( \gamma ^2\delta ^2\left[ \frac{1}{\rho }+\frac{1}{1-\rho }\right] \right) }{2\left[ K_\lambda \left( \gamma \delta \right) \right] }\left[ \rho (1-\rho )\right] ^{\lambda -1},\quad \rho \in (0,1). \end{aligned}$$\end{document}

Moreover, considering the target functionals of the analysis, the most restrictive threshold should be chosen as the value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} .

The resulting density is a function of the product \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \gamma $$\end{document} . To simplify the parameter specification, we consider the special case \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \rightarrow 0$$\end{document} that frees the distribution from the dependence on both parameters and that makes the choice of different \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} s due to different constraining equations immaterial for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\rho )$$\end{document} .

When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \rightarrow 0$$\end{document} , the density (11) can be simplified further by using a small argument approximation to the Bessel K function:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(\rho )\simeq \frac{\Gamma (|2\lambda |)}{\Gamma (|\lambda |)^2}\left[ \rho (1-\rho )\right] ^{|\lambda |-1},\quad \rho \in (0,1). \end{aligned}$$\end{document}

Setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda =1$$\end{document} implies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho \sim \mathcal {U} (0,1)$$\end{document} . If we consider \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi =\frac{\tau ^2}{\sigma ^2}$$\end{document} , a one-to-one transformation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho $$\end{document} , the prior implied by the above choices is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\phi )=(1+\phi ^2)^{-1}$$\end{document} , that is the solution proposed for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi $$\end{document} by Ye (Reference Ye1994) within the reference prior framework (Berger and Bernardo, Reference Berger and Bernardo1992).

The strategy can be summarized as:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sigma ^2\sim GIG\left( \lambda =1, \delta =\varepsilon , \gamma _m\right) ,\quad \tau ^2\sim GIG\left( \lambda =1, \delta =\varepsilon , \gamma _m\right) ; \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _m$$\end{document} is the most restrictive existence conditions for the considered quantities and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon $$\end{document} is some small constant close to 0 (e.g., 0.01). This proposal can be straightforwardly extended to the case \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q>1$$\end{document} assuming that a uniform prior is specified for every \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _s=\tau _s(\tau _s+\sigma ^2)^{-1}$$\end{document} . These marginal priors are retrieved setting all the priors on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _s$$\end{document} as independent and equal GIG distributions with parameters fixed according to the described strategy; i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2_s\sim GIG(\lambda =1, \delta =\varepsilon , \gamma _m),\ \forall s$$\end{document} .

We note that under the described setting, if the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} parameter is set to be positive, a gamma prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {G}\left( \lambda ,\gamma ^2/2\right) $$\end{document} for each variance component is approximately assumed. As a consequence, a normal-gamma prior is specified marginally for the random effects vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {u}$$\end{document} . This prior setting is not new to the literature as it was introduced by Griffin and Brown (Reference Griffin and Brown2010) as prior for the coefficients of a linear model. Frühwirth-Schnatter and Wagner (Reference Frühwirth-Schnatter, Wagner, Bernardo, Bayarri, Berger, Dawid, Heckerman, Smith and West2011) and Fabrizi et al. (Reference Fabrizi, Ferrante and Trivisano2018) already use this distribution as prior for random intercepts. They note that these priors encourage shrinkage of the random intercepts toward the general intercept and more so as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} gets smaller. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda =1$$\end{document} , the gamma distribution degenerates to the exponential distribution, and in that case the normal-gamma is a Laplace distribution. This particular prior is known also as Bayesian Lasso and is characterized by a spike in 0. In general, the degree of shrinkage determined by the prior can be increased setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} near 0, whereas increasing this parameter has an opposite effect.

The main difference between Griffin and Brown (Reference Griffin and Brown2010), Frühwirth-Schnatter and Wagner (Reference Frühwirth-Schnatter, Wagner, Bernardo, Bayarri, Berger, Dawid, Heckerman, Smith and West2011), and the present proposal is represented by the approach used to deal with the scale (or rate) parameter of the gamma prior. In fact, the cited papers specify an hyper-prior on it. This solution is not viable here because of the restrictions on the parameter space due to the posterior moments existence condition.

4.2. Computational Algorithms

An appealing characteristic of the adoption of GIG priors (7) and (8) for the variance components of model (1) is their conditional conjugacy. This can be exploited to derive easy to sample full conditionals for the model parameters in order to implement a Gibbs sampler algorithmFootnote 1 able to generate random samples from their posterior distributions:

(12) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\sigma ^2|\varvec{\beta },\mathbf {u}, \varvec{\tau }^2,\mathbf {w}\sim GIG\left( \lambda _\sigma -\frac{n}{2}, \sqrt{\left( \mathbf {w}-\mathbf {X}\varvec{\beta }-\mathbf {Zu}\right) ^{T}\left( \mathbf {w}-\mathbf {X}\varvec{\beta }-\mathbf {Zu}\right) +\delta _\sigma ^2}, \gamma _\sigma \right) ; \end{aligned}$$\end{document}

(13) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\tau _s^2|\varvec{\beta },\mathbf {u}, \sigma ^2,\varvec{\tau }^2_{-s},\mathbf {w}\sim GIG \left( \lambda _{\tau , s}-\frac{m_s}{2}, \sqrt{\mathbf {u}_s^T\mathbf {u}_s+\delta _{\tau ,s}^2}, \gamma _{\tau ,s}\right) ,\qquad s=1,...,q; \end{aligned}$$\end{document}

(14) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\mathbf {u}|\varvec{\beta },\sigma ^2, \varvec{\tau }^2,\mathbf {w}\sim \mathcal {N}_{m}\left( \mathbf {V}_\mathbf {u}\mathbf {Z}^T\left( \mathbf {w}-\mathbf {X}\varvec{\beta }\right) ,\sigma ^2 \mathbf {V}_\mathbf {u}\right) ; \end{aligned}$$\end{document}

(15) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\varvec{\beta }|\mathbf {u},\sigma ^2, \varvec{\tau }^2,\mathbf {w}\sim \mathcal {N}_p\left( \left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\mathbf {X}^T\left( \mathbf {w}-\mathbf {Zu}\right) , \sigma ^2\left( \mathbf {X}^T\mathbf {X}\right) ^{-1}\right) ; \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {V}_\mathbf {u}=\left( \mathbf {Z}^T\mathbf {Z}+\sigma ^2\mathbf {D}^{-1}\right) ^{-1}$$\end{document} . The sampler has been implemented in C++ within the function LN_hierarchical() in the R package BayesLN.

5. Simulations

In this section, we present two simulation exercises focused on simple models specified on the logarithm of the response variable. In the first place, we consider a special case of (9) where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {x}_{ij}^T\varvec{\beta }=\mu $$\end{document} , that is a one-way ANOVA model. The aim is to assess the frequentist properties of posterior means as predictors of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m=\exp \left\{ \mu +\frac{\tau ^2+\sigma ^2}{2}\right\} $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{c}(v_j)=\exp \{\mu +v_j+\frac{\tau ^2}{2}\}$$\end{document} under different choices for the priors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\sigma ^2)$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\tau ^2)$$\end{document} . We also include summaries of the posterior of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{c}(v_j)$$\end{document} conditional on the variance components, i.e., treating the variances as known, as benchmarks. We devote special attention to the analysis of small samples, where the impact of the priors is more apparent. A second simulation exercise, with a data generating process characterized by the presence of a continuous covariate, aims at assessing the impact of alternative prior choices on the posterior distribution of regression coefficients and posterior predictive distributions. Details about this second simulation exercise are presented in Section S4 of the supplementary material.

In the first simulation exercise, we generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B=2000$$\end{document} samples from model (9) assuming \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mathbf {x}_{ij}^T \varvec{\beta } = \mu $$\end{document} under 24 different scenarios obtained crossing the following choices for the parameters: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_j=(2,5)$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m=10$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi =\tau ^2/\sigma ^2=(0.5,1,2)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2=(0.05,0.25,0.5,0.75)$$\end{document} . The general mean in the logarithmic scale is set to 0, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu =0$$\end{document} . The considered grid of values for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2$$\end{document} is aimed at covering the range log-scale variances most common in applications. The estimates that require Monte Carlo methods are based on 4000 iterations, after the first 1000 iterations are discarded as burn-in. The point predictors we compare are:

1. (i) The posterior means of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(v_j)$$\end{document} when priors are:

   (16) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(\mu )\propto 1,\ \ \sigma ^2\sim GIG\left( 1, 0.01, \gamma _{\text {m}}\right) ,\ \tau ^2\sim GIG\left( 1, 0.01, \gamma _{\text {m}}\right) , \end{aligned}$$\end{document}
   where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\text {m}}=\max \{\gamma _\sigma , \gamma _{\tau ,1}\}= \sqrt{3+3^2m^{-1}}$$\end{document} , according to the suggestions provided in Sect. 4.1 in order to assure the posterior variance existence. The predictors will be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_m^{GIG}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }^{GIG}_c(v_j)$$\end{document} , and the function LN_hierarchical of the BayesLN package is used to estimate the model;

2. (ii) The posterior means of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(v_j)$$\end{document} when priors are:

   (17) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(\mu )\propto 1, \ \ \sigma ^2\sim IG(1,1),\ \ \tau ^2\sim IG(1,1), \end{aligned}$$\end{document}
   that will be labeled as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_m^{IG}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }^{IG}_c(v_j)$$\end{document} . These priors for the variance components are suggested as default choice in the BANOVA package (Wedel and Dong, Reference Wedel and Dong2020). The algorithm for sampling from the posterior distributions is implemented in Stan (Carpenter et al., Reference Carpenter, Gelman, Hoffman, Lee, Goodrich and Betancourt2017);

3. (iii) The posterior means of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(v_j)$$\end{document} under small parameters inverse gamma (“Jeffreys like”) priors (Carpenter et al., Reference Wagenmakers, Love, Marsman, Jamil, Ly and Verhagen2018a):

   (18) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(\mu )\propto 1, \ \ \sigma ^2\sim IG(0.001,0.001),\ \ \tau ^2\sim IG(0.001,0.001), \end{aligned}$$\end{document}
   that will be labeled as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_m^{J}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }^{J}_c(v_j)$$\end{document} . The algorithm for sampling from the posterior distributions is implemented in Stan. An alternative choice of the IG parameters and namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2\sim IG(1,0.001)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2\sim IG(1,0.001)$$\end{document} is also considered. For brevity, results related to these latter alternatives are reported in section S3 of the supplementary material;

4. (iv) A conditional Bayes predictors in which \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2$$\end{document} are assumed to be known for the case of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} prediction:

   (19) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\theta }_m^{c}=\exp \left\{ \bar{w}+\frac{\sigma ^2+\tau ^2}{2}-\frac{3\left( \sigma ^2+n_g\tau ^2\right) }{2n} \right\} . \end{aligned}$$\end{document}
   In line with Zellner (Reference Zellner1971), we can show that (19) reaches minimum frequentist MSE among the predictors of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} having form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\exp {\{\bar{w}\}}$$\end{document} . For benchmarking purposes, a minimum MSE estimator conditioned to the variance components for the functional \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(\nu _j)$$\end{document} is useful too. In this case, a decision to take is the estimator class, since the global sample mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{w}$$\end{document} as the only argument of the exponential function appears to be not appropriated. A heuristic strategy to obtain a conditioned estimator might be based on the derivation of the Bayes estimator under relative quadratic loss, obtaining:
   (20) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\theta }_{c}^{c}\left( v_j\right) =\exp \left\{ \frac{\sigma ^2}{\sigma ^2+n_g\tau ^2}\left( \frac{\tau ^2n_g}{\sigma ^2}\bar{w}_{.j}-\bar{w}\right) +\frac{\sigma ^2}{2}-\frac{3}{2}\frac{\sigma ^2}{\sigma ^2+n_g\tau ^2}\left( \tau ^2+\frac{\sigma ^2}{n}\right) \right\} . \end{aligned}$$\end{document}
   The derivations of these estimators can be found in Section S2 of online supplementary materialFootnote 2.

Bias, root mean square error (RMSE), frequentist coverage and average interval width are reported for estimators of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} (for which we use the generic notation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_m$$\end{document} ). Specifically, we calculate:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bias\left( \hat{\theta }_m\right)&=\frac{1}{B}\sum _{k=1}^B \left( \hat{\theta }_m^{(k)}-\theta _m\right) ;\ RMSE(\hat{\theta }_m)=\sqrt{\frac{1}{B}\sum _{k=1}^B \left( \hat{\theta }_m^{(k)}-\theta _m\right) ^2 };\\ Cov\left( \hat{\theta }_m\right)&=\frac{1}{B}\sum _{k=1}^B \varvec{1}_{\left[ \hat{L}^{(k)};\hat{U}^{(k)}\right] }\left( \theta _m^{(k)}\right) ;\ Wid\left( \hat{\theta }_m\right) =\frac{1}{B}\sum _{k=1}^B \left( \hat{U}^{(k)}-\hat{L}^{(k)}\right) ; \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{L}^{(k)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{U}^{(k)}$$\end{document} are computed as the 0.025 and 0.975 quantiles of the posterior distributions in question. In these formulas, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_m^{(k)}$$\end{document} is the estimate of the true overall expectation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} at Monte Carlo iteration k and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{L}^{(k)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{U}^{(k)}$$\end{document} are the estimated lower bound and upper bound for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$95\%$$\end{document} intervals.

To jointly evaluate the m different estimates for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(v_j),\ j=1,...,m$$\end{document} , an average evaluation of the estimates, that we denote with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\bar{\theta }}_c$$\end{document} , is required. Therefore, the relative absolute bias (RABias), the relative RMSE (RRMSE), the average frequentist coverage (ACo.) and the average interval width (AWi.) are studied.

More in detail we define the quantities:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} RABias\left( \hat{\bar{\theta }}_c\right)&=\frac{1}{J}\sum _{j=1}^J\left| \frac{1}{B}\sum _{k=1}^B \left( \frac{\hat{\theta }_c^{(k)}\left( v_j\right) -\theta _c^{(k)}\left( v_j\right) }{\theta _c^{(k)}\left( v_j\right) }\right) \right| ;\\ RRMSE\left( \hat{\bar{\theta }}_c\right)&=\frac{1}{J}\sum _{j=1}^J\sqrt{\frac{1}{B}\sum _{k=1}^B \left( \frac{\hat{\theta }_c^{(k)}\left( v_j\right) -\theta _c^{(k)}\left( v_j\right) }{\theta _c^{(k)}\left( v_j\right) }\right) ^2 };\\ ACo\left( \hat{\bar{\theta }}_c\right)&=\frac{1}{J}\sum _{j=1}^J\frac{1}{B}\sum _{k=1}^B \varvec{1}_{\left[ \hat{L}^{(k)}\left( v_j\right) ;\hat{U}^{(k)}\left( v_j\right) \right] }\left( \theta _c^{(k)}\left( v_j\right) \right) ;\\ AWi\left( \hat{\bar{\theta }}_c\right)&=\frac{1}{J}\sum _{j=1}^J\frac{1}{B}\sum _{k=1}^B \left( \hat{U}^{(k)}\left( v_j\right) -\hat{L}^{(k)}\left( v_j\right) \right) ; \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{L}^{(k)}(v_j)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{L}^{(k)}(v_j)$$\end{document} are calculated as the 0.025 and 0.975 percentiles of the posterior distributions and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\theta }_c^{(k)}(v_j)$$\end{document} is the estimate of the j-th true group specific expectation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c^{(k)}(v_j)$$\end{document} at Monte Carlo iteration k.

In Tables 1 and S1 (the latter in Section S3 of the online supplementary material), we can see the frequentist properties of the point estimators of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} : problems occurring to posterior means under inverse gamma priors for variance components ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta ^{IG}_m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta ^{J}_m$$\end{document} ) are apparent. In fact, extremely high values for bias and RMSE are detected. These anomalies can be considered as the numerical equivalent of the analytical non-finiteness of posterior moments. On the other hand, under our proposed prior, the estimators reach RMSE values that keep the same magnitude of the ones obtained for the benchmark \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m^c$$\end{document} , showing their reliability.

Moving to results about group means (Tables 2 and S2), we note that observing numerically the analytical problems proved for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c^{IG}(v_j)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c^{J}(v_j)$$\end{document} is harder. In these cases, explosive numerical situations are not evident, even if we can say that our proposal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c^{GIG}(v_j)$$\end{document} systematically outperforms the other considered estimators.

Table 1.

Bias and RMSE for the considered estimators of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} in the different scenarios with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_g=2$$\end{document} .

In the supplementary material, results about the frequentist properties of credible intervals are reported for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} (Table S3) and averaged for the group specific expectations (Table S4). Considering both the inferential problems, we can summarize the results as follows: under all priors, systematic deviations from the nominal coverage level of 0.95 are not evident. Considering the intervals width, it emerges that the ones produced under GIG priors are almost always narrower than intervals produced under inverse gamma priors. This is particularly evident in the case of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} . In particular, larger intervals are obtained under IG(1, 1) prior for variance components: probably it is not an appropriate choice in cases of variance components near to 0, as often happens in log-transformed data.

Table 2.

RABias and RRMSE for the considered estimators of the group-specific expectations in the different scenarios with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_g=2$$\end{document} .

In Section S3 of the supplementary material, results about this simulation setting under three further prior settings are presented. The first two explore the sensitivity of posterior with respect to different choices of the GIG scale parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta $$\end{document} . Specifically, we consider the settings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta =0.1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta =0.001$$\end{document} . It is interesting to note that we obtain results extremely close to those under prior (16). The third simulation setting involves the alternative choice for the IG hyper-parameters described below formula (18). Results point in the direction of non-existence of posterior moments showing also issues in the estimation of the group means.

As far as the second simulation exercise, we mentioned above is concerned, the results (reported in Section S4 of the supplementary material) show that different priors on the variance components do not induce remarkable changes on the estimation of a regression coefficient, whereas the problems affecting the moments of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(v_j)$$\end{document} emerges also for the posterior predictive distribution, in line with theoretical findings.

6. Real Data Application: Reading Times

Several applications in psychology and cognitive sciences have as central output the time requested to perform some tasks. By definition, times are positive numbers and often show a positively skewed distribution: for these reasons, it is common to analyze their logarithmic transformations.

The data we use to apply our methodologies were originally collected by Gibson and Wu (Reference Gibson and Wu2013) in order to investigate the presence of a notable difference between times requested to process a subject-extracted relative clause (SRC) and an object-extracted relative clause (ORC) in Chinese language. In particular, times (in milliseconds) required to read the head noun of a Chinese clause are registered under a repeated measure design characterized by two factors: subject and reading item.

This dataset has been analyzed also by Sorensen and Vasishth (Reference Sorensen and Vasishth2015), that proposed a Bayesian linear mixed model specified for the reading time logarithm. Here, we consider the model formulation with two random intercepts related to the grouping factors:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} w_{ijk}=\log \left( y_{ijk}\right) =\beta _0+\beta _1 x_{i}+u_j+v_k+\varepsilon _{ijk}, \end{aligned}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{ijk}$$\end{document} is the reading time observed for subject \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j=1,...,37$$\end{document} , reading item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=1,...,15$$\end{document} and clause type \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i=1,2$$\end{document} . More in detail, it is fixed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i=-1$$\end{document} in case of SRC, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i=1$$\end{document} for ORC. The random effects are aimed at accounting for the potential within subject and within item correlation, and they are assumed to be independently distributed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_j|\tau _u^2\sim \mathcal {N}\left( 0,\tau _u^2\right) $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_k|\tau _v^2\sim \mathcal {N}\left( 0,\tau _v^2\right) $$\end{document} . Both of them are assumed independent from the error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon _{ijk}|\sigma ^2\sim \mathcal {N}\left( 0,\sigma ^2\right) $$\end{document} . Beyond the usual inference on the model parameters that are related to times in the log-scale, to have a clearer interpretation of the studied phenomenon the estimation and prediction of quantities in the original data scale might be relevant. For example, the expectation conditioned on clause type and marginalized with respect both the random effects:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \theta _m\left( x_i=\pm 1\right) =\exp \left\{ \beta _0\pm \beta _1+\frac{\tau ^2_u+\tau ^2_v+\sigma ^2}{2} \right\} . \end{aligned}$$\end{document}

On the other hand, the expectation specific of a particular subject and item (individual) is:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \theta _c\left( x_i,u_j,v_k\right) =\exp \left\{ \beta _0+x_i\beta _1+u_j+v_k+\frac{\sigma ^2}{2} \right\} . \end{aligned}$$\end{document}

From an interpretative viewpoint, it can be useful to target the expected time conditioned to only a particular random effect, e.g., integrating out only the subject and considering only a particular item:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \theta _c\left( x_i,v_k\right) =\exp \left\{ \beta _0+x_i\beta _1+v_k+\frac{\tau _u^2+\sigma ^2}{2} \right\} . \end{aligned}$$\end{document}

Obtaining posterior summaries of these functionals might help in understanding the phenomenon and communicating results.

More technically, the design matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}$$\end{document} for the random effects is constituted by two blocks, in order to define two distinct random intercepts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}=\left[ \mathbf {Z}_v\ \ \mathbf {Z}_u \right] $$\end{document} . The elements of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}_v\in \mathbb {R}^{n\times 15}$$\end{document} assume value 1 in column k if the observation is related to the item k and 0 otherwise; on the other hand, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}_u\in \mathbb {R}^{n\times 37}$$\end{document} assume value 1 in column j if the observation is related to subject j and 0 otherwise.

As a consequence, the rank deficiency of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}\left( \mathbf {I}-\mathbf {P}_Z \right) \mathbf {X}$$\end{document} is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l=1$$\end{document} and it is due to the fixed effect intercept, which is linearly dependent with respect to both \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}_v$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {Z}_u$$\end{document} .

Hyper-parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} in priors (7) and (8) are set along the lines of Section 4.1 in order to assure the existence of the first two posterior moments. For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^2$$\end{document} , we apply condition (i) in Theorem 1 by setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=3$$\end{document} for numerical stability and calculating the maximum leverage: we obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\sigma }=1.742$$\end{document} . For the random effects variances, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_v\in \mathbb {R}^{2\times 2}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {L}_u\in \mathbb {R}^{2\times 2}$$\end{document} must be computed, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}_o$$\end{document} coincides with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}$$\end{document} since the rank deficiency is due to the intercept. Given that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l=1$$\end{document} , the unique non-null elements coincide with the inverse of the first elements of the matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}^T\left( \mathbf {Z}(\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {C}_v (\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {Z}^T\right) \mathbf {X}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {X}^T\left( \mathbf {Z}(\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {C}_u (\mathbf {Z}^T\mathbf {Z})^{-}\mathbf {Z}^T\right) \mathbf {X}$$\end{document} , where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {C}_v=\text {diag}\left( \mathbf {I}_{15}, \varvec{0}_{37}\right) $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {C}_u=\text {diag}\left( \varvec{0}_{15}, \mathbf {I}_{37}\right) $$\end{document} . The deduced numerical conditions are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\tau ,v}=2.046$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\tau ,v}=2.434$$\end{document} ; therefore, the latter value is chosen for all the GIG priors tail parameters since it is the more restrictive condition. We stress that the available package BayesLN (Gardini et al., Reference Gardini, Fabrizi and Trivisano2020) automatically produces these computations to facilitate the usage by practitioners. The code required to obtain the results presented in this section is available as supplementary material, whereas details on the MCMC convergence diagnostics are reported in Section S5 of the supplementary material.

Table 3.

Posterior means and standard deviations obtained for the whole dataset ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=547$$\end{document} ) under three considered prior specifications.

In Table 3, the posterior means and standard deviations obtained for the complete dataset ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=547$$\end{document} ) under prior settings (16), (17) and (18) are reported. Posterior inference has been carried out both on basic model parameters and some conditional expectations of reading times. In particular, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(x_i=- 1)$$\end{document} represents the expected time requested to process a SRC estimated by the model, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(x_i=1)$$\end{document} is the time expected for an ORC. Another interesting output for these kind of models is the estimation of the response variable expectation within a particular group: for example, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(-1,u_{3})$$\end{document} represents the average reading time for item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j=3$$\end{document} in the SRC case and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _c(1,u_{3})$$\end{document} in the ORC case.

Table 4.

Posterior means and standard deviations obtained for a subset of the dataset ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=110$$\end{document} ) under three considered prior specifications.

Figure 1.

Posterior distributions of the marginal means \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(x_i=- 1)$$\end{document} under different priors for the variance components. The results obtained with the complete and the reduced data are shown.

We note that the issues that affect posterior moments of functionals in the original data scale are masked by the moderately large sample size. In fact, there are no clear symptoms of the fact that posterior results obtained under inverse gamma priors are theoretically meaningless, since they are MCMC estimates of integrals that are analytically not finite, as already noted in the simulation section. We also note that the inverse gamma prior with parameters both equal to 1 can be a largely informative prior for variances when their actual value is near to 0, as it often happens in the analysis of log-transformed data. In this application, the variance components ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2_u$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau ^2_v$$\end{document} ) posterior estimates are substantially higher than the ones obtained under the proposed GIG priors and the small-parameters inverse gamma priors.

Finally, we fit the same model under the three prior settings on a subset of the original dataset: we considered reading time observations from the first three clauses only ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=1,2,3$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=110$$\end{document} ). In Table 4, posterior results are displayed. The aim of this second exercise is to stress again the mathematical inconsistency of the conditional expectations posterior summaries in Table 3: we note that, in this case, the infiniteness of the target integrals is evident also from their MCMC estimates. The cause of this feature appears in Fig. 1 where the boxplots representing the posterior distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m(x_i=-1)$$\end{document} highlight the heavy tails obtained under IG priors for the reduced dataset. On the other hand, our prior specification allows to produce reliable estimates in any case, improving the readability of the log-normal mixed model results.

Supplementary material

In the supplementary material, the following information is reported. In Section S1, we complement the discussion on the choice of prior specification for the hyper-parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} contained in Sect. 4.1 of the main paper. In Section S2, the minimum MSE estimator conditioned to the variance components of the overall mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _m$$\end{document} is derived and its connection to the Bayesian framework is explained. This quantity is used as benchmark in the simulation study. In Section S3, some additional tables concerning the results of the simulation discussed in Section 5 of the paper are reported. Section S4 contains an additional simulation study in which covariates are included in the model, and the frequentist properties of the posterior predictive distribution are investigated. Section S5 reports the information about the convergence diagnostics of the MCMC algorithm used to fit the models compared in the application of Section 6. Eventually, the proof of Corollary 1 and some software details useful to estimate models with dependent random effects are contained in Section S6. All the R code used for the simulations and the application is available in a zipped folder.