Generalized g-Prior Methods

• Generalized g-prior is a Bayesian extension of Zellner's classical g-prior, adapting the normal linear model framework to exponential family, non-Gaussian, and additive models.
• It utilizes Laplace approximations and flexible hyper-prior structures, such as hyper-g, robust, and intrinsic priors, to enhance model selection and estimation in complex scenarios.
• The framework supports practical applications in spline-based function estimation and block-specific shrinkage, ensuring both theoretical robustness and computational tractability.

A generalized g-prior is an extension of Zellner's classical g-prior framework for Bayesian modeling, most notably applied to variable selection and nonparametric function estimation in regression and generalized additive models (GAMs). The construction generalizes from normal linear models to exponential family models and expands the g parameter to be either block-specific, drawn hierarchically, or mixed according to flexible hyper-prior distributions. Modern methodologies encompass continuous mixtures such as the hyper-g, robust, intrinsic, and Beta-prime families, as well as adaptations for penalized splines and non-Gaussian likelihoods, offering both theoretical robustness and practical tractability (Kang et al., 2023, Li et al., 2015, Bové et al., 2011, Fouskakis et al., 2020).

1. Classical g-Prior and Core Generalizations

Zellner's classical g-prior in an $n \times p$ Gaussian linear regression model, $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$, assigns

$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$

where $g > 0$ is a scalar controlling prior information. For $g=n$, the prior matches the information of one sampled data point per parameter (the "unit information prior"). This g-prior yields analytic marginal likelihoods, but its sensitivity to $g$ limits inferential stability and consistency (Li et al., 2015, Kang et al., 2023).

Extensions include hierarchical models with a prior on $g$ ("mixtures of g-priors"), empirical Bayes choices, and block-specific (localized) versions. Block hyper-g priors generalize to

$$\beta_b \mid \sigma^2, g_b \sim N_{p_b}(0, g_b \sigma^2 (X_b^T X_b)^{-1})$$

with independent priors on the block scales $g_b$ (Som et al., 2014).

2. Generalized g-Priors for Generalized Linear and Additive Models

Generalized g-priors adapt the Gaussian construction to non-Gaussian exponential family models, including GLMs and GAMs. Given a canonical link and a model

$$p(y_i;\theta_i,\phi) = \exp\left(\frac{y_i\theta_i - b(\theta_i)}{\phi} + c(y_i,\phi)\right),$$

with an additive predictor $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$0, the approach replaces $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$1 with the (centered) observed Fisher information $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$2 evaluated at the MLE to account for non-normality.

For basis expansion of $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$3 in splines, the generalized g-prior for the $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$4-vector of basis coefficients $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$5 is: $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$6 where $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$7 is the design matrix for the chosen spline basis and knot configuration $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$8 (Kang et al., 2023).

To maintain analytic tractability, Laplace approximations replace exact marginalizations. Inferences, Bayes factors, and credible intervals are then computed using closed-form approximations, with explicit handling of the curvature of the non-Gaussian likelihood (Bové et al., 2011, Li et al., 2015).

3. Mixtures of g-Priors and the tCCH Family

Rather than fixing $$Y \mid X, \beta, \sigma^2 \sim N_n(X\beta, \sigma^2 I_n)$$9, it is endowed with a flexible prior over $$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$0. The truncated Compound Confluent Hypergeometric (tCCH) family for $$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$1 provides a unifying framework: $$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$2 Key special cases describe many prominent g-prior mixtures:

Name	a	b	r	s	ν	κ	Order of $$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$3
Unit info (fixed)	—	—	—	—	—	—	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$4
Hyper-g	1	2	0	0	1	1	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$5
Hyper-g/n	1	2	3/2	0	1	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$6	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$7
Robust	1	2	3/2	0	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$8	1	$$\beta \mid \sigma^2, g \sim N_p(0, g\sigma^2 (X^T X)^{-1})$$9
Intrinsic	1	1	1	0	$g > 0$0	$g > 0$1	$g > 0$2

The marginal model likelihood (Laplace-approximated) under a tCCH-mixed g-prior is: $g > 0$3 where $g > 0$4 is a Wald-type statistic for the fitted model. This yields explicit Bayes factors and supports model search via enumeration or MCMC (Kang et al., 2023, Li et al., 2015).

4. Computational Methodology: Spline Selection and Marginal Likelihoods

Model space consists of knot configurations and smoothness choices for each functional term. The prior over knot configurations $g > 0$5 is constructed so $g > 0$6 is full-rank, with templates:

• Even-knot splines: a fixed grid of knots determined by quantiles or equally spaced locations.
• VS-knot splines (variable selection): Bernoulli prior over inclusion of candidate knots with constraints on maximum number.
• Free-knot splines: fully free knot locations sampled via reversible-jump MCMC.

The marginal likelihood for each configuration is computed using the Laplace/tCCH marginal formula, and the posterior over model configurations is explored by enumeration or MCMC. Sampling from the conditional Laplace-approximated posterior of $g > 0$7 is performed at each selected knot configuration, allowing for full Bayesian model averaging including uncertainty in smoothness and knot placement. The prior on knot count per term is best chosen as a mixture of a point mass at zero with a lightly-tail truncated geometric (Kang et al., 2023).

5. Evaluation, Empirical Comparisons, and Practical Recommendations

Simulation studies in Bernoulli, Poisson, and Gaussian GAMs indicate that fixed g-priors such as the unit information prior ($g > 0$8) systematically underfit due to excessive penalization. Among mixture g-priors, "intrinsic" and "robust" settings achieve superior root mean square error (RMSE) and credible-interval coverage, consistently across sample sizes and model types (Kang et al., 2023).

• "Hyper-g" and "hyper-g/n" priors show a tendency to overfit, often selecting overly complex models.
• Beta-prime and ZS-adapted versions tend to underfit, albeit moderately.
• Even-knot splines are computationally most efficient but lack local adaptivity; variable-selection (VS) knot models are preferred when adaptivity is required.

Recommended defaults for practitioners are the "intrinsic" or "robust" mixtures of g-priors for the scale hyperparameters, coupled with a VS-knot spline strategy for flexible, locally adaptive smoothing. Hyperparameters for the knot count prior are set as a mixture of a point mass at 0 and a truncated geometric with light-tailed decay (e.g., mixture weight ≈ 0.2, geometric parameter ≈ 0.5) (Kang et al., 2023).

6. Connections to Broader Generalized g-Prior Frameworks

Power-Secondary g-priors (PEP) and related "expected-posterior priors" also generalize the g-prior through randomized or power-discounted imaginary data, leading to a family of adaptive, compatibility-enforcing priors which can be framed as mixtures of g-priors with beta-prime mixing on $g > 0$9 (Fouskakis et al., 2020, Fouskakis et al., 2013). These produce analytically tractable prior-posterior updates and Bayes factors mimicking BIC asymptotics, while enforcing stronger parsimony in finite samples compared to both fixed-g and hyper-g settings.

In high-dimensional settings, generalized g-priors retain consistency under suitable growth conditions on $g=n$0 and the dimension of $g=n$1; mixture priors (including hyper-g and Zellner-Siow) have been shown to achieve posterior consistency in regimes where $g=n$2, subject to suitable tail conditions and measurements in $g=n$3 norm (Sparks et al., 2015).

Block hyper-g priors introduce separate shrinkage for groups of coefficients, circumventing issues such as Essentially Least Squares and Conditional Lindley's Paradox that arise when a single g parameter is globally assigned, ensuring robust model selection in the presence of block-sparse or highly heterogeneous signal regimes (Som et al., 2014).

7. Theoretical Properties and Paradigms

Generalized g-priors, in all forms, support analytic or Laplace-computable marginal likelihoods, facilitate rigorous Bayesian model selection and model averaging, and adapt posterior shrinkage to both model dimension and observed data fit. These priors achieve local geometric invariance, curvature matching to the likelihood, and support for both high-dimensional and nonparametric regimes.

The tCCH family subsumes all standard mixtures; the full Laplace-approximation pipeline for spline-based inference is robust across exponential family models, and recommendations for hyperparameters are now standardized based on a blend of theoretical desiderata and empirical performance (Kang et al., 2023, Li et al., 2015). Practical implementations are available in several R packages (e.g., BAS, hypergsplines) (Li et al., 2015, Bové et al., 2011).