Tweedie GLM: Theory and Applications

• Tweedie GLM is a statistical framework within the exponential dispersion family that uses a power-variance relationship to model data with exact zeros and right-skewed positive values.
• The model extends into DGLM by incorporating a dual regression framework that estimates both the mean and dispersion, allowing covariate-dependent variability.
• Recent advances focus on integrating regularization, spatial effects, and Bayesian methods to enhance model performance in fields like insurance and environmental science.

The Tweedie Generalized Linear Model (GLM) is a statistical modeling framework within the exponential dispersion (ED) family, distinguished by its power-variance structure and its facility for simultaneously modeling both the mean and dispersion of a response variable—most notably through the double generalized linear model (DGLM) extension. Widely adopted in actuarial science, spatial statistics, and other domains dealing with semicontinuous or mixed response data, Tweedie GLMs are especially prominent where data exhibit both exact zeros and positive, right-skewed values, as in compound Poisson–gamma distributions. Recent advances focus on regularization, spatial random effects, and algorithmic scalability, as well as Bayesian variable selection and zero-inflation modeling.

1. Exponential Dispersion Models and the Tweedie Family

Exponential dispersion models for a univariate outcome $y$ are defined by the canonical density

$$f(y;\theta,\phi) = a(y,\phi) \exp\left\{\phi^{-1}\left[y\theta - \kappa(\theta)\right]\right\},$$

where $\theta$ is the canonical parameter, $\phi > 0$ is the dispersion parameter, and $\kappa(\theta)$ is the cumulant function. The mean and variance are given by

$$\mu = \mathbb{E}[y] = \kappa'(\theta), \quad \mathrm{Var}[y] = \phi \kappa''(\theta).$$

The Tweedie subfamily is uniquely characterized by a power-variance relationship: $$V(\mu) = \mu^p, \quad p \in \mathbb{R} \setminus \{0, 1\},$$ encompassing normal ($p=0$), Poisson ($p=1$), gamma ($p=2$), inverse Gaussian ($p=3$), and the compound Poisson–gamma ($1 < p < 2$) cases. For $1 < p < 2$, the Tweedie model exhibits an atom at zero and continuous positive support. This is essential for modeling insurance claims, rainfall, and similar data exhibiting spikes at zero together with continuous positive outcomes (Halder et al., 2020, Halder et al., 2019).

2. Tweedie GLM and Double Generalized Linear Models (DGLMs)

The Tweedie GLM assumes $N$ independent responses $y_i \sim \mathrm{Tw}(\mu_i, \phi, p)$, with a link function $g$ such that

$$g(\mu_i) = \mathbf{x}_i^\top \beta, \qquad \mathrm{Var}[y_i] = \phi \mu_i^p.$$

The canonical link is $g(\mu) = \mu^{1-p}/(1-p)$ (power link), with the log link being common in practical applications.

The DGLM extends this by allowing dispersion to vary with its own set of covariates, introducing an additional regression: $$g_1(\mu_i) = \mathbf{x}_i^\top \beta, \qquad \mathrm{Var}[y_i] = \phi_i V(\mu_i), \qquad g_2(\phi_i) = \mathbf{z}_i^\top \gamma,$$ where $g_2$ is typically the log-link, so that $\phi_i = \exp(\mathbf{z}_i^\top \gamma)$ (Halder et al., 2020, Halder et al., 2023, Halder et al., 2019). This dual-regression structure provides flexibility to account for heterogeneous overdispersion related to observed covariate effects.

3. Spatial Effects, Graph-Laplacian Penalty, and Residual Analysis

For spatially referenced data, a region-specific random effect $\alpha$ is introduced to the mean predictor: $$g_1(\mu_{ij}) = \mathbf{x}_{ij}^\top \beta + r_{ij}^\top \alpha,$$ where $r_{ij}$ indicates the areal unit for observation $(i,j)$. To induce spatial smoothness, a quadratic penalty over an adjacency graph $G=(V,E)$ is imposed via the graph Laplacian $\mathfrak{W} = \mathcal{D} - \mathcal{W}$,

$$P(\alpha) = \frac{1}{2}\alpha^\top \lambda_2 \mathfrak{W} \alpha = \sum_{(i \sim i') \in E} \lambda_2 (\alpha_i - \alpha_{i'})^2.$$

The penalized objective for all parameters $\Theta = (\beta, \alpha, \gamma)^\top$ is then

$$F(\Theta, p) = \ell(\Theta, p) + \frac{1}{2} (A\Theta)^\top[\lambda_1 I_0 + \lambda_2 W_0](A\Theta),$$

with $A$ indexing penalized components (Halder et al., 2020, Halder et al., 2019).

Residual analysis, both deviance and Pearson residuals,

$$r_{ij}^{(D)} = \mathrm{sign}(y_{ij}-\hat\mu_{ij}) \sqrt{\frac{d(y_{ij},\hat\mu_{ij})}{\hat\phi_{ij}}}, \quad r_{ij}^{(P)} = \frac{y_{ij}-\hat\mu_{ij}}{\sqrt{\hat\phi_{ij}\hat\mu_{ij}^p}},$$

permits spatial risk ranking and the identification of anomalous spatial units (Halder et al., 2019).

4. Estimation Algorithms: Penalized Maximum Likelihood, Block Coordinate Descent, and Bayesian Methods

The penalized likelihood is minimized via a blockwise coordinate descent algorithm. Given fixed $p$, alternate quadratic majorizations are constructed for the mean-side parameters $\eta = (\beta, \alpha)$ and dispersion-side $\gamma$, with updates: $$\eta^{\text{new}} = [\lambda_1I_0+\lambda_2W_0 + c_1\,\nabla^2_{\eta\eta}\ell(\eta)]^{-1} (c_1\,\nabla^2_{\eta\eta}\ell(\eta) - \nabla_\eta\ell(\eta)),$$

$$\gamma^{\text{new}} = -\frac1{c_2}[\nabla^2_{\gamma\gamma}\ell(\gamma)]^{-1}\nabla_\gamma\ell(\gamma).$$

The Tweedie index $p$ is profiled over a grid. Convergence is certified by a monotonic decrease in $F(\Theta,p)$, with an explicit bound on the descent at each iteration (Halder et al., 2020).

Bayesian estimation can be conducted under a fully hierarchical framework, incorporating spike-and-slab priors for variable selection in both mean and dispersion predictors, and Gaussian process priors for spatial random effects. Posterior inference uses a hybrid of Gibbs, Metropolis–Hastings, and Metropolis-adjusted Langevin algorithms (MALA), exploiting blockwise updates for efficiency (Halder et al., 2023). Choice of prior on the Tweedie index (typically Uniform$(1,2)$ for compound Poisson–gamma) and convergence diagnostics are standard components.

5. Extensions: Zero-inflated Tweedie Models and Nonparametric Estimation

For data with excessive zeros, the zero-inflated Tweedie (ZIT) model augments the standard Tweedie GLM with a logistic mixture: $$Y_i = \begin{cases} 0, & \text{with probability } \pi_i \ T_i \sim \mathrm{Tweedie}(\mu_i, \phi_i, p), & \text{with probability } 1-\pi_i \end{cases}$$ with logit link for $\pi_i$ and additional GLMs for both $\mu_i$ and $\phi_i$ (Gu, 2024).

Parameters in both standard and zero-inflated models may be estimated via generalized EM algorithms, with mean, dispersion, and zero-inflation probabilities all permitted to depend on covariates, fit via gradient-boosted trees. The estimation alternates imputation of latent zero states, and updates of component predictors via boosted minimizing of their respective losses, leveraging decision tree models for flexibility (Gu, 2024).

Gradient boosting approaches such as TDboost enable highly flexible, non-linear, and interaction-rich modeling of the mean structure in Tweedie GLMs, with profile-likelihood estimation for the index parameter $p$ (Yang et al., 2015). Orthogonality properties facilitate the decoupling of mean and dispersion estimation, rendering profile approaches consistent and computationally stable.

6. Applications and Model Selection

Tweedie GLMs, DGLMs, and their extensions (spatial, zero-inflated, boosting) have found primary application in insurance (premium and loss modeling for auto/casualty lines), environmental sciences, and other domains with semicontinuous outcomes. In empirical studies, penalized and spatially-augmented Tweedie DGLMs outperform ridge and unpenalized models in both simulation and real auto insurance applications (Halder et al., 2020, Halder et al., 2019).

Variable selection in Bayesian DGLMs leverages spike-and-slab priors and local false discovery rates for covariate inclusion, with the R package sptwdglm providing practical implementation (Halder et al., 2023). For nonparametric and interaction modeling, gradient-boosted Tweedie models (TDboost, ZITboost) have been shown to outperform parametric alternatives in prediction, particularly on highly zero-inflated insurance claims data (Yang et al., 2015, Gu, 2024).

7. Theoretical Properties and Scalability

The Fisher-orthogonality of exponential dispersion models between mean, dispersion, and index parameters ensures that mild misspecification of $p$ does not adversely impact estimation of spatial or covariate effects in large samples (Halder et al., 2019). Scalability for massive spatial domains is achieved by block-diagonalizing the spatial Laplacian, reducing computational cost while preserving the convergence and accuracy of the estimation algorithm.

Key theoretical properties—power-variance structure, flexibility in modeling overdispersion, support for spatial and zero-inflation extensions, algorithmic guarantees—place the Tweedie GLM with DGLM extension as a central methodology for contemporary semicontinuous and spatial data analysis (Halder et al., 2020, Halder et al., 2019, Gu, 2024).