Distributional Regression Models and Applications

Distributional regression represents a paradigm shift from conventional mean-based modelling to a comprehensive characterisation of the entire conditional distribution of a response variable. By specifying separate regression structures for parameters such as location, scale and shape, these models accommodate heteroscedasticity, skewness and kurtosis, enabling nuanced probabilistic forecasts. Methodological innovations include quantile regression, generalised additive models for location, scale and shape (GAMLSS) and their Bayesian counterparts (BAMLSS), as well as non-Gaussian extensions using flexible families such as sinh-arcsinh distributions. Advances in computation—particularly Markov chain Monte Carlo, backfitting algorithms and gradient boosting—have rendered complex distributional models tractable on large data sets. Applications span precision livestock monitoring, clinical normative modelling, forestry and climate science, where full predictive distributions inform risk assessment, resource management and personalised decision-making. The global reach of these methods is evident in domains as diverse as neuroimaging, ecological mapping and industrial forecasting, underlining their capacity to handle irregular measurement regimes, spatial structure and multi-site variability with rigorous uncertainty quantification.

Research from Nature Portfolio

Recent studies have introduced a three-step framework for irregular intensive longitudinal data that combines normalising transformations to address nonlinearity, broken-stick alignment to regularise observation times and strategic input selection for distributional forecasting. Illustrated in a piglet weight-monitoring context, this approach captures individual growth trajectories, accommodates varying measurement histories and yields accurate full-distribution predictions. The methodology is readily adaptable to diverse applications requiring probabilistic future outcomes under nonlinear dynamics and irregular sampling.

Research from all publishers

In machine learning, a holistic review of predictive uncertainty estimation has synthesised early statistical methods with modern algorithms—ranging from random forests to deep neural networks—under unified proper scoring rules and consistent scoring functions. This work highlights how distributional frameworks can guide the design of novel probabilistic algorithms and standardised evaluation metrics. In clinical neuroimaging, hierarchical Bayesian regression has been extended with the flexible sinh-arcsinh family to model heteroscedastic non-Gaussian data, demonstrating improved normative modelling across multiple sites and enhanced control over distributional shape parameters. Ecological applications have seen Bayesian spatial distributional regression applied to airborne laser scanning data to predict stem diameter distributions in alpine forests: by modelling both gamma distribution parameters with nonlinear smooths and Gaussian-process spatial effects, the method delivers robust spatial interval predictions that inform forest management and conservation strategies.

Distributional Regression Models and Applications publication trend

The graph below shows the total number of articles in distributional regression models and applications across all publications each year (not limited to Nature Index journals).

Technical terms

Distributional regression model: A regression framework that models multiple parameters (e.g. location, scale, shape) of the full conditional distribution of a response variable rather than only its mean.

Quantile regression: An approach that directly models specified conditional quantiles of a response variable, capturing distributional heterogeneity across covariate values.

Heteroscedasticity: A condition in which the variability of a response variable changes systematically with covariates rather than remaining constant.

Hierarchical Bayesian regression: A multilevel Bayesian modelling approach that accounts for nested or grouped data structures by introducing parameters at different hierarchy levels.

Proper scoring rule: A metric for assessing the quality of probabilistic forecasts, ensuring that the best expected score is achieved by reporting the true distribution.

References

1. A review of predictive uncertainty estimation with machine learning. Artificial Intelligence Review (2024).
2. Broken Stick Model for Irregular Longitudinal Data. Journal of Statistical Software (2023).
3. Longitudinal individual predictions from irregular repeated measurements data. Scientific Reports (2023).
4. Non-Gaussian normative modelling with hierarchical Bayesian regression. Imaging Neuroscience (2024).
5. Spatial Prediction of Diameter Distributions for the Alpine Protection Forests in Ebensee, Austria, Using ALS/PLS and Spatial Distributional Regression Models. Remote Sensing (2024).
6. Generalized Additive Models for Location Scale and Shape (GAMLSS) in R. Journal of Statistical Software (2007).
7. BAMLSS: Bayesian Additive Models for Location, Scale, and Shape (and Beyond). Journal of Computational and Graphical Statistics (2018).