Mixtures of multivariate linear asymmetric Laplace regressions with multiple asymmetric Laplace covariates

TL;DR

This paper introduces a mixture model based on multivariate asymmetric Laplace distributions for clustering responses and covariates with skewness, including a robust extension for outlier detection, with implementation in an R package.

Contribution

The paper proposes a novel mixture model framework for skewed data, incorporating outlier detection and leverage point identification, with theoretical and empirical validation.

Findings

Enhanced outlier detection capabilities.

Effective clustering of skewed data.

Implementation available in R package.

Abstract

In response to the challenge of accommodating non-Gaussian behaviour in data, the shifted asymmetric Laplace (SAL) cluster-weighted model (SALCWM) is introduced as a model-based method for jointly clustering responses and random covariates that exhibit skewness. Within each cluster, the multivariate SAL distribution is assumed for both the covariates and the responses given the covariates. To mitigate the effect of possible atypical observations, a heavy-tailed extension, the contaminated SALCWM (cSALCWM), is also proposed. In addition to the SALCWM parameters, each mixture component has a parameter controlling the proportion of outliers, one controlling the proportion of leverage points, one specifying the degree of outlierness, and another specifying the degree of leverage. The cSALCWM has the added benefit that once the model parameters are estimated and the observations are assigned to components, a more refined intra-group classification in typical points, (mild) outliers, good leverage, and bad leverage points can be directly obtained. An expectation-conditional maximization algorithm is developed for efficient maximum likelihood parameter estimation under this framework. Theoretical identifiability conditions are established, and empirical results from simulation studies and validation via real-world applications demonstrate that the cSALCWM not only preserves the modelling strengths of the SALCWM but also significantly enhances outlier detection and overall inference reliability. The methodology proposed in this paper has been implemented in an \texttt{R} package, which is publicly available at https://github.com/arnootto/ALCWM.