The typical method for estimating mixture of regression models relies on the assumption that error components are normally distributed. This assumption makes them highly vulnerable to outliers or data with heavy-tailed errors. This lecture will review some robust alternatives for mixture of regression models. In particular, we will focus on a new robust model introduced by Zarei, which extends the mixture of symmetric α-stable (SαS) distributions to the regression setting.
The SαS distribution is a heavy-tailed generalization of the normal distribution, where an additional parameter, α, controls the heaviness of the tails. A unique characteristic of the SαS distribution is that its variance diverges to infinity when α < 2. This property makes the model exceptionally robust against extreme outliers compared to other heavy-tailed distributions, like the Student's t-distribution.
The model's parameters, except for α, are estimated using a standard Expectation-Maximization (EM) algorithm. The parameter α is estimated separately via a stochastic EM algorithm that utilizes a rejection sampling method. We will illustrate and compare this new model with existing mixture regression models using both simulated and real-world datasets.
23 Settembre 2025, ore 12
Shaho Zarei
University of Kurdistan (UOK)
In person: Room 24 (4th floor) building CU002 Scienze Statistiche
Webinar: https://uniroma1.zoom.us/j/83625004899?pwd=bXCtz0mp759PUh2lkqT0BUoVa0Uegg.1
ID riunione: 836 2500 4899
Passcode: 123456