WEBINAR: A Two-Stage Joint Selection and Compositional Regression Model for Simplex Data with Structural Zeros (Kai-Sheng Song)

25/05/2026

CoDa Association

WEBINAR

 

A Two-Stage Joint Selection and Compositional Regression Model for Simplex Data with Structural Zeros  

Kai-Sheng Song

 

Day: 29-05-2026

Time: 14:30h (CEST: UTC/GMT +2 h)

Link: https://us06web.zoom.us/j/89178651916?pwd=kyiYYqhviTWBAfZ3MzbpaQjfpI3vu8.1

 

 

 

Speaker: Kai-Sheng Song 

Department of Data Analytics and Statistics. Department of Mathematics. University of North Texas

 

Moderator: Anna Maria Fiori

 

Title: A Two-Stage Joint Selection and Compositional Regression Model for Simplex Data with Structural Zeros

 

Abstract: Structural zeros are ubiquitous across empirical fields, appearing when firms select among distinct debt instruments in corporate finance, consumers choose between competing brands in marketing, or agents make discrete trade-offs throughout the social sciences. Rather than reflecting missing data or measurement errors, these zeros represent deliberate, discrete economic choices made by optimizing agents. Standard compositional data techniques typically break down or require ad-hoc adjustments when applied to simplex boundaries where structural zeros occur.

 
In this talk, a novel, unified econometrics framework is proposed that defines a single, coherent joint probability distribution over the entire mixed continuous-discrete space. By bridging Random Utility Models for discrete selection with continuous sub-model simplex densities for conditional allocation, the methodology provides a rigorous foundation for multivariate mixed regression data. Within this joint architecture, the Extended Flexible Dirichlet distribution is introduced as a concrete parametric instance capable of accommodating complex covariance structures and multi-modal boundary behaviors. To make estimation practically feasible, a computationally efficient two-stage profile likelihood routine is derived to factorize the joint likelihood, separating discrete choice configurations from continuous regressions without sacrificing structural or geometric closure properties.
 
Extensive Monte Carlo simulation studies confirm the finite-sample efficiency and consistency of the proposed estimators. An application to corporate debt structures reveals new insights and showcases the importance of recognizing joint compositional features in financial and economic studies.

 

Keywords:  structural zeros, simplex data, compositional regression, Random Utility Models, Extended Flexible Dirichlet distribution, profile likelihood, mixed continuous-discrete data

 
==========
Please note that the webinar will be recorded. Feel free to forward this message to your colleagues who may be interested in attending.