The L^p -quantile regression framework, defined through an L^p -type asymmetric loss with 1
In survival analysis research, the effects of covariates on the time to the event of interest often exhibit significant nonlinear characteristics. To address this issue, this paper constructs an AFTBoost model by integrating the parametric accelerated failure time (AFT) model with the XGBoost algorithm. Considering the limitations of machine learning algorithms in small-sample settings, this paper further introduces a transfer learning method. By incorporating external source-domain data, a dynamic weighted transfer learning AFTBoost algorithm is proposed, aiming to improve the robustness and predictive accuracy of the model on the target domain. Numerical simulation results show that, compared with the basic AFTBoost model, the proposed method achieves higher predictive accuracy and stronger robustness under various settings. Finally, the proposed method is applied to an empirical study of Alzheimer’s disease conversion, and the results verify that introducing external data can effectively improve the predictive performance of the model.
This paper proposes a composite quantile regression estimation method for a multivariate functional partially varying coefficient hybrid model (MFPVCHM). The model accommodates multiple correlated functional predictors defined on distinct domains with different dimensions, scalar covariates with smooth varying coefficients, and errors that follow short-range dependent and strictly stationary linear processes. By combining B-spline approximation of the varying coefficient functions with multivariate functional principal component analysis for dimension reduction, the proposed method provides robust estimation and inference under heavy-tailed and serially dependent errors. We derive the convergence rates of the functional and varying-coefficient estimators under two simultaneously diverging dimensions, and establish their asymptotic normality through a Bahadur-type representation. Monte Carlo simulations show that the proposed estimator achieves competitive efficiency under normal errors and superior robustness under non-normal errors, while an application to U.S. mortality and air pollution data illustrates its practical utility.
Many tensor data have smooth aspects in one or more modes and often contain outliers beyond the noise level. So, considering both smoothness and robustness in the tensor decomposition helps identify continuously varying factors of the tensor data that are less susceptible to contamination. In this paper, we present a novel rank-one approximation approach that integrates these two perspectives for CP decomposition models. The proposed formulation adopts a robust Huber loss and a smooth penalization by filtering in each tensor mode to ensure robustness and smoothness, respectively. We propose a practical algorithm based on the iterative weighted least squares method. We further increase the flexibility of the proposed approach by imposing basis constraints on individual factors of the tensor. Through simulation studies and real-data analysis of Korean sea weather data, we demonstrate the empirical promise of our proposed method over existing non-robust or non-smooth decomposition methods.
Structural Equation Modeling (SEM) is a widely used framework for analyzing relationships among observed and latent variables, with the estimation process representing a critical step. A central component of estimation is the computation of the implied covariance matrix. A key observation motivating this work is an implicit structural constraint arising from the equality between the empirical variances of the manifest variables and the diagonal of the implied covariance matrix. This constraint is not explicitly considered in the classical CB-SEM framework, which relies on Jöreskog’s formulation. An alternative approach based on the Finite Iterative Method (FIM) is proposed for constructing the implied covariance matrix. The proposed method naturally incorporates the variance constraint within its algorithmic structure, resulting in a substantial reduction in the number of free parameters by nearly half, thereby yielding a more parsimonious representation and reducing computational burden. The performance of the FIM-based approach is evaluated through empirical analysis and a simulation study, demonstrating its numerical stability, parsimony, and equivalence with classical SEM estimation.
