Semiparametric Analysis of Heterogeneous Data Using Varying-Scale Generalized Linear Models.

Area of Science:

• Statistics
• Econometrics
• Biostatistics

Background:

• Generalized linear models (GLMs) often assume homoscedasticity, which is frequently violated in real-world data.
• Heterogeneity in data, arising from varying exposures or aggregation levels, necessitates advanced modeling techniques.
• Existing models may not adequately capture complex relationships when variance depends on covariates.

Purpose of the Study:

• To develop a class of heteroscedastic generalized linear regression models allowing nonparametric rescaling of regression parameters.
• To establish efficient semiparametric inference methods for the parametric components of these models.
• To provide tools for adapting statistical analyses to data with heterogeneity.

Main Methods:

• Introduction of a flexible class of models accommodating nonparametric scale function estimation.
• Development of an algorithm for nonparametric estimation of the scale function.
• Derivation of asymptotic distribution theory for semiparametric regression parameter estimates.
• Utilizing bootstrap methods for goodness-of-scale testing.

Main Results:

• The proposed semiparametric estimator for the parametric part achieves the semiparametric efficiency bound.
• The methodology is validated through simulations and real-world datasets.
• The developed algorithm effectively estimates the scale function nonparametrically.

Conclusions:

• The introduced heteroscedastic models offer a powerful framework for analyzing data with complex variance structures.
• Efficient semiparametric inference is achieved, improving upon standard GLM approaches.
• The methodology demonstrates practical utility in diverse applications, including ultrasound safety research.