New strategies for detecting atypical observations based on the information matrix equality

Area of Science:

• Statistics
• Statistical Modeling
• Regression Analysis

Background:

• Traditional diagnostic analyses in regression modeling rely on residuals or local influence measures to identify atypical observations.
• Existing methods may not fully capture the impact of unusual data points on the overall model specification, especially under maximum likelihood estimation.

Purpose of the Study:

• To develop a new approach for identifying atypical observations in regression models estimated by maximum likelihood.
• To define atypical observations based on their disproportionate effect on model specification adequacy.
• To introduce new diagnostic measures derived from the information matrix equality.

Main Methods:

• The proposed approach leverages the information matrix equality, which holds under correct model specification.
• A new definition of atypical observations is introduced, focusing on their impact on the degree of model specification adequacy.
• Various distance measures between symmetric matrices are employed to quantify model adequacy, alongside a modified generalized Cook distance and a new criterion combining modified and unmodified generalized Cook's distances.

Main Results:

• The study presents a novel framework for identifying atypical observations that significantly impact model adequacy.
• Empirical applications demonstrate the utility of the proposed methods in Gaussian and beta regression models.
• The new diagnostic criteria effectively highlight cases that disproportionately affect the assessment of model specification.

Conclusions:

• The developed approach offers a valuable addition to diagnostic tools in regression analysis, particularly for maximum likelihood estimation.
• The focus on information matrix equality provides a robust way to assess model specification and detect influential atypical observations.
• The proposed methods enhance the reliability of regression diagnostics by offering a more sensitive way to identify problematic data points.