Sensitivity of test for overdispersion in Poisson regression

Area of Science:

• Statistics
• Biostatistics
• Data Analysis

Background:

• Count data frequently exhibit overdispersion, where variability exceeds the mean, violating Poisson assumptions.
• This overdispersion leads to biased parameter estimates in standard Poisson regression.
• Dean's score tests (1992) are common for detecting overdispersion but can be affected by extreme observations.

Purpose of the Study:

• To propose novel diagnostic measures for evaluating the sensitivity of Dean's score test.
• To assess how anomalous observations impact the detection of overdispersion in Poisson regression.
• To enhance the reliability of statistical analyses for overdispersed count data.

Main Methods:

• Development of diagnostic statistics to quantify the influence of individual observations.
• Application of these diagnostics to Dean's score test for overdispersion in Poisson regression.
• Illustrative analysis using real-world datasets, including fabric faults and Ames salmonella assays.

Main Results:

• The proposed diagnostics effectively identify observations that unduly influence Dean's score test results.
• Sensitivity analysis reveals specific conditions under which the test is vulnerable to extreme values.
• Demonstrated utility in correctly interpreting overdispersion in practical count data scenarios.

Conclusions:

• The developed diagnostics are crucial for robustly applying Dean's score test to overdispersed count data.
• These measures aid researchers in identifying and handling influential points, ensuring more reliable statistical inference.
• The findings contribute to more accurate modeling and analysis of biological and industrial count data exhibiting extra-Poisson variation.