Statistical Power for Postlicensure Medical Product Safety Data Mining

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Longitudinal observational databases are crucial for public health surveillance.
• Tree-based scan statistics offer a method for data mining in large-scale health datasets.
• Adjusting for multiple testing is essential when analyzing numerous disease outcomes.

Purpose of the Study:

• To provide sample size calculations for implementing tree-based scan statistics in longitudinal observational databases.
• To evaluate the statistical power of both unconditional and conditional Poisson tree-based scan statistics.
• To guide researchers in designing effective data mining analyses for pharmacovigilance.

Main Methods:

• Utilized tree-based scan statistics on epidemiologic datasets with hierarchical outcome structures.
• Assessed statistical power by varying excess risk, sample size, event rates, and healthcare utilization.
• Quantified power reduction due to inappropriate risk window specifications.

Main Results:

• Achieved at least 98% power to detect an excess risk of 1 event per 10,000 exposed individuals with a sample size of 500,000.
• The conditional tree-based scan statistic effectively controlled Type I error in the presence of confounding healthcare utilization, unlike the unconditional version.
• Demonstrated the impact of risk window size on statistical power.

Conclusions:

• Tree-based scan statistics enhance pharmacovigilance by enabling monitoring of numerous outcomes with controlled multiple testing.
• Power evaluations are critical for optimizing the design and implementation of retrospective data mining studies.
• These statistical tools support robust surveillance in large longitudinal health databases.