A preregistered simulation study provided evidence on the appropriate use of data-driven variable selection in...

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Multivariable regression modeling is crucial in epidemiology.
• Data-driven variable selection is common but can lead to biased, unstable models.
• Limited guidance exists for selecting appropriate methods in descriptive modeling.

Purpose of the Study:

• Provide evidence-based guidance on data-driven variable selection in epidemiological regression.
• Focus on appropriate use within descriptive modeling contexts.
• Compare common variable selection methods systematically.

Main Methods:

• Comprehensive simulation study comparing backward elimination (BE) and Lasso variants.
• Evaluated across realistic epidemiological scenarios (outcome type, sample size, signal-to-noise ratio).
• Assessed performance using true positive rates (TPR), false positive rates (FPR), and coefficient bias.

Main Results:

• No single method excels in all scenarios; large samples or strong signals are needed.
• BE with mild criteria (α=0.5) maximizes TPR.
• BE with strict criteria (α=0.05, BIC) minimizes FPR; AIC balances both.
• Lasso methods underperformed on descriptive modeling criteria.

Conclusions:

• Variable selection can supplement, but not replace, knowledge-driven modeling.
• Use is restricted to favorable conditions (large N, high signal-to-noise).
• Method choice depends on desired TPR/FPR trade-off, with a decision framework provided.