Performance Evaluation of Parametric and Nonparametric Methods When Assessing Effect Measure Modification.

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Effect measure modification is crucial for understanding exposure impacts.
• Parametric models offer efficiency but rely on strong assumptions.
• Nonparametric models avoid functional form assumptions but may need larger samples.

Purpose of the Study:

• To compare the performance of parametric and nonparametric models in detecting effect modification.
• To evaluate tradeoffs between model accuracy and sample size requirements.
• To assess models for both binary and continuous modifiers.

Main Methods:

• Simulation study evaluating generalized linear models and doubly robust (DR) estimators.
• Models assessed with and without sample splitting.
• Continuous modifiers modeled using cubic splines, fractional polynomials, and a nonparametric DR-learner.

Main Results:

• Generalized linear models demonstrated highest power for binary modifiers (0.42-1.00).
• Augmented inverse probability weighting showed lowest power, improving by 23% with sample splitting.
• For continuous modifiers, DR-learner matched parametric models for monotonic functions but had lower integrated bias for non-monotonic functions.

Conclusions:

• Parametric and nonparametric models exhibit comparable performance in evaluating effect modification.
• Model choice depends on specific data characteristics and modifier types.
• Findings support the use of flexible modeling strategies for effect modification analysis.