On the role of marginal confounder prevalence - implications for the high-dimensional propensity score algorithm

Area of Science:

• Pharmacoepidemiology
• Statistical methodology

Background:

• High-dimensional propensity score (HDPS) algorithms are increasingly used in pharmacoepidemiology for analyzing large administrative databases.
• HDPS algorithms use covariate marginal prevalence as an indicator of potential confounding impact.

Purpose of the Study:

• To investigate the influence of confounder marginal prevalence on bias magnitude when estimating risk ratios in point exposure studies with binary outcomes.
• To evaluate the role of covariate prevalence in the variable selection process of HDPS algorithms.

Main Methods:

• Applied the law of total probability and an established bias formula.
• Derived and illustrated relative bias boundaries concerning marginal confounder prevalence.

Main Results:

• Maximum bias magnitudes can occur at any marginal prevalence level of a binary confounder.
• Both low and high prevalence confounders can exert substantial confounding effects on estimated risk ratios, especially with rare or common exposures.

Conclusions:

• Prevalence-based covariate pre-selection may result in suboptimal confounder sampling within HDPS algorithms.
• Recommends omitting prevalence-based empirical identification of candidate covariates in HDPS, despite the algorithm's benefits in large-scale studies.