Categorisation of continuous covariates for stratified randomisation: How should we adjust?

Area of Science:

• Biostatistics
• Clinical Trials Methodology

Background:

• Stratified randomization is used in clinical trials to ensure balance in prognostic factors.
• Continuous variables are often categorized for stratification, leading to questions about appropriate analysis adjustment.

Purpose of the Study:

• To evaluate different covariate adjustment strategies following stratified randomization.
• To compare adjustment for randomization categories versus continuous values, considering linear and non-linear covariate-outcome relationships.

Main Methods:

• Data simulation was used to assess adjustment strategies for continuous and binary outcomes.
• Analysis included unadjusted models, adjustment for categories, and adjustment for continuous values using linear models, fractional polynomials, and restricted cubic splines.
• The impact of complete versus missing outcome data was investigated.

Main Results:

• Unadjusted analyses performed poorly.
• Adjustment methods misspecifying the covariate-outcome relationship were less powerful and biased, particularly with missing data.
• Adjustment for randomization categories often resulted in the highest degree of misspecification.

Conclusions:

• Adjustment for randomization categories should be avoided.
• Flexible adjustment methods (fractional polynomials, restricted cubic splines) are recommended for continuous stratification variables to prevent misspecification and bias.