Semiparametric Allocation of Subjects to Cohort Strata

Area of Science:

• Biostatistics
• Epidemiology
• Observational Studies

Background:

• Stratum assignment is crucial for causal inference in observational studies.
• Traditional methods often rely on strong modeling assumptions.
• Small cohort studies present unique challenges for robust stratum assignment.

Purpose of the Study:

• To present a novel method for stratum assignment in small cohort studies.
• To avoid making strong modeling assumptions.
• To improve covariate balance and reduce bias in treatment effect estimation.

Main Methods:

• Utilized off-the-shelf software (rgenoud) for stratum assignments.
• Developed a loss function based on within-stratum and population-adjusted Euclidean distances.
• Employed loss minimization to optimize stratum assignments for covariate balance.

Main Results:

• Minimized Euclidean distance loss reduced covariate imbalance by a median of 99% in simulated data.
• Achieved substantial covariate imbalance reduction compared to propensity score stratification and inverse probability weighting.
• Demonstrated effective covariate balance in a real-world cohort of children undergoing immunotherapy.

Conclusions:

• Semiparametric stratum-assignment algorithms allow for tailored loss functions to meet specific design goals.
• A loss function emphasizing covariate balance proved effective in initial testing.
• This method offers a flexible and assumption-lean approach to stratum assignment in small cohort studies.