experimental-design

Use block_randomization for balanced assignment throughout enrollment, or stratified_block_randomization to balance a prognostic factor like site or sex. Both return seeded pandas DataFrames that can be saved as an auditable allocation schedule.

Simple randomization assigns each unit independently, which can produce arm imbalance with small samples. Permuted-block randomization guarantees the target ratio within each block, keeping arms balanced throughout sequential enrollment.

Use a Plackett-Burman or resolution III fractional factorial design to screen 5 or more factors cheaply, estimating main effects only. Follow up on the important factors with a full factorial or response-surface design.

Yes, pyDOE3 provides central-composite and Box-Behnken generators, which the doe_designs.py script wraps to return designs in real factor units. These fit quadratic models to locate optima, unlike two-level factorial designs.

Pseudoreplication is treating repeated measurements of one unit as independent replicates, such as counting 100 cells from 3 mice as n=300. The replicate must be at the level the treatment is randomized, and the analysis must respect the nesting with a mixed model.

Use cluster randomization when the intervention is delivered to whole groups like clinics or classrooms rather than individuals. Power depends mainly on the number of clusters and the intraclass correlation, and analysis must account for clustering.